AMD's Radeon VII Packs A Low More FP64 Performance Than Expected

[WYP]

Great news for creators professional users.



Read more about AMD's Radeon VII Packing More FP64 Performance Than Expected.

 

[FTLN]

Your source is KO, any comparison to Nvidia for the FP64 compute ?

 

[grec]

NVidia only really has one (sort of almost) consumer card (Titan V) that comes anywhere near comparing since Kepler when they started building their gaming cards on distinct silicon.

That card can pump out over 7TFLOPs, while their gamer silicon like the RTX Titan can barely reach 0.5 TFLOPs.

Before Titan V, the HD7990 was still the top performing consumer card for double precision compute.

 

[WYP]

Your source is KO, any comparison to Nvidia for the FP64 compute ?

Hmm, for whatever reason the website didn't like how the link was originally formatted. Strange.

Anyways, right now none of Nvidia's consumer-grade GPUs offers FP64 compute that is this high, aside from the Titan V, which is the only current-gen Nvidia consumer GPU that delivers half-rate Double Precision Compute.

Nvidia doesn't officially publish numbers for half precision compute for a lot of their graphics cards, so it is hard to make a proper comparison.

 

[grec]

While this won't give exact real world figures we can take the allocation of compute units on the architectures to get a maximum theoretical figure, with Pascal (Excluding GP100) at 1/32 double precision vs single precision (Or 1/64 half precision). Turing also has 1/32 max theoretical FP64 but of course moves to x2 FP16 like GP100 & Vega.
The double precision side just comes down to core count, a fully enabled TU102(RTX) has 2 DP units per SM for 144 total, GV100(Volta) has 32 DP units per SM for 2688 total.