Risk mitigation, maintenance, resell value (?), support, reliability. The custom AI accelerators you mentioned, how long have those been in business for? How many units have they moved? How many generations of hardware have they produced? Will they still be around to replace or upgrade units in 5-10 years? How flexible are they in their workload?

That's a lot of factors to keep in mind when you're spending millions on a supercomputer. I'd go for an established hardware provider as well. I'm not knocking the custom AI chips, but I wouldn't try and max out my budget with those - keep them for smaller applications for now.

Your supercomputer won’t last more than a couple years, 5 at max. Then it will simply be unjustifiably expensive to run when newer, more power efficient hardware becomes available.

What matters is the software platform and compatibility: you don’t want to retool and port all your existing programs. That’s where having a platform - CUDA - becomes a deal-maker.

What I don’t understand is how consumers - big, institutional consumers- don’t insist on viable vendor neutral platforms like OpenCL. Obviously they’re the only ones that have an interest in cross-vendor compatibility.

Not sure why you're downvoted, HPC systems don't last more than ~6 years before they are decommissioned because they've become too power-hungry to justify their continued use. I just checked some of the systems I'm familiar with and they were all decommissioned after around 6 years of use (e.g. SuperMUC, 2012-2018; two smaller cluster at the university where I work: 08/2012-04/2017, 01/2014-03/2020).

I also attended a presentation at IPDPS 2018 by someone from one of the big US national labs where they talked about how absolutely huge of an undertaking it was to port their codes to make efficient use of GPUs. You don't just re-do all of that if the payoff isn't enormous.

Out of curiosity, do you know what happens to the decomissioned hardware? is it scrapped for useful parts like perhaps the PSUs? or is gold and other metals extracted from the chips and the metal sold for recycling?

Is there anything useful someone could do with it or it's just too much of a problem to set it up and repurpose it?

> Out of curiosity, do you know what happens to the decomissioned hardware? is it scrapped for useful parts like perhaps the PSUs? or is gold and other metals extracted from the chips and the metal sold for recycling?

I cannot state with 100% clarity what happens after the systems are acquired by surplus vendors, but I can say that we have received certificates of what was "recycled". Once the surplus vendors take possession of the hardware, it's theirs and they can do what they want with it. In fact, we recently had to purchase EOL'ed, refurbished Infiniband switches to continue to support an interconnect fabric still in use (~2016). Interestingly enough, some of the switches still had "core" and "edge" labels on them.

> Is there anything useful someone could do with it or it's just too much of a problem to set it up and repurpose it?

In my opinion, it really depends on the node. If the chassis supports hot swappable & redundant hardware (HDD, PSU, etc.), then we'll typically cannibalize several chassis to create an administrative and/or infrastructure node. Case in point, a large portion of our older 12 core nodes have been put aside to serve as administrative nodes, all fully redundant (RAID1 HDD's, dual PSU's, ECC memory, etc.). Since we have a stack of these, we're fairly confident that these will serve us for the next few years, worry free, given the abundance of parts lying around.

Interesting! thanks for the insight.

CUDA is valuable and seen as a positive, even if it's vendor locked. It's battle tested. Engineers are familiar with it. Support is provided because it's all under one roof. Performance targets, roadmap timeframes, etc can be assured.

The same reason a big institutional consumer will happily pay Microsoft for Windows Enterprise, instead of demanding Microsoft rebuild Windows using open source technologies.

A supercomputer or DC cluster has a fixed lifespan and amortization period, if you can get something in a contract, with everything you want for a price you are happy with, why would you introduce risk to it by demanding open standards for no benefit?

I’m not saying CUDA is not valuable. Heck, proprietary has a ton of value... see how Intel owes much of its position to its compiler (the FORTRAN one in particular) suite.

Of course you don’t get to dominate a field to such an extent without some redeeming positive quality.

What I’m arguing - or more appropriately, what I’m wondering - is why on Earth aren’t consumer consortiums not insisting on open-standard platforms.

I get how you want to eek every drop of performance out of your tools without making it your only mission (that would be the research you’re using the tool for), but tying yourself to a single proprietary vendor defeats that purpose imho

TPUs are perfectly positioned to capture this market. There are an endless number of reasons why, but to keep it short: Wanna see a magic trick? https://twitter.com/theshawwn/status/1286426454171975680

GPT-2 117M training at 1 million tokens/sec.

Now, I don't have experience with DGX clusters, so I'm not going to make a firm statement. What I will say is that I, as an outsider, managed to achieve a performance level that is ~unheard of for GPT-2 training. And you can too; TPUs are pervasive.

A TPUv2-512 isn't even as far as the gas pedal goes, either. v3-512 can train all of imagenet to 75.9% accuracy in 4 minutes: https://twitter.com/theshawwn/status/1223395022814339073

v3-1024 can do it in 2 minutes: https://twitter.com/theshawwn/status/1234654848114520065

I once attached a debugger to a training run during startup, after the infeed loop began (meaning it was feeding inputs to the TPU, but no training was happening yet; it was "winding up") and was shocked to discover that when I hit c to continue, it trained on all of imagenet in like 54 seconds. That blows the lid off of every perf result here (under "image classification"): https://mlperf.org/training-results-0-6

(It's not a fair comparison, but it was quite astonishing to see the raw horsepower in action.)

So, nVidia has some catching up to do. And I don't know if they'll be able to. The TPU ecosystem may be clunky at the moment, but boy is it effective. Your options are to invest your time in this ecosystem, which will likely be around in ten years, or in DGX-cluster-type knowledge, which ... might be less pervasive in 10 years.

The distinguishing feature of a TPU is that it has a CPU on board. In fact, it has a CPU with 300GB of memory for every 8 cores. Friggin' love these things.

Last time I worked on TPUs, a lot of very pervasive LA operations (e.g. Cholesky decomposition) were unoptimized and slow compared to the NN-style operations. I’m sure that can be fixed, but for the time being, it seems inappropriate for anything beyond the obvious NN operations.

If "this market" means supercomputers (as 'Cthulhu_ was talking about), then TPUs are certainly not perfectly positioned to capture this market, since they are only capable of doing bfloat16 arithmetic with high performance, and AFAIK incapable of fp64 at all. That means a HPL score of zero.

TPUs will probably dominate AI training, but not so much supercomputing.

1 milion per second of anything isn't very fast. Unless you're running that on a mobile phone.

Many people use HPC for things that are not AI/ML, such as CFD. Those applications are typically written using MPI/CUDA

Do those custom AI accelerators work with PyTorch and Tensorflow? The graphcore website says that support with PyTorch aten is available in 2020, but it's not -- and Aten is inference only, no autograd, and no extensions. Without those, we can't train on it.

Graphcore claims they do. I don't know true that is though.

I reimplemented one benchmark using TF and the speed up wasn’t the 10x or so hey we’re claiming, but more like 1x..

The only non-NVidia device you can actually use that has broad support outside basic models in software is Google's TPUs.

Graphcore looks great! But no one outside Graphcore has used them so who knows. Intel's Nervana looked great on paper, right up until they dumped it.

There are some interesting accelerator options around for inference. But for training it's NVidia for almost everything, and TPUs as a good option is a few cases. But you can't buy TPUs (except for the inference-only Coral board), and universities like to own hardware.

> But no one outside Graphcore has used them so who knows.

This part really confuses me. I'd love to try out their hardware, but the only cloud offering they had last time I checked made you rent a whole month's worth (for many thousands of dollars).

It seems like getting it in the hands of devs should be a top priority.

From my very limited understanding the Graphcore machines outperform CUDA only significantly in inference, in training the improvements might not be sufficient to switch technology.

There is a trend towards larger models in the state of the art of deep learning research. The total cost of the on-chip memory on the NVidia GPU is still a good value proposition compared to current custom deep learning accelerators. This cost to get to large on-chip memory, combined with the flexibility of CUDA for other types of scientific applications, makes it harder for such accelerators to compete with Nvidia in such large academic research clusters. I hope that this situation changes over the next couple of years.

Is this really true? At least for inference I've seen data showing the trend is rather for smaller models that exploit sparsity.

This is a chart from before GPT-3: https://www.microsoft.com/en-us/research/blog/turing-nlg-a-1...

Yes, in NLP big models rule and they are getting bigger.