Nvidia's 900 tons of GPU muscle bulks up server market, slims down wallets

It's not as much of an issue about supporting old versions, it's just that new versions of Linux breaks the drivers since they don't know what's in it.

Correction: unwilling, not unable

It depends on the model architecture you are using. Once you cannot fit a single instance of your model on a single GPU or at minimum on a single node, things start becoming very complicated. If you are lucky and you have a generic transformer model, you can just use Deepspeed with their transformer kernel. But if you have another architecture it will likely not be compatible with Deepspeed or Fairscale or any of the other scaling frameworks and you will end up having to write your own CUDA kernels.

So the per GPGPU RAM is quite important.

But a lot of people (including me) run 3090/4090 rigs at home, or rent them at runpod/vastai to save a buck.

It depends on the workload, but generally for large neural network model tasks, no, while for most other accelerated tasks, yes.

> Is there really no way to partition the workload to run with 16gb memory per card?

With regards to the neural networks in the news, you can partition the workloads. The 8xH100 machines are also partitioned in the same way. Nonetheless the performance is very poor on the consumer cards, because they lack the very high speed interconnect of the H100 that makes partitioning performant. Additionally, the size of the RAM on the H100 is matched to the capabilities of the chip itself - it sits in at the peak of an upside down U of a performance curve - which is why it is not possible to "just" add more RAM to these chips.

In a recent paper I was reading they were bragging they "only" needed 8 x A100 to train their model.

Renting a 4090 24GB costs around $.40/hr. Renting A100 80GB costs around $3.4/hr. For 3x memory and 2x bandwidth.

To fit a 13B model in a 4090 you need to quantize it. Imagine a 130B or 200B model. And never mind training, which needs a lot more memory.

The divide is crazy. If you are not FAANG you are "GPU-poor".

It really depends and this can get really complicated really fast. I'll give a tldr and then a longer explanation.

TLDR:

Yes, you can easily split networks up. If your main bottleneck is batch size (i.e. training) then there aren't huge differences in spreading across multiple GPUs assuming you have good interconnects (GPU direct is supported). If you're running inference and the model fits on the card you're probably fine too unless you need to do things like fancy inference batching (i.e. you have LOTS of users)

Longer version:

You can always split things up. If we think about networks we recognize some nice properties about how they operate as mathematical groups. Non-residual networks are compositional, meaning each layer can be treated as a sub network (every residual block can be treated this way too). Additionally, we may have associative and distributive properties depending on the architecture (some even have commutative!). So we can use these same rules to break apart networks in many different ways. There are often performance hits for doing this though, as it practically requires you touching the disk more often but in some more rare cases (at least to me, let me know if you know more) they can help.

I mentioned the batching above and this can get kinda complicated. There are actually performance differences when you batch in groups of data (i.e. across GPUs) compared to batching on a single accelerator. This difference isn't talked about a lot. But it is going to come down to how often your algorithm depends on batching and what operations are used, such as batch norm. The batch norm is calculated across the GPU's batch, not the distributed batch (unless you introduce blocking). This is because your gradients AND inference are going to be computed differently. In DDP your whole network is cloned across cards so you basically run inference on multiple networks and then do an all reduce on the loss then calculate the gradient and then recopy the weights to all cards. There is even a bigger difference when you use lazy regularization (don't compute gradients for n-minibatches). GANs are notorious for using this and personally I've seen large benefits to distributed training for these. GANs usually have small batch sizes and aren't getting anywhere near the memory of the card anyways (GANs are typically unstable so large batch sizes can harm them), but also pay attention to this when evaluating papers (of course as well as how much hyper-parameter tuning has been done. This is always tricky when comparing works, especially between academia and big labs. You can easily be fooled by which is a better model. Evaluating models is way tougher than people give credit to and especially in the modern era of LLMs. I could rant a lot about just this alone). Basically in short, we can think of this as an ensembling method, except our models are actually identical (you could parallel reduce lazily too and that will create some periodic divergence between your models but that's not important for conceptually understanding, just worth noting).

There is are also techniques to split a single model up called model sharding and checkpointing. Model sharding is where you split a single model across multiple GPUs. You're taking advantage of the compositional property of networks, meaning that as long as there isn't a residual layer between your split location you can actually treat one network as a series of smaller networks. This has obvious drawbacks as you need to feed one into another and so the operations have to be synchronous, but sometimes this isn't too bad. Checkpointing is very similar but you're just doing the same thing on the same GPU. Your hit here is in I/O, but may or may not be too bad with GPU Direct and highly depends on your model size (were you splitting because batch size or because model size?).

This is all still pretty high level but if you want to dig into it more META developed a toolkit called fairseq that will do a lot of this for you and they optimized it

https://engineering.fb.com/2021/07/15/open-source/fsdp/

https://github.com/facebookresearch/fairseq

TLDR: really depends on your use case, but it is a good question. I/O is an asshole