Polama wrote:
the individual tensor cores are 4x, but there's half as many of them. From your numbers, you saw 0.5x cores and 2.1x more tflops total throughput (meaning about 4x
per core). And as you note, benchmarks reported by the manufacturer can be misleading.
Ok, right. 4x is the theoretical order of magnitude per core, 2.7x is what Nvidia promises but probably is only an upper bound so the 2.1x more tflops total throughput for 3080 compared to 2080 is somewhat closer to the truth.
However, a first ALU cores test puts the promised 2x in relation. Nvidia selected 8 sample 3D games and their test resulting in an average 1.8x improvement from 2080 to 3080. Given Nvidia's bias, that must be an upper limit, too. Since 2080 TI is circa 1.3 as fast as 2080 for 3D games, we get 1.8 / 1.3 ~= 1.4 as the factor from 2080 TI to 3080 for 3D games.
For tensor cores, it might be a bit more.
Similar guesstimates for 3090 give circa 1.7x as the factor from 2080 TI to 3090 for 3D games.
So I doubt that 3080 or 3090 can quite reach 2x compared to 2080 TI for deep learning.
Nevertheless, close to 2x might be good enough: At the EGC Pisa, which ended on 2018-08-05, a professor of computer science from, IIRC, San Francisco (sorry, forgot his name) said that 2x 1080 TI (or was it 2x 2080 TI) roughly equalled the 4 TPUs of AlphaGo Zero. Since 2080 TI was launched only afterwards on 2018-09-27, I think what he must have said was 2x 1080 TI. Hence, if 3090 is circa 2x 1080 TI, a 3090 would be good enough, although 2x 2080 TI would still be faster but only for programs actually using SLI.
Then there is the option to await 3080 TI hoping it to have SLI but I guess we speak of 2x net $1100 or $1200 to achieve very roughly 1.35x the speed of 2x 2080 TI.
So far my current kaffeesatzleserei. The principle options for alleged >9p play are:
2x 2080 TI (currently used only in the USA with aproximately reasonable prices)
3080 (probably not enough, although more than good enough for kyu learners)
2x 3080 (presumes the programs to use them despite missing SLI)
3090 (probably not quite, but maybe good enough nevertheless; advantage of avoiding SLI troubles)
2x 3080 TI (if this will have SLI)
2x 3090 (clear case but too expensive by far)
EDITs