I wasn't sure where else to put this, but I wanted to inquire about incorrect stats being listed for some GPUs.
notably, the FP64 spec/ratio listed for Nvidia RTX A-series (A4000/A5000/etc) cards. the FP64:FP32 ratio is listed as "1:32" with seemingly no basis and in contrast to the official nvidia docs which show the Ampere GA10x (excluding GA100) architecture to be a 1:64 ratio with no special treatment of the RTX A-series cards. where did 1:32 come from? every other mention of this spec across the internet/reviews seems to come from you guys and as it looks to be incorrect, it should be corrected in your database.
independent testing by myself concludes that these cards have a 1:64 ratio just like all the other Ampere Geforce cards. (A4000 does ~300GFlops in pure FP64 benchmark/load)
used a benchmark called gpu-burn (here: https://github.com/wilicc/gpu-burn) ran this on several RTX A4000s and several RTX 3070Tis, which have the same GPU core counts, ran similar clockspeeds under the FP64 test (-d argument). both scored the same ~300GFlops for doubles (FP64).
notably, the FP64 spec/ratio listed for Nvidia RTX A-series (A4000/A5000/etc) cards. the FP64:FP32 ratio is listed as "1:32" with seemingly no basis and in contrast to the official nvidia docs which show the Ampere GA10x (excluding GA100) architecture to be a 1:64 ratio with no special treatment of the RTX A-series cards. where did 1:32 come from? every other mention of this spec across the internet/reviews seems to come from you guys and as it looks to be incorrect, it should be corrected in your database.
independent testing by myself concludes that these cards have a 1:64 ratio just like all the other Ampere Geforce cards. (A4000 does ~300GFlops in pure FP64 benchmark/load)
used a benchmark called gpu-burn (here: https://github.com/wilicc/gpu-burn) ran this on several RTX A4000s and several RTX 3070Tis, which have the same GPU core counts, ran similar clockspeeds under the FP64 test (-d argument). both scored the same ~300GFlops for doubles (FP64).
Last edited: