H2o vs. Ln2 - 3x Titan V - 1sec 499ms GPUPI - 1B

[gigioracing]

waiting for firestike and other 3d :-)

[Alan_Alberino]

Same, hope to see the power of this card on a good bencher hand and not a random reddit guy

[mahaudi]

Very strong System

[Splave]

Do you think it will improve much with cuda?

[H2o vs. Ln2(BANNED)]

Do you think it will improve much with cuda?

 

Positive, don't need the 4th cards for the actual wr. Matt work on this.

[speed.fastest]

What about not CUDA, you know right Titan V have very good compute performance from Tensor Core?

[yosarianilives]

Is it an issue of sending the workload to the gpu at this point? Would more cpu speed potentially help?

[_mat_]

What about not CUDA, you know right Titan V have very good compute performance from Tensor Core?

The Tensor Cores are for tensor mathematics and are not general purpose like the other cores. They won't play a part in the calculation of Pi.

 

Is it an issue of sending the workload to the gpu at this point? Would more cpu speed potentially help?

The issue is with the memory reduction to calculate the final sums of the billions of partial results. I don't know what's happening but the kernel seems to be incompatible.

 

I already reported the issue to NVIDIA and asked to borrow me a sample. No answer as usual. Anybody has an NVIDIA contact in the US that actually knows what he's talking about?