Hacker News Reader

Comment by coffeebeqn | Hacker News Reader

I can’t imagine the SOCs are too far away for PCs. It does go against the upgradeability philosophy but I’m not sure how many users really care about that

View on HN

Is this different from integerated GPUs from Intel and AMD?

As exemplified in my post above, AMD integrated GPUs don't have unified memory, at least my 5000 series I bought in 2022. Maybe the new 7000 series changes this.

They have had it for about 10 years, since Kaveri. See eg https://hexus.net/tech/news/cpu/54709-amds-heterogeneous-uni... or https://www.guru3d.com/articles-pages/amd-a10-7800-kaveri-ap...

Not exactly. While that may have been the case way back in the past when AMD APUs were designed as a homonegnous CPU-GPU unit from the start, I can tell you that's not the case with the relatively modern 5000 series.

I need to go in the Bios and specify explicitly how much of the system RAM I want allocated exclusively of the integrated GPU and the rest stays available for system RAM.

The reason is because Ryzen 5000 APUs seem to be a job rushed out the door so they're just a Zen 3 CPU with a separate Vega GPU glued together on the same die, but they're not a homogenous design, designed to work as one unit, like the APU of the PS5, so memory wise they're unaware of each other, even though AMD calls them APUs, they're not really, but more like separate CPU and GPU on the same die.

I wish I knew about this limitation at the time, as Intel chips with integrated graphics have unified memory.

I wonder if a clever enough driver+firmware could make that hot-adjustable (or truly unified, rather) with nothing but a software patch. It sounds like the pieces are all there.

My understanding is that the Steam Deck, which uses a Zen 2 APU, dynamically allocates system RAM to VRAM. It only works in SteamOS (Arch Linux) though. Folk that run Windows on their Steam Decks need to go into the BIOS and bump the static mapping to 4GB in order to get good performance in many games.

"Unified" doesn't mean the CPU and GPU can store different data in the same memory at the same time. The BIOS setting is just assigning how much are visible for each to use in the absence of OS support for doing it dynamically. Clearly if you can change the amount it isn't a hardware limitation, and I'm not even sure if it needs some special support. The iGPU surely supports DMA, so if it does a load from memory "assigned" to the CPU does it perform any differently than from memory "assigned" to the GPU?

Sorry but this sounds like jumping to conclusions, going off only the fact that you can fence off some memory in the BIOS. See eg this windows gaming site review where they couldn't find a difference in the performance between allocating 64MB or 2GB save for satisfying some games that need to detect a certain amount of GPU memory to start: https://www.techspot.com/article/1578-amd-raven-ridge-reserv...

But I don't doubt there still exist software (Firmware, OS, driver, graphics API) problems where unified memory ends up being used badly. There doesn't seem to be evidence that much if any of it are because of hw limitations.

If you put the memory on the same package as the CPU it can have higher bandwidth without needing an unreasonably high number of pins in the CPU socket.

But there is no reason you can't do both. Intel is already by putting HBM on the Xeon Max. It still has separate memory slots, because they're for two different things. HBM is fast and expensive. DDR costs less so you can have more of it for the same amount of money and the large majority of workloads aren't memory-bound.

You could then have systems with DIMM slots that aren't populated because the CPU has enough HBM to satisfy you, others that cost less with no HBM and the same amount of DDR, and still others that have say 16GB of HBM and 128GB of DDR. Both of which are still upgradable, but if you want more HBM it comes with a new CPU.

Who really upgrades pcs anymore. By the time you need to upgrade everything from socket to power standard is outdated

A mid-lifetime upgrade where you pop in a new GPU, fill up any unpopulated ram slots, or add in a new drive, can be nice. CPUs last a really long time nowadays.

So do gpus though. Are you really going to pop in a 800$ gpu into 200$ worth of components?

Depends how good the original was if it’s worthwhile. GPUs don’t need that much from the rest of the machine as long as it’s somewhat sensible for gaming. Especially at 1440p or 4k the GPU is a huge portion of the performance

> Are you really going to pop in a 800$ gpu into 200$ worth of components?

Why not? For the right workloads, the GPU is the important thing anyways. It's not that different from rotating disks through a machine.

Right. At which point, why pay the cost of the rest of the PC being modular? Both the $800 and the $200 would be smaller amounts in the case of an integrated system.

You don’t have the option to pay to upgrade only one of the components in the integrated system…

Sure. But if the entire integrated system can cost a similar amount to the most expensive component you might upgrade in a modular system, there’s little practical difference.

My thought as well. The system is basically integrated anyway because all the components have to be matched at the current gen

My current PC is using an old 4th gen i7 with an RTX 3060Ti and it keeps trucking along just fine, even with my 144Hz display. It used to have a GTX 960 but its 2GB of VRAM were anemic for some ML stuff I wanted to run + the newer Nvenc is a massive upgrade in quality. I don't see any reason to upgrade anything else.

I would hope that this is more common than GP thinks, and this is also something I look for when I'm buying laptops- if they don't have an okay amount of extra PCIe lanes laying around, I don't want it- because they can become really fun projects (ie strap a GPU to it and get a few more years of Minecraft out of it, or use it as a network appliance with a built-in UPS)

That’s a cool idea. How do you get to the PCIe lanes?

I’ve been hoping thunderbolt would make GPU enclosures more of a thing, but they seem to still be bottlenecked by even that pretty good interface (last I checked, at least).

In the past, the thing to look for was laptops with multiple mini PCIe connectors (generally for WiFi cards and sometimes cellular) but now the thing to look for is M.2 slots with PCIe on them. Generally the easiest thing to do is to yank out the wifi card and add your own thing there, and then for networking you can just use either onboard or a decent type-c gigabit dongle. Some laptops have spots for multiple SSD's and you can just use one of those instead.

Keep in mind that even a basic PCIe card can pull up to 75 watts and that you'll probably need some sort of external power supply. There are two main routes you can go- use a second external laptop power brick, or what I did- slap an entire computer power supply in there.

I bought an SFX power supply from EVGA, who has an awesome B-stock program where you can get returned / refurbished stuff for cheaper than normal and it still comes with a warranty. If you do decide to use a computer PSU, I'd highly recommend one that's modular, because you won't need a bunch of the wires and they take up a lot of space quickly.

To tell the power supply to turn on, you can either use the 'paper clip trick' to jumper it on, or get an "ADD2PSU" board (basically a relay on a PCB) and use a USB cable to power the relay and you should be good to go.

If you use a laptop power supply, you'll either need something that's 12V or get step down boards or a Pico PSU or something like that.

The thing to search up is "M.2 NGFF to PCIe" or something like that. I'd recommend getting ones with some kind of cable and an open back of slot design so that you don't wind up needing two adapters for one thing.

Adapters and cables are available from the usual suspects- Amazon, eBay, and Aliexpress will all have what you want.

This is a lot of information, but this is legitimately pretty easy to do as long as you make sure you check all the boxes- PCI lanes, external power, and something to keep the new stuff you're strapping on safe.

Certainly not the case with AMD AM4. Stats show that plenty of people have upgraded their AM4 builds.

The Intel only crowd is such a sad bunch

I'd disagree if AM5 follows the path of AM4.

AMD did pretty well with the AM4 socket. At least 3 CPU generations on the same socket.

Amd socket AM4 supports 5 generations of Ryzen processors. AM5 came out with 7000 series and it will be supported by 8000 series as well. Upgrades are easy and give a significant performance boost.

Sockets yes but you don’t support the faster ram it needs so what’s the use

Power standards aren't changing that much. Yes, there's 12VHPWR now, but that's only on the top end gpus, and adapters exist. ATX 12VO might become a thing, but feels like OEM standardization.

Before that, the last big change was Haswell low power states ten years ago. If your old PSU didn't meet that standard, you would probably be ok if you just disable the lower power sleep states.

AM4 has been great for this. Upgraded from a 3700x to a 5950x last year, then from 32GiB of RAM to 64GiB this year.

Reply on news.ycombinator.com