Big labs like OpenAI and Deepmind have big clusters that support this kind of bursty allocation for their researchers, but startups so far have had to get very small clusters on very long term contracts, wait months of lead time, and try to keep them busy all the time.

Our goal is to make it about 10-20x cheaper to do an AI startup than it is right now. Stable Diffusion only costs about $100k to train -- in theory every YC company could get up to that scale. It's just that no cloud provider in the world will give you $100k of compute for just a couple weeks, so startups have to raise 20x that much to buy a whole year of compute.

Once the cluster is online, we're going to be pretty much the only option for startups to do big training runs like that on.

In 2023 you can barely get a single TPU for more than an hour. Back then you could get literally hundreds, with an s.

I believed in TRC. I thought they’d solve it by scaling, and building a whole continent of TPUs. But in the end, TPU time was cut short in favor of internal researchers — some researchers being more equal than others. And how could it be any other way? If I made a proposal today to get these H100s to train GPT to play chess, people would laugh. The world is different now.

Your project has a youthful optimism that I hope you won’t lose as you go. And in fact it might be the way to win in the long run. So whenever someone comes knocking, begging for a tiny slice of your H100s for their harebrained idea, I hope you’ll humor them. It’s the only reason I was able to become anybody.

Um. Can't you order them from coral.ai and put them in an NVMe slot? Or are the cloud TPUs more powerful?

One TPU (not even a pod, just a regular old TPUv2) has 96 CPU cores with 1.4TB of RAM, and that’s not even counting their hardware acceleration. I’d love to buy one.

A single TPUv2 chip has 1 core and 8gb of memory. A single device comes in the v2-8 configuration with 8 cores and 64gb of memory.

Pod variants come in v2-32 to v2-512 configurations.

I know just enough about the architecture to facilitate using TPUs for research training runs but I'm not sure what's so special about the host?

Sure it's beefy but there are much beefier servers readily available.

The TPUs you rent that are being discussed here are capable of training, consume hundreds of watts and have a heatsink bigger than your fist and really spectacular network links. They're analogous to Nvidia's highest end GPUs from a "what can you do with them" perspective.

Both are custom chips for deep learning but they're completely different beasts.

As for the rest of them, list them on Amazon and let them do the fulfillment. That $10k of hardware isn't going to sell itself from your closet. (Yet. LLMs are making great strides.)

And that's a good idea, thanks. I've been dreading the idea of using ebay.

This is the nicest thing anyone has said to us about this. We're gonna frame this and hang it out on our wall.

> So whenever someone comes knocking, begging for a tiny slice of your H100s for their harebrained idea, I hope you’ll humor them.

Absolutely! :D

I'm not encouraging the false belief that everything you do will work out. Instead I'm encouraging the realization that the greatest accomplishments almost always feel like long shots, and require significant amounts of optimism. Fear and pessimism, while helpful in appropriate doses, will limit you greatly in life if you let them rule you too significantly.

When I look back on my life, the greatest accomplishments I've achieved are ones where I was naive yet optimistic going into it. This was a good thing, because I would have been too scared to try had I really known the challenges that lay ahead.

I argue that realism trumps optimism. It's perfectly normal in a realist farming to see something difficult, acknowledge the high risk and failure potential, and still pursue something with intent to succeed.

I've personally grown tired of over optimism everywhere because it creates unrealistic situations and passes consequences of failure in an inequitable way. The "visionary" is rewarded when the rare successes occur, while everyone else suffers the consequences for most failures. No contingency plans for failure, no discussion of failure, and so on. Optimism just takes any idea, pursues it and consequences be someone else's problem and be damned.

Pessimism isn't much better, you essentially think everything is too risky or unlikely to succeed so you never do anything. You live in a state of inaction because any level of risk or uncertainty is too much.

To me, realism is much better. You acknowledge the challenge. You acknowledge the risk. You make sure everyone involved understands it, but you still charge forward knowing you might succeed. Some think if you're not naively optimistic (what most people in my experience refer to as "optimism") you don't create enough pressure. I think that's non-sense.

Sam Altman talks about this quite frequently, that it's not intelligence or luck necessary for an enduring innovation. It is persistence in the face of inevitability, and a high tolerance for being proven wrong and still persisting

In this context, I tend to read the parent claim as something like, "great success requires willingness to sometimes take worse-than-even odds or pursue modestly-negative-EV opportunities". I'm not sure I agree with the strongest version of that, but I think it's likely that the space of risky paths to great achievement is richer than that of cautious ones.

Mostly agree, except the market is not an optimistic place — it’s the market.

There are a multitude of reasons you lose your optimism, mostly because people take it away — your optimism is their money

My understanding was that this situation changed drastically depending on what sort of email you had or how popular your Twitter handle was.

Are you affiliated with an academic institution? Otherwise I'm not sure why they're been more generous with me, my projects have been mildly interesting at best.

They're certainly a lot stingier with larger pods than they used to be though.

Check out this list of recent TRC-supported publications: https://sites.research.google/trc/publications/

Demand for Cloud TPUs is definitely intense, so if you're using preemptible capacity, you're probably seeing more frequent interruptions, but reserved capacity is also available. Hope you email the TRC support team to say hello!

This may feel like an anime betrayal, since you basically launched my career as a scientist. But it’s important for hobbyists and tinkerers to be able to participate in the AI ecosystem, especially today. And TRC just does not support them anymore. I tried, many times, over the last year and a half.

You don’t need to take my word for it. Here’s some unfiltered DMs on the subject: https://imgur.com/a/6vqvzXs

Notice how their optimism dries up, and not because I was telling them how bad TRC has become. It’s because their TPUs kept dying.

I held out hope for so long. I thought it was temporary. It ain’t temporary, Zak. And I vividly remember when it happened. Some smart person in google proposed a new allocation algorithm back near the end of 2021, and poof, overnight our ability to make TPUs went from dozens to a handful. It was quite literally overnight; we had monitoring graphs that flatlined. I can probably still dig them up.

I’ve wanted to email you privately about this, but given that I am a small fish in a pond that’s grown exponentially bigger, I don’t think it would’ve made a difference. The difference is in your last paragraph: you allocate reserved instances to those who deserve it, and leave everybody else to fight over 45 minutes of TPU time when it takes 25 minutes just to create and fill your TPU with your research data.

Your non-preemptible TPUs are frankly a lie. I didn’t want to drop the L word, but a TPUv3 in euw4a will literally delete itself — aka preempt — after no more than a couple hours. I tested this over many months. That was some time ago, so maybe things have changed, but I wouldn’t bet on it.

There’s some serious “left hand doesn’t know that right hand detached from its body and migrated south for the winter” energy in the TRC program. I don’t know where it embedded itself, but if you want to elevate any other engineers from software devs to researchers, I urge you to make some big changes.

One last thing. The support staff of TRC is phenomenal. Jonathan Colton has worked more miracles than I can count, along with the rest of his crew. Ultimately he had to send me an email like “by the way, TRC doesn’t delete TPUs. This distinction probably won’t be too relevant, but I wanted to let you know” (paraphrasing). Translation: you took the power away from the people who knew where to put it (Jonathan) and gave it to some really important researchers, probably in Brain or some other division of Google. And the rest is history. So I don’t want to hear that one of the changes is “ok, we’ve punished the support staff” - as far as I can tell, they’ve moved mountains with whatever tools they had available, and I definitely wouldn’t have been able to do any better in their shoes.

Also, hello. Thanks for launching my career. Sorry that I had to leave this here, but my duty is to the open source community. The good news is that you can still recover, if only you’d revert this silly “we’ll slip you some reserved TPUs that don’t kamikaze themselves after 45 minutes if you ask in just the right way” stuff. That wasn’t how the program was in 2019, and I guarantee that I couldn’t have done the work I did then under the current conditions.

> But it’s important for hobbyists and tinkerers to be able to participate in the AI ecosystem

Totally agree! This was a big part of my original motivation for creating the TPU Research Cloud program. People sometimes assume that e.g. an academic affiliation is required to participate, but that isn't true; we want the program to be as open as possible. We should find a better way to highlight the work of TRC tinkerers - for now, the GitHub and Hugging Face search buttons near the top of https://sites.research.google/trc/publications/ provide some raw pointers.

I'm sorry to hear that you've personally had a hard time getting TPU v3 capacity in europe-west4-a. In general, TPU availability varies by region and by hardware generation, and we've experimented with different ways of prioritizing projects. It's possible that something was misconfigured on our end if your TPU lifetimes were so short. Could you email Jonathan the name of the project(s) you were using and any other data you still have handy so we can figure out what was going wrong?

Also, thanks for the kind words for Jonathan and the rest of the TRC team. They haven't lost any power or control, and they are allocating a lot more Cloud TPU capacity than ever. However, now that everyone wants to train LLMs, diffusion models, and other exciting new things, demand for TPU compute is way up, so juggling all of the inbound TRC requests is definitely more challenging than it used to be.

It would be funny if someone set gpt-2-15b-poetry (our project) in some special way to prevent us from making TPUs that ever last more than a few hours, but from what I’ve heard from other people, this isn’t the case. That’s what I mean about the left hand doesn’t know what’s going on with the right hand. It’s not a misconfiguration. Again, pretend to be some random person who just wants to apply for TPU access, fill out your form, then try to do research with the TPUs that are available to you. You’ll have a rough time, but it’ll also cure this misconception that it’s a special case or was just me.

Again, no need to take my word for it; here’s an organic comment from someone who was rolling their eyes whenever I was cheerleading TRC, because their experience was so bad: https://news.ycombinator.com/item?id=36936782

I think that the experience is probably great for researchers who get special approval. And that’s fine, if that’s how the program is designed to be. But at least tell people that they shouldn’t expect more than an hour or two of TPU time.

By default, the TRC program grants both on-demand quota and preemptible quota. If you are able to create a TPU VM with your on-demand quota, it should last quite a bit longer than a few hours. (There are situations in which on-demand TRC TPU VMs can be interrupted, but these ought to be rare.) If your on-demand TPU VMs are being interrupted frequently, please email TRC support and provide the names of the TPU hosts that were interrupted so folks can try to help.

When there is very high demand for Cloud TPUs, it's certainly possible for preemptible TPU VMs to be interrupted frequently. It would be an interesting engineering project to make a very robust training system that could make progress even with low TPU VM uptime, and I hope someone does it! Until then, though, you should have a better experience with on-demand resources when you're able to create them. Reserved capacity is even better since it provides an expectation of both availability and uptime.

I'm sorry that you had such a frustrating time and that we weren't able to sort it out via email while it was happening. If you decide to try TRC again and run into issues like this, please be sure to engage with TRC support!

> Notice how their optimism dries up, and not because I was telling them how bad TRC has become. It’s because their TPUs kept dying.

Unless I'm misreading this they sound pretty happy and you sound pessimistic? Their last substantial comment was "I'm sure Zak could hook you up with something better"?

As for their comments, the third screenshot is the key; they’re agreeing that the situation is bad. They’re a friend, and they’re a little indirect with the way they phrase things. (If you’ve ever had a friend who really doesn’t want to be wrong, you know what I mean; they kind of say things in a circular way in order to agree without agreeing. After awhile it’s pretty cute and endearing though.)

I was particularly pessimistic in those DMs because it came a couple months after I thought I’d give TRC one last try, back in January, which was roughly a year after I’d started my “ok, I’m losing hope, but I’ll wait and see” journey. In the meantime I kept cheerleading TRC and driving people to their signup page. But after the TPUs all died in less than two hours yet again, that was that.

I have a really high tolerance for faulty equipment. This is free compute; me complaining is just ungrateful. But I saw what things were like in 2019. “Different” would be the understatement of the century. If my baby wasn’t being incubated in the NICU today, I’d show the charts where our usage went from thousands of cores down to almost zero, and not for lack of trying.

It also would’ve been fine to say “sorry, this is unsustainable, the new limits are one tpu per person per project” and then give me a rock solid tpu. We had those in 2021. One of our TPUv3s stayed online for so long that I started to host my blog on it just to show people that TPUs were good for more than AI; the uptime was measured in months. Then poof, now you can barely fire one up.

I'm just pointing out that your summary of the DMs ("Notice how their optimism dries up, and not because I was telling them how bad TRC has become. It’s because their TPUs kept dying") is the opposite of what the DMs show.

Frankly, it sounds to me like they're having severe yield+reliability problems with the TPUv4s that aren't getting caught by wafer-level testing, and have binned the flakiest ones for use by outsiders.

A lot of yield issues show up as spontaneous resets/crashes.

The Googlers maintaining the TPU Github repo also just basically don't care about your PR unless it's somehow gonna help them in their own perf review.

In contrast with a GPU-based grid, you can not only run the latest & greatest out-of-the-box but also do a lot of local testing that saves tons of time.

Finally, the OP here appears to be offering real customer engagement, which is totally absent from my own GCloud experiences across several companies.

Oh come on, colab gives TPU access in the free tier for a whole half day. No need to exaggerate the shortage

554 5.7.1 <[email protected]>: Relay access denied

  roomservice.dev. 60 IN MX 5 alt1.aspmx.l.google.com.
  roomservice.dev. 60 IN MX 5 alt2.aspmx.l.google.com.
  roomservice.dev. 60 IN MX 1 aspmx.l.google.com.
  roomservice.dev. 60 IN MX 10 alt3.aspmx.l.google.com.
  roomservice.dev. 60 IN MX 10 alt4.aspmx.l.google.com.
  roomservice.dev. 60 IN MX 15 4ig53n4pw7p3cuxm7n7xi7dpuyq6722aipexvhkngzbd2e4mudmq.mx-verification.google.com.

Where will the cluster be hosted ?

May I suggest that you get your IP transit from he.net ?

Right now, it's pretty easy to get a few A/H100s (Lambda is great for this), but very hard to get more than 24 at a reasonable price ($~2 an hour). One often needs to put up a 6+ month commitment, even when they may only want to run their H100s for an 8 hour training run.

As a graduate student, thank you. Thankfully, my workloads aren't LLM crazy so I can get by on my old NVIDIA consumer hardware, but I have coworkers struggling to get reasonable prices/time for larger scale hardware.

https://open.spotify.com/track/3VIMS1p3sNifH0RQnmDf7s

No idea how capacity at lambdalabs actually looks like though. Does anyone have insight how easy it is to spin up more than 2-3 instances up there?

I have never seen a GPU crunch quite like it is right now. To anyone who is interested in hobbyist ML, I highly highly recommend using vast.ai

Also, many of the available options clearly are recycled crypto mining rigs which have somewhat odd configurations (poor gpu bandwidth, low cpu ram).

I am guessing you mean at some point just buy your own 3090 as it will be cheaper than paying a cloud per second for a server-grade Nvidia setup.

What I mean is that you can rent out 4 3090 GPUs for much cheaper than renting an A100 on aws because you are not paying Nvidia's "cloud tax" on flops/$

For H100s and A100s - lambda, fluidstack, runpod. Also coreweave and crusoe and oblivus and latitude

For non a/h100s: vast, Tensordock, also runpod here too

Neither Alex and I are currently VCs, and this has no affiliation with any venture fund.

We want to be a customer of the sf compute group too!

If not and you got the loan from a bank, super curious how you were able to get the bank to trust that renting out the GPUs would cover the loan or if some other reasoning convinced them. Assuming you aren’t trying to turn this into a big business, that knowledge might help a lot of other players run similar programs and further democratize SOA GPU access.

Anyone have an insight?

When was the last time you spoke to a chatbot?

Basically twitter devolves into the Colos of the late 90s :-)

Make money off your GPU with vast.AI

https://cloud.vast.ai/host/setup

> Ubuntu 18.04 or newer (required)

> Dedicated machines only - the machine shouldn't be doing other stuff while rented

well that's certainly not what I expected. ctrl-f "virtual" gives nothing, so it seems they really mean "take over your machine"

> Note: you may need to install python2.7 to run the install script.

what kind of nonsense is this? Did they write the script in 2001 and just abandon it?

Anything AI/ML is a hot mess of cobbled-together bits and pieces of Python barely holding together. I recently read somewhere that there should be a new specialization of "ML DevOps Engineer"... and hell I'm supporting that.

Do you mean MLOps? Nothing new about it. We have two full-time MLOps engineers at our startup.

Python is terrible because they built what people wanted.

A comment in that discussion mentions yet another competitor in this space that I've never heard of: https://www.qblocks.cloud/ -- I just tried Q blocks out and the new user experience wasn't as good for me as with vast.ai: you have to put in $10 money to try it, instead of getting to try it initially for free; there is a manual approval process before you can try data center class GPUs; you only see that your instance is in Norway (say) after you try to start it, not before; it seems like there's no ssh access, and they only provide Jupyter to connect; neither pytorch nor tensorflow seemed to be installed. They could probably update their pages too, e.g., https://www.qblocks.cloud/vision is all about crypto mining and smartphones, which feels a bit dated... :-)

It's unique in that you can set your own prices, it's a true spot marketplace.. I've grabbed 2x3090 for $0.02/hr before.

Probably no good for training (can be interrupted any time with zero warning ssh just drops and that's it) but for my inference usecases it lets me spot heavy compute for pennies.

Price and market depth are very different things

[1] We have not gone the way of the Xerces blue [2] yet... we still exist!

[2] https://en.wikipedia.org/wiki/Xerces_blue

But I do think a lot of our customers will be out here —- SF is still probably the best place to do startups. We just have so many more people doing hard technical stuff here. Literally every single place I’ve lived in SF there’s been another startup living upstairs or downstairs

Good idea to host some in person events!

now that's a hot take if I ever saw one

In theory, this sounds almost identical to the business model behind AWS, Azure, and other cloud providers. "Instead of everyone buying a fixed amount of hardware for individual use, we'll buy a massive pool of hardware that people can time-share." Outside of cloud providers having to mark up prices to give themselves a net-margin, is there something else they are failing to do, hence creating the need for these projects?

They want to do that themselves, and keep the customer relationship and the profits, instead of giving them to a middleman or the customer.

You can rent a 2-socket AMD server with 120 available cores and RDMA for something like 50c to $2 per hour. That’s just barely above the cost of the electricity and cooling!

What do you want, free compute just handed to you out of the goodness of their hearts?

There is incredible demand for high-end GPUs right now, and market prices reflect that.

The instances with the 4th generation "Genoa-X" processors (HB176rs_v4) cost about $2.88 per hour. The HX176rs_v4 model with 1.7 TB of memory is $3.46 per hour.

https://learn.microsoft.com/en-us/azure/virtual-machines/hbv...

https://learn.microsoft.com/en-us/azure/virtual-machines/hbv...

https://learn.microsoft.com/en-us/azure/virtual-machines/hx-...

It's just business and I'd do the same if I was in charge of AWS.

Source required

A large part of the profit comes from the upfront risk of buying machines. With this you are just absorbing that risk which may be better if the startup expects to last.

Secondly, there is a fundamental question of resource sharing here. Even with this project by Evan and AI Grant (the second such cluster created by AI Grant btw), the question will arise — if one team has enough money to provision the entire cluster forever, why not do it? What are the exact parameters of fair use? In networking, we have algorithms around bandwidth sharing (TCP Fairness, etc.) that encode sharing mechanisms but they don’t work for these kinds of chunky workloads either.

But over the next few months, AWS and others are working to release queueing services that let you temporarily provision a chunk of compute, probably with upfront payment, and at a high expense (perhaps above the on demand rate).

I would srgue this has always been a common case for cloud GPU compute

1) Margins. Public cloud investors expect a certain margin profile. They can’t compete with Lambda/Fluidstack’s margins.

2) To an extent also big clouds have worse networking for LLM training. I believe only Azure has infiniband. Oracle is 3200 Gbps but not infiniband, same for AWS I believe. GCP not sure but their A100 networking speeds were only 100 Gbps I believe rather than 1600. Whereas lambda, fluidstack and coreweave all have ib.

3) Availability. Nvidia isn’t giving big clouds the allocation they want.

Where’s the capital for upgrades, repairs, and replacements coming from?

Of course it doesn't always work, and it may be even harder to make it work in the current macroeconomic environment, but it's still pretty standard play.

Sincere question.

Lambda has "Lambda Sprint" which is kinda similar,[1] but Sprint is $4.85/GPU/hr instead of <$2.

So if you want 128 GPUs for a week, you can't use Lambda reserved (3 year term), you can't use Lambda on-demand (can't get 128 A/H100s on-demand), your options are Lambda Sprint or SF Compute, and SF Compute is offering significantly lower prices.

[1]: https://lambdalabs.com/service/gpu-cloud/reserved

The most expensive part would be the land, but honestly there is some pretty cheap land outside the cities.

And when we are talking about low margins, a 5-10% difference in cost is very significant.

Perhaps its also a way for freshly applying grad students to look at a university looking to do research in LLMs that requires scale...

Who is funding this?

Cause if it’s VC then it’s going to have the same fate as everything else after 5-7 years.

I hope y’all have as innovative of a business model. You’ll need it if you want to do what you’re doing now for more than a few years

Not everything has grow to have the appetite of Galactus and swallow a whole planet. Making single digit millions of dollars over a couple of years is still worthwhile, especially if it helps others and moves humanity forwards.

This project isn't ever going to want to try and compete with AWS, so no, it's not a billion dollar question. $20 Million, yeah.

That’s why I’m asking because a “bootstrapped” company like you describe has a future…

One backed by VC doesn’t

I mean they may have a future but not like you describe

[1] https://andromedacluster.com/

But for AI training? If the public cloud isn't competitive even for bursty AI training, their margins are much higher than I anticipated.

OP mentions 10-20x cost reduction? Compared to what? AWS?

I've never had to buy very large compute, but I thought that was the whole point of the cloud

Is it accurate to say you’re willing to go into ~20,000,000 USD debt to sell discounted computer-as-a-service to researchers/startups, but unwilling to go into debt to sponsor the undergraduate degrees of ~100-500 students at top-tier schools? (40k - 200k USD per degree)

Or, you know, build and fund a small public school/library or two for ~5 years?