Independent F@H Numbers for 3080

[Gorgon]

Phoronix has posted some numbers testing the RTX3080FE on the Phoronix Test Bench.

 

So really nice performance uplift in terms of PPD/$ but Pascal levels of efficiency (PPD/W) rather than Turing.

 

No sign of a review from AnandTech on either the 3090 or 3080

[TVwazhere]

27 minutes ago, Gorgon said:

Pascal levels of efficiency (PPD/W) rather than Turing.

So theoretically, 2060's/2070's are still the best PPD/watt?

[FloRolf]

12 minutes ago, TVwazhere said:

So theoretically, 2060's/2070's are still the best PPD/watt?

According to his link that would actually be the standard 1660.

[Gorgon]

I'm waiting for the 3070 results.

 

I'm not, however, waiting for Big Navi as there is still the issue that the OpenCL Compute numbers fall far short of the Theoretical maximum from that which the FP32 TFLOPs number suggest. This is a driver issue and well let's face it AMD does not have a very good record when it comes to fixing driver issues for Gaming and Compute is most likely an even farther priority for them.

 

So even if AMD does come out with a stellar card with lots of potential for compute the whole drivers issue combined with the optimized CUDA core will likely make it a non-starter.

 

The other factor is that these cards are running closer to the edge of the power envelope so they might still be much more efficient at 60-70% of the Power Limit than Turing Cards are. i.e. if you drop their power limit by 20-30% you'd see a smaller decrease in PPD than you might see with Turing.

[Rybo]

Interesting.  This contradicts field reporting from Lars Systems GPU stats.  There's no holistic view yet, but the 3080 seems to be 2nd place in PPD per watt (1st place is 3090, which pays you money to fold).  Maybe it's because most users in the wild actually power limit their system, yielding better returns than stock settings?  

 

EDIT: Darn, he fixed the 3090 bug.  It no longer pays you, but it does cost the least per million points to run!