Avantek's Arm Workstation: Ampere EMAG 8180 32-Core Arm64 Review
anandtech.com
anandtech.com
I’ve been looking into Nvidia’s Xavier NX of late but, of course, it’s not an SBSA-compliant system, leaving the user totally at the mercy of the producer (Nvidia, in this case, as is indeed tangentially mentioned by the article itself).
Whenever I mention this people usually throw in irrelevant statements such as picking up a Raspberry Pi 4 (of which I have a couple) or another SBC, but really I’m looking for something which offers PCIe and all the creature comforts (even the HoneyComb LX2K lacks PCIe beyond a single M.2 slot for the SSD and offers no wireless capabilities unless you add a USB3 dongle).
Is there a reason why somebody hasn’t (oh, I don’t know) thrown a Snapdragon 8cx on a board with some of these accoutrements and provided it in a mini-ATX form-factor?
Because there is no profitable market for it. The number of people who would be willing to pay the $1000+ that it would cost to produce a board like that in low volume just aren't there, in part since you could build a dramatically more powerful x86 machine for less money.
If the phone/networking chip were designed upfront to to be less cost optimized for its target market (say putting graphics on a networking chip, or more PCIe lanes on a phone soc) that would help. As would assuming your going to sell 100k of them, and assuring the design and pricing is competitive with a $300 x86 motherboard+cpu combination. The honeycomb could be a good board, but its obvious they have designed it for some target market that isn't actually a generic mITX board. The CEX, and x4 SFP cages add cost, but little value vs a desktop board so you have a board that costs $800 and competes poorly with $150 AMD boards that take AM4 processors unless you happen to need 4 ports of 10G ethernet. The honeycomb minus all the networking junk would be selling in the hundreds of thousands of units if it were $100 to $250. The rpi doesn't sell millions of units because its good, its sells that many because its $50.
So really the "flaw" is all the extreme upfront cost cutting to create a product that is laser focused at a single big fish customer, then failing to provide a generic and flexible version for the open market that can be customized less expensively.
For various reasons it’s something I want. Do I need it? No, I don’t... but then again none of us really need anything beyond oxygen, hydration, and caloric sludge.
Let’s put it this way: given my predicament (locked down in a building that only has WiFi) a wireless solution is pretty much required.
As for the PCIe... the 8x slot clearly is required for a graphics card. I have several other peripherals (an SDR and an FPGA card) that would be homeless.
AKA, if the board is missing something fundamental (like graphics, sata, or even USB!) then the board is useless for a lot of cases which require both the missing bits + some other PCIe board. Its the same thing with soldered on RAM, 4G should be enough for everyone!
Its quite silly in many regards too, since adding a few $$ to the BOM for a graphics chip/usb/networking/whatever to bring it up to feature parity with $120 boards from asus/etc never seems to cross their minds.
It's not always the BOM that costs you, instead it's the additional testing, validation & support.
There isn't any way, that every chip+motherboard+ram combination possible on the x86 market is tested. What you test for is compliance to the spec, and a few really common cases to sanity check everything.
Then the OEM/final integrator is on the hook for validating/supporting a design. And it works the majority of the time you can buy RAM, or pcie cards, or whatever and plug them in. That is the point of standards. Plus, on the bleeding edge the standards almost always have fallback modes which are considered easier to get right (lower clocks, well understood previous generation, whatever). So when something isn't fully conformant it still works, just at a degraded speed.
The problem is most of the arm vendors can't built compliant parts to save their lives. From nonstandard usb, pci, ram and networking gear. It only tends to work because they hack it together in that one configuration that happens to work because its got a couple capacitors sprinked on the board, or a software quirk that bypasses the broken functionality. Just look at all the stuff wrong with the rpi.
Adding an actual system OEM/integrator means another person needs convincing there is a market and adds additional markup.
The struggle to put together functional boards is just indicative of the complexity of the problem. Even some x86 boards are garbage.
None of those options really work for the majority of the arm SBC market which are repurposed phone SOCs. Yes, they are flexible as long as a large part of the target is touchscreen + wifi and attaching a USB SDR. Or they get misused as massively overspeced micro controllers that happen to need a network connection (aka IoT).
So its really that latter case where they are successful because they are "generic" microcontrollers. Many of the uses cases would be far better/cheaper handled with a cortex-M and a custom board. But using a $35 rpi to read a GPIO and send a packet is the "generic" solution for someone not willing to spend $500k on a custom design to create 1k devices at $2 each.
The microcontroller and kiosk markets are actually relatively large, and a good niche to be in. Adding the complexity needed to make a full-blown x86 competitor is extremely risky.
There are integrators out there using ARM in networking and NAS gear, but they are building the board to their spec, validating it against a small sampling of add-in hardware, screwing the box shut with a "warranty void" sticker on it, and charging a healthy markup for the service.
And just to extend the "whitebox" market is the largest slice of the x86 market. But even so, I just saw a number the other day that the desktop market is ~12B in revenue for intel. So, its might be small for intel, but that would be a huge segment for many of companies in the arm ecosystem. Think of what would happen if a company like rockchip managed to acquire 20% of that market.
PS: I'm aware of the vendors building ARM NAS's, security appliances, and such, and if you look at their product lines, overwhelmingly they are x86. The arm parts form a small fraction of their designs. Frequently the lowest end products with the highest volume and tightest margins. The higher end products are just boring old pentium/xeon parts with fat margins.
ex:
https://www.synology.com/en-us/products/series/smb vs https://www.synology.com/en-us/products/series/home (and only a portion of those are arm)
PPS: I think a lot of people are really confused about the "desktop" market. The desktop market is an artificial segmentation of a generic technology stack. Intel puts it in a different bucket for business reporting reasons, but the core technology is the same between a random desktop board they make $ on and a random board like those in the synology appliances above. The largest differentiator tends to be whether or not the part supports ECC (and some of the desktop parts come with that feature) and who supports it. So from a technology perspective its the same fundamental thing in a 2U chassis with 16 spinning drives as it is in the tower under a desk. Which is why in the parent article you see a 32 core "Server" being sold as a workstation. It could just as well be sold as a NAS. The problem here, is like those arm NAS's the arm vendors are only shooting for the high volume markets, which naturally happen to also have the lowest margins because they are the most competitive... Or the ones with the highest margins. See the problem? AMD unlike intel doesn't do as much artificial product differentiation (which is why a lot of people like them). Their server class parts are basically the same as the desktop, they differ only in scale. So they make the margins in the areas that are truly low volume/high technology while scooping up volume with thin margins elsewhere.
You got that response because your post sounded like the LX2K only has the M.2 slot and no other PCIe at all, when it in fact has a regular ×8 PCIe slot.
> the 8x slot clearly is required for a graphics card.
The HoneyComb LX2K is not a desktop or mITX replacement. It's a network/routing platform, as shown by the onboard 100GE QSFP28 port. The ×8 slot is clearly required for more networking :P
> I have several other peripherals (an SDR and an FPGA card) that would be homeless.
... where are you putting those on an x86 Mini-ITX board with one PCIe slot?
e.g. (21 TOPS) Xavier NX is for devices "constrained by size, weight, and power budgets."
The honeycomb (or even the mcbin) are better choices if you just want a PCIe slot.
Regarding the boot times: If you watch the serial console when the machine is booting, you see that its running a native arm boot loader (uboot?), and then that boot loader boots the EFI firmware. So I think the "double bootloader" is a large part of the long boot times.
The process is fast on desktop machines. Big servers spend minutes training and testing all the damn RAM.
In my case, I was booting FreeBSD, and I don't remember having boot much slower on this box than it does on Intel Xeon or AMD Rome once the EFI bios finally handed off to the FreeBSD efi boot loader.
Last time I booted a serious server, the OS came up in about a minute or so. The application took something close to two hours to be fully primed and within performance guidelines. Years ago a RabbitMQ decided to take 9 hours to fully restart. We had to add queue server redundancy to the application more quickly than I would advise anyone to try.
You're not going to be making a very appealing workstation by just throwing old mobile focused ARM licensed IP together (or by something that looks and performs a lot like it). Shame the only party thinking in the direction of workstation like performance for ARM is Apple. Then again, making good workstation chips requires a lot of money. Guess they have to have a market first.
The new Ampere "Altra" chip is going to be like an Amazon Graviton2, but higher clocked, so that's going to be really fast.
The latter (https://www.anandtech.com/show/12571/gigabyte-thunderxstatio...) is a pretty serious piece of developer hardware, with much better single thread perf, and 64 cores. But its also quite spendy, and you have to be a really serious arm developer to be able to justify one.
I'm not quite sure what you mean by that..
You can't test SVE (Scalable Vector Extensions) using programs on an eMAG or any other processor currently available other than the A64FX. Regular software, yes, it's all AArch64.
If the company put 20 or so online people could log into with 1 hour sessions or something, that'd be pretty useful to see if it's a good fit
I was able to login, install rust via rustup, install htop, build ripgrep showing nearly all of the cores being used, use ripgrep and kill the instance in about the same time on large x86 box.
Well, I was just curious and not in the market for a new workstation. I would appreciate a hint for a fan-less MB (Arm or otherwise) with (3.3V) PCI sockets ...