Do you get different results from TPU? Or how can you tell it's a bandwidth issue instead of something else?
I own a GTX980, the answer is yes.
Do you get different results from TPU? Or how can you tell it's a bandwidth issue instead of something else?
There is no memory OC on a GTX980 vs stock in the TPU you linked. So the answer is yes.
The 980 Ti (the Gigabyte one) has the overclocked memory, not the 980. Everything else is at stock.
We are not talking about the GTX980TI but the GTX980.
You have to remember that any X70/X80 would have 8 Ghz GDDR5 though so you would get higher bandwidth than what the 980 has.
2/3 of GP100's shaders at <1/2 the die size. Probably very agressively clocked too.
Bodes well for a future 'GP102' gaming GPU. We could be looking at up to 5120 SPs.
Vega 11 is the competitor to GP104. Polaris 10 is just too small to compete and it will face off versus GP106 as expected. If NV releases in retail Q3, they have a few months where there's no competition for the $400 - $550 segment.
GP100 is fully graphics capable. They aren't drawn on the diagrams, but it has display controllers, ROPs, etc. And I agree a Quadro is a good bet at some point.
Who is Ryan Smith?
Software running on the P100 can be preempted on instruction boundaries, rather than at the end of a draw call.
This means a thread can immediately give way to a higher priority thread, rather than waiting to the end of a potentially lengthy draw operation. This extra latency – the waiting for a call to end – can really mess up very time-sensitive applications, such as virtual reality headsets. A 5ms delay could lead to a missed Vsync and a visible glitch in the real-time rendering, which drives some people nuts.
By getting down to the instruction level, this latency penalty should evaporate, which is good news for VR gamers. Per-instruction preemption means programmers can also single step through GPU code to iron out bugs.
Spot on with the predictions earlier, @JDC & others who nailed it.
2560 SP (4x GPC, GP100 has 6x GPC with FP64), ~300mm2 size. This is a lean mean gaming chip.
They are going to need some major improvements in their memory compression.
But expect ~Titan X +15% performance at ~150W. In newer titles and GCN-optimized/DX12 games, you can expect a much bigger lead, potentially Titan X + 25-35%. It's definitely going to look more better against GCN in the new era of games than Maxwell, or LOL-Kepler.
With that down, GP106 will be 1280 SP and GP107 will be 680 SP.
^ The amazing thing, AMD's (Polaris 10: 2560 SP, Polaris 11: 1,280 SP) and NV's chips will have identical core counts overall at each segment here and in the same arrangement of cores per SM/TMU/ROP block!
Oh, btw, guys!! Good news!!
Pascal has basic Async Compute support on the hardware level!
http://www.theregister.co.uk/2016/04/06/nvidia_gtc_2016/
BOOOM! That's your real DX12 hardware, your real VR hardware. Any VR gamer that buys a Maxwell for VR, expect to upgrade to Pascal for vastly better motion to photon latency and less puke-factor.
NV is going to bank it big time on Pascal because of how gimped Maxwell is in VR for latency, people will realize soon enough, they bought obsolete GPUs for VR and will all upgrade.
What genius marketing PR. I am impressed. They held back talking about Async Compute all this time, promising it in their drivers... when Pascal comes, Async Compute is going to be "enabled" in the drivers, for Pascal.
There's gonna be a lot of VR coverage from NV for Pascal, a LOT. You'll see blog posts about how improvements in Pascal enable Async Timewarp latencies to drop below that magical 20ms barrier that is deemed acceptable. Much better than Maxwell. All the early VR adopters will upgrade in droves and NV is gonna laugh all the way to the bank. Pure genius.
Do you even understand what you are talking? Preemption has nothing to do with Async Compute. :\
Do you understand one of the crippling problem for NV's current GPU is their crushing context switches at draw call boundaries whenever compute needs to run?
This new pre-emption feature basically nullifies that weakness.
When an Async Compute task is called, like Async Timewarp (go read the Occulus blog), on Maxwell, it can't do it, it has to wait for the current graphics in the pipeline to finish first. Stall. It's going to miss the timewarp window and the user sees a major stutter/lag.
With their change in hardware to allow priority context switches immediately, they don't have to wait for the graphics rendering to finish, that async timewarp call is processed instantly.
This is a major change for VR, and while they may or may not be able to run graphics + compute in parallel, they won't be as neutered losing performance when Async Compute is in play.
You wait and see. I will be 100% correct on this.
Spot on with the predictions earlier, @JDC & others who nailed it.
2560 SP (4x GPC, GP100 has 6x GPC with FP64), ~300mm2 size. This is a lean mean gaming chip.
They are going to need some major improvements in their memory compression.
But expect ~Titan X +15% performance at ~150W. In newer titles and GCN-optimized/DX12 games, you can expect a much bigger lead, potentially Titan X + 25-35%. It's definitely going to look more better against GCN in the new era of games than Maxwell, or LOL-Kepler.
With that down, GP106 will be 1280 SP and GP107 will be 680 SP.
^ The amazing thing, AMD's (Polaris 10: 2560 SP, Polaris 11: 1,280 SP) and NV's chips will have identical core counts overall at each segment here and in the same arrangement of cores per SM/TMU/ROP block!
They dont let it run concurrently with the graphics pipeline. And there is no context switch with Async Compute because the Compute queue will be put after the graphics queue...