Raspberry Pi 4 PCIe bridge “chip” | Hacker News Reader

Raspberry Pi 4 PCIe bridge “chip”

blog.zakkemble.net

236 pointsfanf2177 comments

I'm starting to think RPi owners just don't want to acknowledge that there are other embedded linux-capable dev boards out there that can do all the things they want (eMMC, PCIe, multiple USB3, eSATA) without going through all of these insane hacks.

If it's for the adventure, cool. If it's because you are pissed at the RPi folks for not including it then holy crap widen your horizons....

I think the "spirit" of the RPi is what encourages this hacker/maker mentality....which is perfectly fine by me.

Funny. I thought the original spirit of RPi was to be a BBC Micro to a new generation of schoolkids and let them explore a modern desktop programming system for cheap.

The "let's hack this into a Beowulf cluster!" stuff was secondary. Now it seems to be primary and the education stuff is, um, not even mentioned anymore?

It is still mentioned, in the context of UK education.

The education market may have been seen as larger and more "addressable" when this was all in the planning stages.

This is what happens when you let people explore something for cheap: they do all kinds of things you didn’t expect.

The outcome couldn’t have been better.

This is exactly it.

There's a group of guys I build hardware/firmware for. They are experienced engineers, but their expertise is in machine tools.

We met because they made a post on an Arduino forum a few years ago basically saying "we need to do X. We know it has to be possible in a cost-effective way because of all the cheap Arduino, Raspberry Pi, etc. hardware we keep seeing for sale but we don't have any experience on those platforms. Can anyone help us?"

This illustrates a point that a lot of people don't see: while Arduino or RPi were released for a specific purpose, the absolutely massive ecosystem that grew up around them enabled a slew of entirely new classes of applications that weren't feasible before due to cost.

Back to "my" guys. If I quoted them an upfront design for what they needed, starting from scratch, it would have reached into the five figures easily. Instead I was able to say "hey, I'm working on something that has a lot in common with that. I'll build you a prototype of what you need for the cost of the hardware" and we went from there. Now that project makes me a few $k per year.

> Now it seems to be primary and the education stuff is, um, not even mentioned anymore?

Except at the very top of the website?

i think it still is that, but achieves it so easily as to be unremarkable. rpi has probably introduced many people to linux and debian. it might seem trivial, but anything that gives ordinary people a good experience with open computing should be considered a massive victory.

Do you mind to name the brand/series of the products in the same price range?

*really wants to know, not trolling

Yeah, I know you're not trolling but that comes up every.single.time. alternatives to RPi are mentioned. We get it: it's cheap. I can get a running Linux board for $30 so kiss off.

There's no other board in this range because the other chipmakers aren't subsidizing their boards or ecosystems in this fashion. So there's no answer to that.

But I know what my time is worth and I'd rather spend $150-$200 more on a Beagle X-15 or Boundary Nitrogen or Toradex Colibri and get the documentation and schematics and all the source I need to make it do things like eMMC and PCIe. If I spent more than 2-3 hours on hacking chips into the board I've already lost.

Well, for hobbyists just messing around on their own time, "just pay $200, your time should be worth more than that" isn't really a sensible answer.

If you're a software engineer with the skills to mess with low level interfaces like that, $200 doesn't seem that insane, especially since those are general purpose boards that can easily be repurposed. I mean of course I realize that for many people $200 is a lot of money but I'm willing to bet that for the average HNer it's really not that big of a deal.

Besides TFA's chip requires desoldering and resoldering an SMT chip from the RPi's tiny and densely populated motherboard, not exactly Soldering 101.

Moreover if you're really just starting with low level programming, want to mess with basic electronics and things like that then don't get into PCIe, actually don't even get into Linux. Just get a cheap-o ARM microcontroller board, a multimeter and a $10 saleae digital analyzer clone from ebay and play with that, you'll learn a lot more that way.

$200 isn't a big deal for something you really want, but at the margin, on projects that you don't even know you're going to bother finishing, it's the difference between being able to justify taking a punt on it and not.

That's why the Raspberry Pi is so popular.

I agree with everything you said about using cheap microcontrollers to experiment with basic electronics projects.

Not if you want to install such a computer in every single room of your house.

Or if you want to build a cluster, like this guy:

https://www.youtube.com/watch?v=78H-4KqVvrg

I've played around with a few other boards besides Pi, and I'm always let down by the kernel support. A lot of them are stuck on ancient kernels or you're at the mercy of their BSP drops.

The Beagles at least have mainline kernel support, but things like pine64 and odroid still rely on a lot of patches or binary blobs.

Thanks for pointing out the X-15, though.. I've not looked at Beagles since the Beaglebone Black. It looks like an impressive board with mainline kernel support.

The Allwinner A64 has nearly complete mainline kernel support at this point, hence the release of the PinePhone: https://linux-sunxi.org/Linux_mainlining_effort#Status_Matri...

Until you need to buy thousands of boards. And having a maintained software stack for a long time has a very high value in my humble opinion.

> There's no other board in this range because the other chipmakers aren't subsidizing their boards or ecosystems in this fashion.

Can you elaborate on this? Who is subsidising the Pi, and why?

Broadcom. They get a lot of kickbacks from the British government, AFAIK.

Edit: Not sure why downvoted. Broadcom very much is able to write off a ton as donations to charity, since the rpi foundation is a British charity. They also get sweet heart deals from the UK based on that. The UK gets the ability to put a passable personal computer on every child's desk, like they've been attempting to do since the BBC micro days.

> Broadcom very much is able to write off a ton as donations to charity

If I donate $1000 dollars so that I can "write it off," do I come out ahead financially?

https://www.youtube.com/watch?v=XEL65gywwHQ

If you count your engineers' spare time as billable hours donated to charity, and then sell the chips to that charity at cost, but write off the profit as charitable donations to reduce your overall tax burden, you sure do. Particularly for chips that weren't going to be sold anyway (the original SoC was from a Roku like device that had been EoLed and Broadcom had been left with a surplus of the chips they didn't know what to do with).

Which I'm not counting as a negative. This is successful example of charitable giving where everyone wins.

Just because Kramer didn't know the process behind writing things off, doesn't mean that there isn't a real process and it isn't being used in this case.

This is the whole reason why the rpi foundation is a charity, to allow these financial structures.

> Just because Kramer didn't know the process behind writing things off, doesn't mean that there isn't a real process and it isn't being used in this case.

I'm not saying it isn't a real process. I'm saying it isn't magic. Reducing your taxable income by $1,000 doesn't mean you save $1,000.

I never said it did.

But marking your engineers' spare time as donated since your contract owns everything they produce anyway, and selling chips at cost and marking the profit as a donation when you were probably going to take a loss on them does in fact net you ahead.

My gut says these SoCs are being sold below cost.

IDK, even rpi4 is still on 28nm which is the cheapest process per gate right now. And they really have very few gates allthibgs considered. I could absolutely see somewhere between $5 and $10 at cost, not counting R&D.

Where are British schoolchildren getting these Pis? Mine aren’t. The school where my sister works isn’t giving them out, either.

The closest I’ve seen to what you describe was a BBC Microbit my son was given as part of holiday club but that was sponsored by Maxell, not the British government, Broadcom, or anyone else.

https://www.computerworld.com/article/3435277/how-the-raspbe...

Also the Microbit was designed by BBC R&D with government grants. Maxell is just the distributor.

Disclaimer: I am a STEM Ambassador (volunteering position) in West Sussex, working with local schools and libraries on IT projects, including using the Raspberry Pi.

The Raspberry Pi isn't a giveaway item and so schools need to make an investment in hardware or obtain sponsorship to kit out their STEM resources with them.

1M BBC micro:bits were distributed free to Yr 7 pupils as part of a project in 2016 - schools had to register/apply at the time, so if your school did not participate their pupils would not have received anything, however, as you have mentioned, other sponsorship also made units available.

Best have a word with your school's STEM/Technology people to see how they are supplementing your children's curriculum with technology resources.

The RockPro64 has an x4 PCIe slot. The 2gb model is 60$ and the 4gb one is 80$. Add ~20$ for accessories and you have a nice little board that should be faster than the RPI and has more features.

If it is anything like the Rock64, it will also be unstable, negating much of the benefit.

I have one that's run well for a year. It's only been offline when I've powered it off myself.

Mine hard locks regularly requiring a power cycle. It is even worse with more recent kernels.

Rockchip's PCIe is kind of a joke, you can't use a GPU there because there's not enough BAR space. But yeah, you could run an NVMe drive or a big NIC, I guess.

Also, the heterogeneous 2xA72+4xA53 CPU setup is ehh.. :/ What's really cool about the RPi4 is that there's four A72 cores. And now 8GB RAM. And https://rpi4-uefi.dev.

I've put a 4-port SATA controller in mine and use it as a NAS. Haven't benchmarked it but it works well.

$200 for an ARM Board? No thanks.

Yes, but for the cases where you don't need to do that, the Pi can blow it out of the water.

I have an STM32-based controller board next to me (I'm a firmware developer) and I had originally suggested that instead of us spinning up a custom board, we use a BeagleBone or RPi Compute Module and build out the I/O and take advantage of all the OSS out there. I got a lot of pushback that I wasn't willing to fight (I said it a couple years ago with the same results), but if we had gone that route, the entire project would have been done months ago. Same would have been true the first time I brought it up on that other project.

Sure the Pi's are cheap and attractive, but it's the ecosystem that provides the real value. And these days, with software usually being the main cost of the project, you can't beat OSS.

I recently did an embedded job where we essentially used a mkrzero as our dev board. It saved us a ton of time on writing the firmware, since we didn’t need to wait for and debug prototypes. And the final schematic took way less time as well, since we could reference the mkrzero for things that might require guessing what the datasheets actually require.

There’s a lot to be said for having a commercial product with multiple revisions to avoid painful guesswork.

The Rev4 I referred to in this post: https://news.ycombinator.com/item?id=23680157 used an OSHW board as a reference for a chip I had never used before and I was able to incorporate the Arduino library for it. Just brought the new hardware up yesterday. OSS/OSH saved me tons of time.

>> There's no other board in this range because the other chipmakers aren't subsidizing their boards or ecosystems in this fashion. So there's no answer to that.

So the answer is "No, there is no alternative in the same price range."

The places I know of to get comparable SBCs for comparable prices are these.:

https://shop.popcorncomputer.com/ https://store.pine64.org/

That's assuming you don't want to buy a cheap phone and dismantle it (comes with battery, charging circuitry, touchscreen, audio, flash storage, etc).

Beaglebone Black or Pocketbeagle.

And, as a plus, everything is documented. And TI will talk to you.

The software support and community of the Raspberry Pi is several orders of magnitude better than anything that exists for the other boards you're mentioning. "Embedded linux capable" is different than having many linux distributions, and other operating systems, treat the boards as a fully supported target.

It's true that RPi owners don't know about a lot of that other stuff, but there are reasons they don't know about them.

And it kind of has to be when Broadcom won't answer their pleas for help, right? I've worked with Broadcom and they won't even pick up the phone unless you're moving 100K EAU. (and, protip: when you do get technical help or source it's complete crap)

Your case is maybe kind of illustrative of what I was talking about. The number of RPi users who work at the level where they might have questions for Broadcom is vanishingly small, and that's okay. Almost all people want to write software in userspace, even if it's software that does cool embedded, IoT type stuff.

What's the best right now in terms of openness and support for someone working at the driver level?

I've been a Freescale/NXP customer for decades and I like their stuff. Their documentation is comprehensive, their technical notes and errata are free to read. They're not perfect but they work at it. Their Github presence is decent. iMX is a good line for Linux, LPC for microcontrollers but the new iMXRT parts are looking great for FreeRTOS.

I don't work as much with TI but I've used Panda in the past for OMAP4 and some personal work with Beagle/AM335x and it's solid stuff. I'd like to think Beagle's community is as good as RPi's, just a little more low-key (and the chipmaker participates as well)

iMX8 ist the most confusing product line ever, tough.

Yeah not gonna argue with you there. NXP has kind of muddied up all the product families with variants on core count and coprocessors.

The iMX6 used to be single, dual, (solo/duallite), quad. Done. Now there's what - 5 other variants? 7?

I've always been a fan of the iMX line - they're my goto reference designs for low volume embedded work along with STM32Fx for MCUs - but the BSPs are just not comparable for hobbyists.

With RaspPi you can download an SD card image fully set up with a simple "wpa_supplicant.conf" for WiFi and a Bounjour discovery service to register a hostname like 'pi.local'. Takes like an hour with a crappy connection and then another hour at most for apt-get update. Need to hook a security camera? Download a prebuilkt MotionEyeOS image. Need to control a CNC with GCode and a .Net core API? Grab the Duet 3 and a raspberry pi.

If you can use Multistrap then all you need is a linux kernel and you can build a full Debian system for your Linux chip.

Thanks for your reply. I've been working with Raspberry Pi and Nvidia Jetson hardware lately, and there is a pretty interesting (ahem) dichotomy in terms of friendliness of support and community. (maybe Nvidia's paying customers get more) I can see how one working at a very low level might regard both as being not sufficient.

I have some good memories of using TI embedded hardware (MSP430) and the documentation being pretty good. That's a different paradigm than a whole Unix system, though, even if it's a system that's going to be used for embedded stuff. It's nice to hear the Beagle situation is good. I remember those boards getting a hearty endorsement from Andrew Tanenbaum but otherwise I don't know too much about them.

I suggest giving TI (Beagle) a serious look.

It should also be said that Raspberry Pi was started by a Broadcom employee who is still on the payroll to this day, afaik.

Absolutely. The main reason I made a Raspberry Pi cluster was because the community support and pathways previously walked for me were there. It is _difficult_ to start on a less-common SBC.

And then when I knew more I found aliexpress and the Orange Pi Zero LTS.

This is why I'm not a fan of GNU/Linux. There are 27 different distros which all support different hardware, so you can't just pick one and run it across whatever hardware you want.

I run NetBSD, and the same exact software that runs on a Pi will also run on a RockPro64, just like the drivers for a zillion PCIe devices which work on other platforms will also work on a RockPro64.

GNU/Linux really should share more between distros instead of always needing distros to be able to differentiate themselves from each other.

I'm not really sure what you're talking about. There's two cases:

Drivers are open source and part of the kernel. In that case, if a distro hasn't enabled support in their kernel build, you can build your own. Or compile it in with dkms if it's out of tree.

Or it's a proprietary driver, and then it might only work with certain distros, but in this case I seriously doubt the HW would work with NetBSD either.

You can pick one and run it across whatever hardware you want. The problem is that ARM SBCs require device specific builds and that is purely a problem with SBC and SoC vendors and has nothing to do with Linux distributions.

The caveat to "fully supported" is that once you start replacing chips on the board it really isn't.

That's not true at all. There are boards that have mainline support quicker after release, there are SoCs really supported by the manufacturer, there are SoCs with much better documentation, etc.

> It's true that RPi owners don't know about a lot of that other stuff, but there are reasons they don't know about them.

You're just trying to explain away ignorance, and prejudice really.

I’ve worked with rPis, Beaglebones, that NTc Linux board, and nanoPis. As somebody who is not a Linux Guru, the RPi is extremely nice because I can google anything I want to do or any error I get + raspberry pi and get a decent jumping off point. The other boards just don’t have the sheer volume of “stupid Linux-ignorant questions” that the rPi does.

The other nice thing is the massive community means that software gets tailored to the Pi’s capabilities - things like octoPi have been invaluable to me. Sure, I could assemble the same thing myself, but having somebody else do the heavy lifting is really nice.

The final nice thing is that rPis are easy to buy. With other boards I have to pay for shipping from the states or wait for shipping from China. With the rPi, I have one distributor that’s local to me, and a second where shipping is pretty cheap.

Seems a bit like self-inflicted wound though. If I look at the /boot dir on my Rpi2/Arch Linux ARM I don't recognize anything there. No u-boot, not boot.scr, nothing standard.

If I wanted to configure display, read thermals, cpufreq I had to use some rpi specific tools in /opt, intstead of using standard kernel interfaces.

Trying to add a TFT display to RPi involved googling and finding a bunch of outdated/contradictory advice and then still having to figure it out myself with experimentation.

Locally bought Rpi2 at the time was also 3x the price of the better specced Orange Pi PC. And if I wanted to use video decoding I had to pay even more.

I mean, it's great it works for you. My interest was in learning, and it felt like the things I would learn about Rpi would not help me that much with any other boards, due to quirkiness of Rpi's SoC design, and nonstandard OS interfaces. So I bailed early, and focused on something else. I also found the SoC to be rather closed.

I think the Odroid folks have some pretty decent software support.

Certainly and performance/$ is usually better too.

I haven't found anything that's been able to match the price point of the pi for the amount of things you can do with it. I'd love an x86 system of similar footprint and cost. Are there some I don't know about?

Here's a great place to start your search https://www.hackerboards.com/

I think this actually highlights part of the problem. There's certainly a ton of capable products there, lots of stuff that looks very fun to play with. You can probably find something with whatever specific combo of features you need for your project.

However, there's two big issues: price and availability.

Price is obvious. If you want to spend $200 or more you have a massive number of options. However, this is a far cry from the $35-75 for a Pi, plus for both you still need a power supply, case, etc.

I can also order a Pi from a couple different sites, and probably pick one up locally in a pinch, too. It seems like there's lot of other boards where you can't actually buy one. Some are in the category of "how many thousand would you like?" It's just a totally different world than the Pi's distribution and ecosystem, to the point it's not really fair to consider for the average hobbyist.

There's dozens of readily available case options and other add-ons for the Pi.

I also know if I build software for the Pi I'll still be able to buy another one anytime I want for the next few years, and due to its popularity chances are for the next decade (if not a real Pi, someone will make a clone).

It's great there's options, but the fact a product merely exists doesn't instantly make it a viable alternative.

The sheet quantity of boards here is another problem - I'd personally much prefer a curated list, with recommendations for different use cases.

There are two major distributions aimed at more than one board, they're Armbian and DietPI, I would suggest reading through their forums because of the good number of people familiar with different boards, so those might likely be the best places to ask if task X is better accomplished by board A or B etc. . Sometimes however boards are similar to each other for employing the same SoC, so it's not rare finding an image that boots successfully both a board from one vendor and a TV Box from another vendor. There are many ready images for tested hardware, but users are encouraged to rebuild the images and make more test to unknown hardware then report findings.

https://www.armbian.com/

https://dietpi.com/ (warning: puke inducing site colors! :-)

What is puke inducing there? The green?

I'd call that color scheme Liquorice-Absinthmeister :-)

> What is puke inducing there? The green?

Yes, I find it extremely uncomfortable to look at; literally a punch in the eye.

$10 to $75 for a Pi - the RPi Zero can be gotten for $10 and it remains very capable. The Pi has no competition among SBCs on bang-for-your-buck: I have extensively checked. You can get lower prices (and lower memory) on microcontroller systems, but for hobbyists, you'll be additionally paying with your time

The RPi Zero has been a joke for two full years: produced in extremely small quantities and sold at loss just to keep customers away from the competition. Of course there are a few people who somehow managed to get one or two without being forced to purchase it bundled with unnecessary overpriced stuff, and they usually were very quick to point that out in forums (while downvoting any criticism), but they were a really small minority. Things may have changed, but too late: I've removed the RPi from the list of the boards I could purchase for anything but video playing. My video player has been a RPi1, then 2, then 3 now 4, but all the other older Pi boards lie now unused in a drawer while I rather use other stuff (Odroids, Orange PIs, NanoPIs, probably getting soon a Pine64B from the excellent pine64.org guys, etc.). The RPi has a lot of competition, almost all of it more powerful or more cheap, sometimes even both.

And almost none of it available in my country without paying more in shipping than the actual product.

Are you sure about that? Many of these boards are sold by Chinese vendors, and usually shipping from China/HK is very cheap because their govt subsidizes shipping. You also have the fast courier option which of course will cost a lot more, but that's just one of the options. When contacting them just be sure they use a tracked method, and much better if they support PayPal which employs an effective user protection against missed or damaged deliveries, faulty products, fakes etc. That's what for example pushed me away from buying more Orange PI boards years ago since although the products are really good their manufacturer and vendor gave no other option than their Aliexpress shop that didn't support PayPal.

> Many of these boards are sold by Chinese vendors

Which means my country’s very good consumer protections don’t apply, which immediately rules them out for me.

No risk? No fun! Boring...

Seen pragmatically your "protection" in buying such small things, where the transport costs more than the things themselves is denying the seller a good rating, until one item that is functional arrives. AND/OR disputing via whatever payment service, or simply writing it off.

I don’t know why you equate fun and risk; it’s my limited expendable cash we’re talking, not yours.

In countries where consumer protections are a complete joke, and a bad purchase is as good as a write-off, I could perhaps understand this sentiment.

However, I don’t have money to splash out on things that might be broken on arrival. I’ve already bought a few Raspberry Pis that needed returning and doing so was hassle free as soon as I quoted the Consumer Guarantees and Fair Trading Acts. At most places, I don’t need to bother.

(Another benefit is I never need to buy extended warranties; legal protections here tend to last longer than those anyway!)

If I can get the product in New Zealand via a local supplier, and I tend to prefer to support local businesses anyway, I can have just as much fun without any risk.

I know what you mean, and didn't intend to make fun of you and understand your reasoning.

I just wrote that because of my experience in Germany, where I'm from, one tends to see the sourcing of gadgets from China like that.

I once ordered a NanoPi Neo2 with 1GB RAM, some 'Sata-Hat' for it, and a fitting metal case made of aluminum. Was about 50 US$ with the wall plug, directly from the vendor, waiting some weeks for delivery, and the possibility of hassle with customs, though unlikely for something of that price.

So i applied your reasoning and sourced it locally. BANG! 99EUR!!! Didn't make that much sense anymore, economically, but i wanted it. Then the delivery service fucked up, it got returned/canceled by the delivery service because I complained about their reason "could not find address", while having seen the driver speeding away at the end of his tour from my window, which was obvious from the online tracking with real time map which they offered.

Had to tell the vendor I did NOT cancel, please send again, this time with a more reliable delivery service, if possible.

Finally I got it! (That was the 'fun', the rest I could have done without)

Would it break now, i'd have no consumer protection at all because i bought it way more than half a year ago, which is how long ours last.

Since I'm no power shopper, this made me think about my/your reasoning, and the conclusion that if you want something from "there", either the 'no risk, no fun' thing is the way to go, or it doesn't exist, since sourcing it locally at least doubles the price, or it isn't available at all.

Ordered a few things since then, had no hassle, everything was good, and at least 3 times cheaper.

Even have Alipay connected to my bank/checking account and identity card, which was possible to set up via browser at the time.

The idea of supporting local businesses is appealing in theory, but seems to hail from bygone ages. Again, as seen from Germany, because nowadays many here are either clueless and/or scammy, which takes the 'fun' out of it.

So, coming back to the little expendable cash, the only remaining question for me is: Does it make sense to have some man in the middle delivering my cheap things for at least 50% more?

Odroids are consistently more expensive than nearest Raspberry equivalent, older Orange Pis (Orange Pi Zero) are even harder to find that RPi Zeros! "Harder" is the wrong word - Orange Pi Zeros are impossible to find. The upcoming wifi cam from Pine64 is the best-next-thing to the RPi 0 for my use case and price range, but its not yet out.

For now, nothing beats the RPi 0 on price, as well as features-per-dollar. The closest is an Onion M2, IIRC. It's $9 before shipping, but only has 64MB memory. I'm prototyping /piloting on RPi Zeros, but if I end up scaling to 1000's of sites (which is why I cant use a $200 SBC per site. It's a self-funded passion project), I'll likely have to switch to the Onions due to the limited availability of RPis you mentioned

Popcorn Computer is very close to shipping. https://shop.popcorncomputer.com/

Pine64 has a number of SBCs and also other products that build around them. https://store.pine64.org/product-category/pine-h64-boards/

Both offer things in similar price ranges as Raspberry Pi. I have some of each (if you consider the Popcorn to be a true successor to the CHiP as the seller does).

I really wanted the CHiP to succeed and was going to grab a pocket chip when I had the spare cash. Unfortunate what happened, but if popcorn ship, it looks like a good successor.

This list is missing a very important column -- "quality of Linux support"

There is a great difference between a board which uses mainline kernel and is supported by major distros, and a board which can only run specific old version of OpenWRT and a patched 2.6.x kernel.

https://www.explainingcomputers.com/sbc.html lists a few other x86 single board computers, but the only ones that come close to the $35 price point that I know of are the Atomic PI (quad core Atom, 1.4GHz boosting to 1.9GHz, started at $35 and rose up to $50 per the video review) and the LattePanda (similar specs, more USB ports and needs a cooling solution) for ~$90.

The Raspi hasn't been anywhere near $35 for the decent versions for a while now, though. And while the 4gb pi4 is $55, the 8gb version (which anyone who wants a pi would want because that memory makes a ton of difference) is $75.

Sure, the 2gb version is cheap, but it's also borderline useless if you want put a pi4 to use as a computer instead of fancy arduino.

And while they're both "computer-board-things", we can't really compare a LattePanda to a Pi: the only reason LattePanda boards exist is because they allow the same kind of prototyping/embedding as the pi, but specifically for windows. There is very little reason to buy one unless you _need_ a board that runs regular windows (because you need to run closed source, windows-only tools, for instance).

I wish to point out that everyone's use case and performance targets are different.

I have been happily using my RPi2 as a home dns, file and media server for at least a year now. While I agree it is constrained by memory, IO and compute, it is adequate for my purposes. Though I am upgrading to an RPi4 4GB, it is a nice upgrade, not a requirement (for me).

I presume there are some people who wish to run x86 binaries on Linux or Windows, for which a LattePanda might be a good fit at a modest cost. I wish them success, which is more fun than failure.

> Sure, the 2gb version is cheap, but it's also borderline useless if you want put a pi4 to use as a computer

I remember when EMACS was "Eight Megabytes And Constantly Swapping"; there is something deeply fucking wrong with modern software if two thousand megabytes isn't enough for it to run smoothly.

I ran a complete KDE3 desktop back in the day with 256MB of RAM shared with Kopete and inline videos, Konqueror with several tabs open AND Amarok in the background.

These kids are truly spoiled. Even with 1GB today could be enough if you know how to set Unbound + a light browser.

You say that as if you never bought a raspberry pi because it lets you use Mathematica. There's many of us who did, because a pi is <$100 and a home use license for Mathematica is $300+

If you need a pi because you want to run something on it that'll run just as well on a pi 2, then sure 2GB is perfectly fine. But if you actually need a pi 4, you need it because it is vastly more powerful and offers actually useful amounts of memory.

Why would I? Mathematica sucks, as you should be able to infer just from (in your own words):

> a home use license for Mathematica is $300+

That's just a bizarre thing to conclude, I think we're done here.

You must be young, kid. My GF's laptop is 2GB and runs everything with Slackware and even Chrome really fast. Double Celeron, Intel Mobile 4 series.

You must be confused, kid. An intel laptop CPU performs nothing like a raspberry pi. That's like saying "your complaint about going uphill on a bike is nonsense because my GF's motorcycle does that just fine".

Eh, GF's dual core Celeron craps on your Rpi2 and 3 probably, but with 2GB of RAM is powerful enough to even watch JS filled pages if use Ublock Origin. Cough, Medium, Cough.

Is no SandyBridge either.

Right, so you agree that comparing a Celeron to an R2/R3 is apples to oranges.

The devices are wildly different, with the amount of RAM available being far less important than the fact that your GF's laptop, with a vastly more complex motherboard, actually lets the OS efficiently use a swapfile. 2GB on a laptop goes a lot farther than 2GB on an Raspberry Pi.

Not the same price point, but I have had my eye on the UDOO X86 II ADVANCED PLUS [1].

[1] https://shop.udoo.org/udoo-x86-ii-advanced-plus.html

I'm sure a person who does this amount of work on the project did a 2 minute Google search for similar products even if they somehow weren't aware of other boards.

Also, did you forget you're on HackerNews? Sometimes people do things just because it's interesting or they're curious.

I think it's like Arduino. If you're not well versed into the world of embedded boards it can be tricky to navigate all the offerings. RPi is popular, massively well supported, you can find tons of projects online using it for almost anything, it seems like the safe choice.

I mean, I have the strong suspicion that the vast majority of RPi owners wouldn't be able to cite the name of a single alternative. Not a one. RPi's popularity eclipsed everything else (like Arduino in its heyday).

And even in your comment you haven't mentioned any alternative yourself, only that they exist. If you care about people using alternatives you should start with that. Actually, I should too, but the only Embedded Linux boards with a PCIe interface I've used are super expensive FPGA-based ones, so not really in the same category.

Even as someone well versed in the embedded world (my previous job had me as a lead for a proprietary high availability RTOS, which I personally ported to several ISAs), I still prefer RPi to the other boards. I normally get these boards with an endgoal in mind that isn't just "play with the board". While none of these boards have anything approaching the real support you'd get as a first rate developer, at least the RPis have an ecosystem of exactly the kind of stuff in the article so you have an equivalent to support. The Pine64 peeps seem close, but I haven't dug too deep there. Everything else seems to be "oh look, you can technically run Linux, but don't touch anything that the vendor provided for you".

Exactly, I tried using Omega Onion’s but it turns out you can really only use their version of Linux. They claim support for booting from an SD card, but not really. The Chinese based Orange/Nano’s use Allwinner soc’s and the mainstream Linux support for those have gotten better, but nowhere near the pi’s.

Though must say TI AM3* boards (beagle’s) have good mainline Linux support and they appear to do all their work on public got repos. Now the iMX* boards seem nice but I haven’t really found a maker board variant.

Have you ever looked at beaglebones? TI's support is pretty nice in my experience, and the hardware is a LOT more solid and reliable. Certainly the USB stack has been a weak point on the RPi, which is problematic since the ethernet adapter is USB.

For most embedded work which needs to be reliable, the RPi is to much designed to a price, with enough tradeoffs making it an unattractive and ultimately more expensive option. In a PoC we did, we had to replace RPi's weekly in the field, racking up intervention costs. This completely disappeared the moment we switched to beaglebones.

I quite like the RPi for what it is, a cheap & fun platform offering a lot of possibilities to the masses, but I wouldn't consider it to be a serious embedded platform to target you can depend on. But I've used it as a video/slideshow player in real-world situations, where if they'd drop out, there aren't any real consequences (which most of the time were cards failing, and yes, the root-fs was read-only...)

Worth looking at NUC and second hand Lenovo M600 fanless units as well. 900% less crap and I reckon 90% of users don’t plug anything directly into IO ports on their Pi.

But then you're looking at double or quadruple the price of even a fancy RPi. Even something like the Odroid H2plus[1] is going to be pretty spendy once you add ram and a ssd (though it's x86, so I would expect it would need a lot less fiddling with)

1: https://www.hardkernel.com/shop/odroid-h2plus/

Yes but the amount of time you’re not wasting is worth it.

Perhaps for you, but 2-3x the cost is an entirely different product.

You tend to have to buy 2-3 times as many SoC boards in the same lifetime as the other item from experience. Thus the savings up front are artificial.

I’ve got two dead pi’s here ceased to function while in unattended operation.

I don't think you'll see much positivity with your comment. This is old school hacking in the new age of computing; I have this one thing, and I want to hook it to another thing, let's figure out how to do that. I think it's amazing, and shows just how flexible and hackable the Pi can be. Of course we can throw bookoo bucks at a problem to make it disappear, or buy a specific built machine to do what we want. But there's no challenge in that, no engineering, and it's a bit of a cop out to lament that people are hacking hardware this way...

Yeah, I know. I always regret commenting on RPi projects because when you do this kind of stuff day-in-day-out for a living you want to suggest alternatives that could save people a lot of busywork and lost psychic energy. Because that's what I would do for my coworkers.

But I'm pissing into the wind at this point. Got my cheap ARM board, fuck you. And that's fine. Enjoy the adventure. I'm all for hacking, but reinventing the wheel isn't my thing.

The Raspberry Pi just works and there's tons of documentation. It's not about cost. I don't want to spend my time digging into the ins and outs of a specific board, which is exactly saving me a bunch of busywork and time.

If there is a new version or my existing one breaks, I'm relatively certain I can get a replacement, and if I need to upgrade, it will be painless.

That said, there is nothing wrong with trying to shoehorn features into something you've got already, vs buying something unknown.

See, to me that sounds like RPi is the only board with a working software ecosystem, active contributors, and available source code to help get your project done.

But that's not really the case.

Depends what you mean by “ecosystem”. If you’re at the level of a hobbyist who’s considers a blog post, Hackaday project, or YouTube tutorial about setting up some server or building some type of project part of the ecosystem, then the difference between the size and accessibility of the RPi and its competitors is staggering.

Sure, for any one project an alternative _may_ be the right solution, especially if you’re spending serious time on it, but when you’re constantly creating little projects and proto-projects, many of which involve little more than copying from a few blog posts, writing a bit of code for an evening or two, and deciding not to go much further, it is just SO much less risky in terms of time and probability of major hiccups to start out on an RPi. Plus, if you’re like me there’s probably one lying around somewhere so you get instant gratification and zero incremental cost to give it a try, which is often the difference between scratching the itch and not on that particular project.

As someone who used to work on embedded hardware day-in and day-out (though it's been close to a decade since I've worked in that sector), I can assure you that really is the case, if you consider a more long-term view.

Every time I look at one of the alternatives to the RPi, I just don't have confidence in the tools and community. I don't want to have to do what I used to do when I worked on embedded hardware for a living. It was fiddly and annoying (though it did have its truimphs), and mainly only tolerable because I was getting paid to do it.

I want a stable, working, well-supported, upstreamed platform that I can build on without too much fuss, and also feel confident that I can keep upgrading the OS and kernel to whatever is the latest, for the next 5-10 years, even if the company behind it disappears.

I play with this stuff in my spare time, and fighting with a chip vendor's nearly-always-shitty BSP is not what I consider a good use of my time. You (being someone who currently works on this stuff day-in and day-out) are in a very unique position that you can leverage to make it easy and productive for you to build on random hardware platforms, especially the ones that aren't even mainstream RPi competitors; most people posting here about this sort of thing are... not in that position.

FYI: the phrase is beaucoup bucks - appropriated from French

It's not unusual to spell "beaucoup" phonetically on purpose, to convey an Americanized pronunciation.

A few years ago I tried switching off Raspberry Pi's after getting fed up with all the proprietary Broadcom nonsense and SD card (un)reliability. Specifically to the OrangePi. Wow the experience was terrible and back to Raspberry it was.

Recently doing another project and ran into the same limitations on the pi4. This time tried the NanoPi m4v2. Wow is it amazing. Native PCIe slot with a NVMe/SATA/USB3 riser hats are fantastic: https://www.friendlyarm.com/image/catalog/description/M4v2_e... - with the SATA hat you can build the equivalent of Synology DS918+ NAS for 1/10th the price.

Won’t be going back to Raspberry. Specifically this is for the use case of a few hundred SBC's deployed in the field for commercial/light industrial - but if you’re a casual Raspberry user do reevaluate the NanoPi. (edit: or the OrangePi apparently, see below!) I’ve found the Raspberry Pi software support to be an ocean wide and a frisbee deep. For anything more involved than a casual weekend project you’re going to have to bushwhack either way.

Most of the OrangePi lineup has full mainline kernel support now, thanks to projects like Linux-SUNXI and Armbian.

Here is the status matrix: https://linux-sunxi.org/Linux_mainlining_effort#Status_Matri...

Rockchip generally cooperates with Linux driver developers. The last time I checked, Linux-SUNXI was still stuck reverse engineering everything because Allwinner is actively hostile to driver developers. Has that changed?

I wouldn't call that full at all. That matrix doesn't show full support for any of the listed SoCs, and some of them are missing some pretty critical things IMO.

And other things raise big red flags for me, like:

> sunxi-musb driver lacks DMA support (with current driver, USB gadgets are limited to PIO, limiting speed to 10MiB/s and causing large CPU)

That... seems like a pretty big deal for anyone who wants to do anything with USB.

It's also telling that it appears that all the upstreaming efforts seem to be community-driven, which is an even bigger red flag to me; the main page of the wiki you linked calls out Allwinner for their hostile attitude toward open source and the GPL[0]. It's fantastic that such a community has formed and works hard to fill in the gaps, but I'm not sure I want to (for example) build a NAS out of an OrangePi when I'm not confident that in 5-10 years I'll be able to run a future-current OS and kernel on it.

The RPi's support isn't perfect, and not everything has been upstreamed, but most of the critical components are, and everything else is either in progress or at least has an out-of-tree open source driver.

[0] Not that Broadcom wins any points here...

Thanks for that. I've been eyeing the FriendlyArm boards for a while. I've already got a NanoPi Neo2 (and a spare) because I wanted to run WLAN Pi[1 2].

Good to hear that the M4V2 lives up to its billing.

I also have a few Orange Pis. For me the sweet spot for them is in low power ethernet connected: the Orange Pi One. Maybe as a GPS-based NTP server, or some such.

[1] https://www.wlanpi.com/ [2] https://github.com/WLAN-Pi/wlanpi/releases

I got excited for a minute, and then I looked into the m4v2, and their software support just feels... hacky? Like, they support several flavors of Ubuntu, and a... (at the time they released it last fall) 2 year old version of Android?

I guess the question here is one I have with every non-RPi board: if the company behind NanoPi disappears tomorrow, can I continue to run a current version of a distro, forever, using a vanilla Linux kernel? If not... I'll pass, thanks.

While the RPi Foundation has their own fork of Debian and the Linux kernel, you can run stock Debian with a stock kernel on it just fine, and some of the missing drivers are being upstreamed all the time. The community has built images based on various other distros as well.

I recently went through the difficult process of taking an old (commercial) NAS that was running Debian squeeze and an ancient 2.6.31 kernel, and upgrading it to Debian buster and a 5.6 kernel. I was incredibly lucky that a hobbyist had upstreamed support for its hardware a bunch of years ago. If it hadn't been for that, I would have been stuck. I expect that sort of thing from a commercial appliance, but for a board I'm buying for its hackability, I expect to not be tied to the company that makes it for software updates.

https://www.armbian.com/nanopi-m4-v2/

Oh, this is truly excellent, thank you.

Is there an ARM based board in a ITX/ATX form factor with a few PCIe slots? Because that's what I'm looking for.

This PCIe on the Raspberry Pi is a neat hack, but still a way of from being a real desktop.

Only if you have deep pockets.

SolidRun MACCHIATObin has one PCIe slot. Enough for a GPU! I have this, it's pretty good. Though I have to resort to USB-Ethernet since FreeBSD doesn't support the onboard NIC (and the driver isn't the easiest to port) :D

SolidRun HoneyComb LX2K has a bigger full size slot and an M.2 slot. I don't think bifurcation works but you do have two slots, you could run a GPU and a NIC or HBA at the same time, using an M.2 to PCIe-slot adapter. And/or use extension boards with PLX switches if you need more cards?

$750 is perhaps a bit more than I was willing to pay, but that is pretty interesting.

Not ITX or ATX, but the RockPro64 has a single PCIex4 slot. $60 for 2GB model or $80 for the 4 GB model right now.

https://store.pine64.org/product/rockpro64-2gb-single-board-...

This is at a completely different budget, but I used an Nvidia jetson TX2 dev kit for a while and it's reasonably responsive and has a PCIe slot. I think the dev kit carrier board is also ITX-shaped, but am not too sure.

I want a plug and play hardware interface that can handle digital and analog io in a hackable package, capable of running cpp (easy to compile for) and if possible python.

If there are convenient alternatives to rpi, especially if they're cheaper for comparable performance, I'm all ears!

Pine64 has a bunch of well-supported SBCs, some of which (Rockpro64) include a PCIe port, for instance. I think beagleboard was in the business even before the pi foundation, but I don't think they care about mainlining as much as the pine64 community.

I tend to agree, but I think part of the problem is that there are so many other options out there. Each individual alternative option gets little attention by the market and it is really difficult to tell which ones are stable enough to be usable.

My problem is I've had a huge problem with getting tensorflow running on anything but a raspberry pi :(

The eco system around the pi is really amazing and I feel like its more alive. Which makes sense since its a learning platform.

From my experience competitors often implement hardware features but their support matrix for software is not quite the same as the RPis. Most of the time you are stuck on the maintainer modified distribution and their own patched kernels and if you want to go beyond that you are on your own. The ARM world is not as standardized as x86 and AMD64.

Last I checked, a problem with the other boards is that they aren't as open.

I've looked at a few of the other Pi-like SBCs out there that have beefier hardware features, and the one question I can't find an answer to is: how well is it supported, and what happens if the company that makes it disappears?

With the RPi, I know that enough stuff is upstreamed that even if the RPi Foundation were to dissolve, I could still run a stock Debian with a stock kernel on it and get most of the functionality working (where in my case, "most" is everything that I personally have a need for). While the RPi isn't 100% open hardware-wise (stupid Broadcom), the stuff that is open is good enough for me.

But I don't know how to get similar confidence about NanoPi, RockPi, OrangePi, CubieBoard, etc. How open is the hardware? Is support for everything I need upstreamed? What's the community like? A cursory look at their various websites didn't give me enough information to feel comfortable trying one out. I have a limited amount of time for this stuff, and I'd like to have a pretty high level of confidence before buying one.

I'll admit that my hesitance is maybe a little unfair and colored by a single bad experience: a few years ago I bought a ReSpeaker (the original v1 hardware). It was, frankly, garbage. The hardware itself was designed poorly and was buggy (attempting to run anything off the SD card would result in random no-reason SIGILL/SIGBUS/SIGSEGV, for one thing; that's a sign of broken hardware design). Support was minimal[0], and after they started working on the v2 hardware, they more or less stopped caring about v1. I actually threw the board away because it was useless. I wasted a good 20+ hours trying to get the ReSpeaker to do something useful, and I'm never getting that time back. I don't want a repeat.

In contrast, the RPi has always been a reliable platform to build on, with a great, responsive community.

[0] And the only suggestion was "only run things off the built-in flash"... which, IIRC, was double-digit-megabyte tiny.

No doubt that RPi has been a game changer, but I can never convince myself to build a light file/media server with a Pi. Do I really want to spend $100+ buying a Pi and accessories, or should I just spend $300 on a Synology... which actually has an NVMe slot.

Or, like, repurpose your previous desktop into a server when upgrading the desktop.

Even old laptops can make good servers. Built-in UPS and screen for recovery. Depending on just how old, you might even be able to add more storage using USB3 ports, or with a controller in an ExpressCard slot or in place of the wireless card, or by replacing an optical drive.

This! I have a highly underpowered laptop (which I bought brand new but at a very low cost for travel purposes) that I just didn't know what to do, because it was so damn annoying and awkward for any personal use. The keyboard was quite awkward. The trackpad was even worse. The display was so basic that even I didn't want to use it for YouTube/Netflix. The CPU... well you get the point. Finally, just a few weeks ago, I thought that I would really like to build a Linux server and thought of my Raspberry Pi 3. I remembered that I would need to connect an external HDD to make it even halfway useful for data storage. I would also need to find a USB wifi adapter (because my home setup currently isn't wired). Then I looked at this laptop which had been collecting dust for like 2 years. I realized it was perfect for my needs: has a 1 TBB HDD (not SSD) and a built-in wifi adapter. I installed Lubuntu (one of the least resource intensive distros) and then started installing numerous services on it: Nextcloud, IRC bouncer, SMB server. I only have Nextcloud running at the moment as it's the only thing I've been able to make serious use out of, but it just works.. perfectly.

Power usage.

Hit the nail right on the head. That and size.

My home server is an $80 laptop I bought from a dude on craigslist with a 2nd gen i5 and it's great.

A light file/media server definitely doesn't need NVMe.

I agree, but I was stressing it's there if I choose to recycle an old disk.

If you're to the point of having spare NVMe drives, I suppose, but USB adapters are $20-25 and they'll perform fine for that use.

Edge servers need NVMe.

Not light ones. They'll do fine despite the extra microseconds.

The Tindie store seems to be out of stock. But, if you need one of these, the GitHub repo for the project includes an Eagle brd file.

Note that most PCB houses want Gerber files. PCBWay has an online converter for brd files. It does require creating a userid on their site, but it will give you the Gerber file without buying anything...so you're free the buy the PCB from anywhere you want.

Great work, this opens new flood gates for Raspberry Pi community and I'm glad that the maker made it available on Tindie.

Off topic: When I heard about Apple silicon on iMac, I was wondering whether it will retain PCIe or will Apple will 'courageously' kill PCIe on iMac (/s).

Very clever to use the same pinout as the PCIe "flexible risers" that already use USB cables, so only the chip replacement is necessary.

I long for the day we can buy proper ATX-sized arm boards. This is nice and all but arm can do way more, and without all that historic 16/32 bit mode clutter x86 has for compatibility.

it has to be coming soon, apple is going to be making arm macs and aws has m6a instance types running 64bit arm right now.

You can buy mini-ITX sized ones from SolidRun. They even run open source firmware (EDK2) that exposes most things as generic ACPI hardware.. well there's still quirks and hardware bugs of course :D

Is edk2 something that should be run in production? I thought it was more of development platform for UEFI.

Well, commercial vendors usually use AMI's modified version that has a proprietary fancy setup UI and whatever other crap. But for hobbyist custom firmware (these ARM systems, Chromebooks, etc) completely raw upstream EDK2 is totally fine.

Fun fact, System76 now runs slightly modified EDK2 (as a Coreboot payload) with a custom setup UI written in Rust: https://github.com/system76/firmware-setup

Discussion from the original PCIe hack: https://news.ycombinator.com/item?id=20399555

If you can plug a keyboard into the power socket, I'd love a Raspberry Pi mod that put an NVMe drive on the PCI bus. _That_ would be a useful little computer.

Doesn't the chip-ectomy also lose Ethernet?

Edit: apparently not, but how is the Ethernet connected on the pi4? Is it on the BCM SoC?

Yeah, the Pi 4 finally has an onboard Ethernet controller on the SoC itself.

This says the Ethernet controller is "natively attached" but the SoC datasheet doesn't mention Ethernet: https://www.raspberrypi.org/documentation/hardware /raspberrypi/bcm2711/README.md

So maybe it's a separate chip?

Maybe. The chip closest to the Ethernet port has BCM on it..

You have to search for RGMII in the bcm2711 datasheet, which is a media-independent link.

In theory you could rip the PHY chip out and replace it with a fibre PHY, so you'd get a fibre Pi.

There's a couple Raspi clones with NVMe, with the most expensive being the Nvidia Jetson Xavier NX carrier board.

The jetson nano doesn't have NVMe

Comment updated. The Jetson Nano dev kit does have an key E m.2 slot, which could be used with a standard key M SSD using an adapter cable, but I don't think anyone who doesn't already own a Nano would do that.

I find this fascinating, and I have a few follow-up ideas.

How many PCIe lanes are on this connector? Is this a full 16x slot? I assume since it has a very limited amount of PINs I'd assume it's just a 1x.

The next obvious step is hooking up this to a PLX chip. Duplicating the number of lanes available even if there's no additional bandwidth could really open up doors.

From the article it's 'just' a 1x PCI-E 2.0 lane (theoretical 5Gbps throughput). This is all the VL805 expects so all that's broken out to it's pads and 'easily' accessible.

It is 1x, obviously. But I highly doubt that the PCIe bus is going to be the bottleneck here ;-)

I guess it would be possible to design an ATX size motherboard with a RPi4 compute module slot (once CM4 becomes available). The PCIe bus would allow all sorts of peripherals. Don't expect any performance though.

Have you seen https://turingpi.com ? I can't imagine an ATX/ITX/STX compatible breakout board for everything on the Pi 4 would be an impossible task.

Well, yes, something like that, but not in a cluster form, but just a single CM4 on a mini-itx form factor board, with PCIe slots, USB ports, ethernet, etc.

You could build a fully blown Linux desktop peripheral wise, just don't expect desktop performance.

It's 1 lane, but a 16x connector to allow plugging in any device. PCIe devices can be downgraded into a lower number of lanes. Many motherboard 16x connectors aren't even a full 16x.

> The next obvious step is hooking up this to a PLX chip. Duplicating the number of lanes available even if there's no additional bandwidth could really open up doors.

I did this with an external PCIe 'splitter'/switch board: https://twitter.com/q3k/status/1225560443927891971

It worked, kinda. After a kernel rebuild with the drivers I wanted and dtb bumps to allow for more PCIe MEMBAR space i was able to get multiple devices (IIRC 10GbE and SATA controller) to work simulataneously, just very slowly.

> bumps to allow for more PCIe MEMBAR space

huh! huh! can you have enough BAR space for an AMD GPU to run?

Rockchip RK3399 is infamous for having too little for GPUs..

One of the biggest limitations of the RaspberryPi is that it’s only got one network port, so that rules out running it as an OpenWRT router for practical reasons. Does anyone know of a similar miniature board with multiple network ports that could be used as an OpenWRT router? I’d like to replace my crumbling router with a low power one that is more configurable than my old one.

I'd love a computer like the Raspberry Pi -- but one where every chip on the system board is "plug-and-play", that is, all of the chips have pins, and all are pushed into sockets such that any one of them could easily be pulled and replaced by another chip that matches the pinout for the given chip, with no desoldering or re-soldering...

Not just the adventure for some. We use the RPI3/4 as a plug-in module to my clients range of industrial controllers because we want reliable availability, forward/backward compatibility and the most important recently is we want to diversify our supply chain from China it fits the bill nicely on all three.

so one can connect a pcie sata adapter and connect sata disks even if on the same x1 lane?

Given it is x1 PCIE 2.0 you are likely to see little performance benefit vs using a USB solution for connecting sata drives while introducing potential stability issues.

Reply on news.ycombinator.com