Separate names with a comma.
Discussion in 'nVidia Flavor' started by Dayaks, Jul 25, 2017.
I like the old style green pcb too .
I wish those Ai developers would of run some benchmarks with the cards just for the hell of it.
The just released PGI v17.7 (compiler and tools) now supports Volta
the PGI page:
Volta V100 DGX System spotted on geekbench.
If that is real Volta hit its x2 perf/watt over Pascal at least in compute.
Those R&D $$$ hard at work
Does this thing need 1.21 Gigawatts of power?
I'm excited, going to hurt the bank account a bit, but imho best titan is the first titan.
Probably about the same as a Vega or two... so yeah.
I hope it can take me back in time to buy some NVDA
GV100 compared to fastest GP100 and fastest Vega in convenient table form.
Link to individual scores:
Shit that is crazy ever single test is more than x2 other than SFFT, that is not something I expected, SFFT I would expect that, compression algorithms and what not have limitations on registers and DP units.
Just shows how far paper specs and reality is as always.
I am just not sure what GV102 will look like.
Look at GV100, remove Tensor cores or reduce? Certainly remove FP64. Replace HBM2 with GDDR6
I know can extrapolate from that though
We might be back to the days of >= 70% performance improvement much to the chagrin of AdoredTV.
A new fat chip that is more than double the performance of my GP102-400 would be a perfect match with a PG27UQ .
Assumptions being assumptions. Here we go.
GV102 (Titan) 5120 anywhere from 12-16gb of VRAM.
Assuming it clocks around 2ghz it should hit a theoretical maximum of 20Tflops.
Thats if I remember the calculation correct though. Core count*2*clock/1000
5120*2*2000/1000 = 20tflop
Memory is 12 or 24 since it´s 384bit, 11 or 22GB for a cutdown Geforce.
Titan 2.0 will be 5376SP
The xxxxTI version will likely be 4864SP with 296 TMU and 88 ROPs.
Your TMUs are off and would be 336 for 5376 SP and 304 for 4864 SP.
While not optimal, it is still possible to do 16GB on a 384-bit bus. NVIDIA has done 2GB on a 192-bit bus in the past. But I don't think there is any need for more than 12GB in games at this point unless you're running 4K Surround.
I dont think we will see that solution on such a card.
But who knows with demand, I think Scorpio will kick the VRAM bucket with its +4GB.
The Power9 Rollout Begins With Summit And Sierra Supercomputers
16 GB HBM2? Nice.
I hope your title is no homo, Razor.
oops, that was meant for Razor
If this GPU has the gold logo saying Volta instant buy can't stand that green GTX logo even with the lights off it bothers me it's so ugly.
Once upon a time I had a EVGA 980ti ACX the logo on their wasn't too big and it just looked good. I think ti's machine or lazer etched in the metal it looks so good.
Plus you usually have to wait around for the Nvidia drivers to mature once a new card comes out when the 1080ti came out at the same time as Windows 10 creators update it was a blurry eyestrain mess for an entire month. I thought it was my data or DVI connection it was just the drivers.