With sparse matrix and 8-bit precision, you can effectively get over 100x the performance on an A100 than with FP64, plus the memory requirements are more than an order of magnitude less. Factor of 128 improvement in processing speed plus at least a factor of 8-16 or more improvement in tokens per memory, depending on sparsity.
https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Cent...
I think your compute cost figure is both very old and probably assumes double precision.
> Let's say the cost of compute is 1 cent per 100B flops (probably a lowball)
That's like 0.5ms worth of compute on A100. VMs with an A100 cost a few dollars per hour at cloud providers, i.e. about a cent per second or 0.001 cents per ms. OpenAI would obviously get a far better deal than somebody renting a single GPU for an hour.
Look, we know what OpenAI is charging for their commercial service. We don't know how it maps to their actual serving costs; maybe they're breaking even, making a small profit, or making a small loss. But just basic common sense should tell you that they are not operating at -10000% gross margins.
I suspect that just working backwards from the prices to a reasonable range of gross margins is going to produce much better estimates than your methodology would even after the inputs are tuned to be more realistic.
I sourced my cost/flop estimate from here: https://aiimpacts.org/2019-recent-trends-in-gpu-price-per-fl...
I'll acknowledge that a fleet entirely made up of A100 GPUs running at full capacity 100% of the time with Azure's 3-year commitment pricing (likely very low margin) would make this 5 orders of magnitude cheaper. But the sequential nature of transformers makes it difficult to achieve perfect utilization, and there aren't enough A100 GPUs in the Azure fleet to serve ChatGPT's billion users, so they're certainly running on mostly lesser hardware. Also, my estimate was for GPT-3 not 4, and neglected to address any memory or bandwidth costs, not to mention the costs to train and iterate on the models in the first place, including legions of data curators and many researchers commanding million dollar salaries.
If ChatGPT plus is $20/mo and each user uses it ~10 times per day, then MS/OpenAI would break even if my estimate were 2 orders of magnitude too high. Within the year, we'll have a much better idea of the real cost of these models.
The other thing to note is that OpenAI limits paying users to X queries per Y hours in GPT-4, something like 3 per 5. Which means the GPT-4 fleet is quite limited.
It also keeps saying "expect significantly lower caps, as we adjust for demand", but it's been three weeks now and they haven't lowered it yet.
> I'll acknowledge that a fleet entirely made up of A100 GPUs running at full capacity 100% of the time with Azure's 3-year commitment pricing (likely very low margin) would make this 5 orders of magnitude cheaper. But the sequential nature of transformers makes it difficult to achieve perfect utilization,
I don't think you've thought this through. There's no realistic level of utilization that's going to make up for 5 orders of magnitude. An unimaginably crappy 1% utilization would still leave you off by 3 orders of magnitude.
This very submission about OpenAI having to suspend selling subscriptions makes it pretty clear they're not having trouble with low utilization...
> If ChatGPT plus is $20/mo and each user uses it ~10 times per day, then MS/OpenAI would break even if my estimate were 2 orders of magnitude too high.
You do understand that this doesn't actually support your initial estimate, right? It just makes it more obvious how ridiculous it was.
Also:
> Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.
https://news.ycombinator.com/newsguidelines.html
If you're going to accuse me of using ChatGPT, at least bother to say why.
>> ... suppose you're describing this to a technical audience - software engineers, specifically. There might be some way to talk about percentages concisely in that particular context.
> In that case, you could describe the situation to a technical audience by saying, "The company's profit margin is effectively -∞% due to offering the product for free while incurring expenses."
I’ll be honest, I laughed out loud at this.
Pretty sure I read somewhere that Bing had around 1 billion searches per day BEFORE the hype of bingchat. At $7/query that makes MSFT’s $10 billion investment basically peanuts. Less than 2 days of Bings GPT usage payments. Lmao.
If they charged 1 cent per 100B flops that pricing would be $35.000 / 1K tokens. They may be taking a loss, but I doubt it would be at a 1 to 17,500 ratio.
Even GPT-4 with 32k token window is only $0.12 / 1K tokens
I've heard from insiders that the cost is roughly a couple cents per ChatGPT message but that was back with the legacy GPT-3.5 model, not the Turbo model or GPT-4
I might be totally misunderstanding this, but from the source you posted elsewhere[1], GPU Sale Price per Flops is as a little bit higher than 1*10^-11 dollars per flop, or 10^11 flops per dollar, or 10^9 flops per cent. So 1GFlops per cent. But Flops are Floating point operations per second. So you aren't paying 1cent for 100B flops, you're paying 1 cent for a card capable of doing 100B floating point operations per second. So let's say the card lasts 3 years and runs at 100% utilization (I know, it doesn't we can multiply by some efficiency metric later), there's about 9.5*10^7 seconds in 3 years. So that dollar cost isn't for 100B flops, it's for 100B * 9.5 *10^7 ( *some efficiency number). So it's many orders of magnitude cheaper than you're saying.
In reality, the cost of running these things isn't the cost of buying the GPUs, it's the cost of the power to run the GPUs, and that's a totally different calculation we can do.
[1]: https://aiimpacts.org/2019-recent-trends-in-gpu-price-per-fl...*
Let's just look at GPT 3. Each forward pass requires 2N=350B flops per token. The computational overhead of the attention mechanism is negligible but the memory overhead of the attention for all users is not, but we can ignore that for now. Let's assume each query involves ~200 tokens of combined input and output. That's 70T flops per query. Let's say the cost of compute is 1 cent per 100B flops (probably a lowball). That's $7 per query for compute alone.
And that's just GPT 3.