Due to exponential growth in the computing industry, we can assume that the majority of flops were shipped in the last few years. NVidia and AMD shipped over 12 million graphics cards last year. (Note this analysis excludes laptop chips, but includes datacenters.) If we take an NVidia GeForce RTX 2080 (20 Teraflops for half-floats) as a typical card, we can estimate that less than 12M * 20T * 5 = 1.2e21 flops were shipped as GPUs in the last five years, and we can assume that is the vast majority of GPU flops in active use today.
How are those GPUs utilized? Presumably, most are used in either personal computers or datacenters, with datacenters breaking down into ML (research/deployment), cryptomining and other uses I’m not aware of. Cryptomining may or may not be a significant fraction of the total. Gamers are very pissed at miners for “hoarding all the cards,” and all the industry players have reasons to obscure what is really going on in this industry, so we might not have accurate information.
I wanted to run a Fermi estimate on x86 CPUs for comparison but I couldn’t find the right information. (I didn’t look very hard.) I did find that about 27 billion ARM chips were sold in 2020. ARM chips vary widely, with at least some having >1000 Gflops but others having <10 Gflops. Using a conservative estimate, 27B * 10B * 5 = 1.35e21, suggesting that there are more flops in low-power embedded applications than in GPUs.
Afaik most low-power embedded ARM chips have more like 1Gflops than 10 - at least a few years ago, most didn’t even have a dedicated floating point unit, and you could probably couldn’t even get 100Mflops out of them.
Also, this is somewhat beside the point but if you tried to distribute any ML across such chips, you’d have enough latency problems ect. that they don’t fully count towards “compute capacity for mid-to-large AIs”
Due to exponential growth in the computing industry, we can assume that the majority of flops were shipped in the last few years. NVidia and AMD shipped over 12 million graphics cards last year. (Note this analysis excludes laptop chips, but includes datacenters.) If we take an NVidia GeForce RTX 2080 (20 Teraflops for half-floats) as a typical card, we can estimate that less than 12M * 20T * 5 = 1.2e21 flops were shipped as GPUs in the last five years, and we can assume that is the vast majority of GPU flops in active use today.
How are those GPUs utilized? Presumably, most are used in either personal computers or datacenters, with datacenters breaking down into ML (research/deployment), cryptomining and other uses I’m not aware of. Cryptomining may or may not be a significant fraction of the total. Gamers are very pissed at miners for “hoarding all the cards,” and all the industry players have reasons to obscure what is really going on in this industry, so we might not have accurate information.
I wanted to run a Fermi estimate on x86 CPUs for comparison but I couldn’t find the right information. (I didn’t look very hard.) I did find that about 27 billion ARM chips were sold in 2020. ARM chips vary widely, with at least some having >1000 Gflops but others having <10 Gflops. Using a conservative estimate, 27B * 10B * 5 = 1.35e21, suggesting that there are more flops in low-power embedded applications than in GPUs.
Afaik most low-power embedded ARM chips have more like 1Gflops than 10 - at least a few years ago, most didn’t even have a dedicated floating point unit, and you could probably couldn’t even get 100Mflops out of them.
Also, this is somewhat beside the point but if you tried to distribute any ML across such chips, you’d have enough latency problems ect. that they don’t fully count towards “compute capacity for mid-to-large AIs”