This article says OpenAI’s big computer is somewhere in the top 5 largest supercomputers. I reckon it’s fair to say their big computer is probably about 100 petaflops, or 10^17 flop per second. How much of that was used for GPT-3? Let’s calculate.
I’m told that GPT-3 was 3x10^23 FLOP. So that’s three million seconds. Which is 35 days.
So, what else have they been using that computer for? It’s been probably about 10 months since they did GPT-3. They’ve released a few things since then, but nothing within an order of magnitude as big as GPT-3 except possibly DALL-E which was about order of magnitude smaller. So it seems unlikely to me that their publicly-released stuff in total uses more than, say, 10% of the compute they must have available in that supercomputer. Since this computer is exclusively for the use of OpenAI, presumably they are using it, but for things which are not publicly released yet.
Is this analysis basically correct?
Might OpenAI have access to even more compute than that?
100 petaflops is ‘only’ about 1,000 GPUs, or considerably less if they are able to use lower precision modes. I’m guessing they have almost 100 researchers now? Which is only about 10 GPUs per researcher, and still a small budget fraction (perhaps $20/hr ish vs > $100/hr for the researcher). It doesn’t seem like they have a noticeable compute advantage per capita.
This article says OpenAI’s big computer is somewhere in the top 5 largest supercomputers. I reckon it’s fair to say their big computer is probably about 100 petaflops, or 10^17 flop per second. How much of that was used for GPT-3? Let’s calculate.
I’m told that GPT-3 was 3x10^23 FLOP. So that’s three million seconds. Which is 35 days.
So, what else have they been using that computer for? It’s been probably about 10 months since they did GPT-3. They’ve released a few things since then, but nothing within an order of magnitude as big as GPT-3 except possibly DALL-E which was about order of magnitude smaller. So it seems unlikely to me that their publicly-released stuff in total uses more than, say, 10% of the compute they must have available in that supercomputer. Since this computer is exclusively for the use of OpenAI, presumably they are using it, but for things which are not publicly released yet.
Is this analysis basically correct?
Might OpenAI have access to even more compute than that?
100 petaflops is ‘only’ about 1,000 GPUs, or considerably less if they are able to use lower precision modes. I’m guessing they have almost 100 researchers now? Which is only about 10 GPUs per researcher, and still a small budget fraction (perhaps $20/hr ish vs > $100/hr for the researcher). It doesn’t seem like they have a noticeable compute advantage per capita.