Hacker News Reader

Comment by karencarits | Hacker News Reader

The top comment says a lot

> It's pretty much the rumored size of GPT-4. However, even when quantized to 4bits, one would need ~800GB of VRAM to run it.

View on HN

It's a sparse model, which means you only need to load the "experts" that you're going to use for a given input.

But you still need to page in the weights from disk to the GPU at each layer, right?

You should only need the weights for the experts you want to run. The experts clock in at around 400 MB each (based on the 800 GB figure given elsewhere). A 24 GB GPU could fit around 60 experts, so it might be usable with a couple of old M40s.

You can get a second hand RTX 3090 with $800. So in theory you could build a cluster with less than $40k for all hardware. Not that this is pocket change, but given the capabilities of such a model the expense doesn’t sound that much.

Or one can try to rent a cluster in the clouds.

you don't have the nvlink, how would this work then

You don’t need the entire VRAM to be accessed as one. Everything is parallelized by the framework. You lose some speed, but it’s the only way to do it. There’s no freaking way to have 800 GB of unified VRAM.

GH200 would like a word ;-)

Hahaha I love this.

As I understand it, NVLink isn't actually required for much of anything. It allows P2P transfers and atomics. That same functionality can be used over PCIe on some GPUs with reduced performance. And the P2P features are not transparent and are lower-performance than staying inside one GPU, so generally software avoids P2P transfers even when they are available.

The Nvidia documentation on this is absolutely horrible, the features vary wildly by model of GPU, and I claim no real expertise.

3090's probably aren't gonna cut it because of the generational improvements in architecture (even if you could hack the drivers together to work)

It's no big deal to run an EPYC host with 2TB of ram. It'll run slow. But the Falcon 180B model runs at several tokens per second that way and its nowhere near as sparse as this should be.

I also have 960GB of vram in my garage (40x P40), though I suspect getting this model distributed in a useful way might be annoying and not worth the effort particularly since it would probably be close to turn-key (if a bit slow) on a high ram epyc host.

be careful, the Compute Enforcement Administration might be reading

How many motherboards did you need to supply PCIe lanes to all of those P40s?

I have a similarly-sized pile of MI25s and recently managed to get a eight of them running in a single plentiful-on-ebay supermicro motherboard (custom bifurcating fanout-riser, x8 to each card). It was a "learning experience".

Falcon 180B, 4-bit quantization fits in VRAM and works, but at under 10tokens/sec because llama.cpp's support for multi-GPU on AMD is kind of an afterthought.

I have them across 4 hosts with pci-e switches. Really only 9 per host are maximally useful due to nvidia mailbox limits in cuda (also because one per host is on 8x, IIRC).

> I have them across 4 hosts with pci-e switches.

Nice!

Sorry to pester you, last question: did you find a source for those at a reasonable price?

Pricing from last time I investigated that would've been more than the P40s, so I'm assuming you did. Apparently the actual ICs are/were incredibly expensive and only Pericom sells them as standalone chips AFAICT. (Edit: oh wow, Pericom got bought by Diodes Inc; this explains a lot)

Thanks!

Wow... I won't be running anything like that locally anytime soon unfortunately...

Reply on news.ycombinator.com