Nvidia Pascal Announcement

HBM2 isn't on-die, and the HBM2 is reportedly smaller than a GDDR5 controller. So the die size would be still somewhat plausible, when cutting on the register files compared to GP100.

So you expect the GP104 already has cut register files as it is 7.2B transistors ?
Personally I would think they keep that the same across a generation.
 
So what are we looking at here compared to 980? 1.6-1.7x perf/W increase? Both 180W, 1080 perf around 980 SLI (60% scaling).
 
Looks like rumors saying Pascal being Maxwell on 16FF were mostly true. The 1080 has 25% more compute resources and 55% higher clocks, which lead to 2x performance rather linearly.

Kudos on them for apparently accelerating GDDR5X.
 
Looks like rumors saying Pascal being Maxwell on 16FF were mostly true. The 1080 has 25% more compute resources and 55% higher clocks, which lead to 2x performance rather linearly.

Kudos on them for apparently accelerating GDDR5X.
Yes, opportunity for AMD to catch up.Dissapointing the Epic tech demo by the way, like PBR werent already possible...
 
Last edited:
I'm certainly not disappointed. Even the 1070 is supposed to be faster than the TitanX and given that it's raw core specs are almost identical (fill rate - assuming 64 ROPS, texture throughput, shader flops) but it only has about 76% of the bandwidth, then they must have made some efficiency improvements. The VR improvement looks awesome too, 1080 up to 2x faster than a Titan X in VR? Shut up and take my money!
 
Epic only having their inevitably uninteresting free-to-play FPS to show wasn't very cool.
What happened with Unreal? Is that thing now dead?

Why do you say that?

Actually, my numbers were wrong and maybe Pascal's performance-per-clock-per-compute is lower?
According to anandtech, nVidia is claiming 65% higher performance in 1080 than 980.
The 1080 has 43% higher clocks and 25% more compute resources. If multiplied linearly, this would result in a > 78% performance boost.
Of course, it seems like only compute and texture mapping resources were increased, so it's reasonable to assume the lesser performance gains are limited by fillrate and/or memory bandwidth.
Regardless, we're not observing any boost from "architectural improvements" (other than the obvious 16FF passage).



They're claiming asynchronous compute, for example. If true, I would expect this to cost a significant amount of power and area.
Like they claimed asynchronous compute for Maxwell 2?
 
Last edited by a moderator:
Perf announced on stage at the event is plain out 2x without any differentiation. (While the slide in the background has "VR performance" on the axis.)
Actually there was differentiation, they said it's "faster than 980 SLI" in regular gaming, while showing the slide where it was bit over 2x GTX 970 speed, ~1,75x GTX 980 speed and ~1,2x Titan X speed
 
I'm certainly not disappointed. Even the 1070 is supposed to be faster than the TitanX and given that it's raw core specs are almost identical (fill rate - assuming 64 ROPS, texture throughput, shader flops) but it only has about 76% of the bandwidth, then they must have made some efficiency improvements. The VR improvement looks awesome too, 1080 up to 2x faster than a Titan X in VR? Shut up and take my money!
Actually 1070 is supposed to be faster than Titan X "in VR when using simultaneous multiprojection", not in general.
 
What? Only 20% faster at almost 100% more clock rate? Is that a joke?

Time to bargain for a cheap 980Ti in wait for the real GeForces.
Need to be put into context though.
The base clock is 1607 Mhz and the boost is 1733MHz, so performance peak is that boost.
What they also show is that the reference (Founders Edition) can boost up to 2100MHz when OC with reasonable temps on air (academic how fast the fan is working as context being this is now possible), which surprises me and this is going to be interesting when it comes to AIBs.
At least this time NVIDIA has given the 1080 model a vapour chamber, so it should make a reasonable card, albeit at an eye watering premium $100 more than when the later AIBs version release.

Those massive clock speeds relating to 14nm; I wonder if this is why AMD has also gone narrow and fast this time round for Polaris as it just fits much better.
It will be interesting to see what speeds AMD Polaris manage.

Cheers
 
Actually there was differentiation, they said it's "faster than 980 SLI" in regular gaming, while showing the slide where it was bit over 2x GTX 970 speed, ~1,75x GTX 980 speed and ~1,2x Titan X speed
An approximation taken with a huge pinch of salt:
Suppossing 180 watts(1080 GTX) / 1,2 (because is 20% faster than Titan)= 150 watts for same performance as Titan X.
If Titan is 250 watts then 250/150=1,66.So would be something like 66% more efficient.In Polaris marketing terms it would be 1,66 times brighter.
 
Last edited:
I'm certainly not disappointed. Even the 1070 is supposed to be faster than the TitanX and given that it's raw core specs are almost identical (fill rate - assuming 64 ROPS, texture throughput, shader flops) but it only has about 76% of the bandwidth, then they must have made some efficiency improvements.

Close to nothing has been said about 1070's performance and I sincerely doubt it'll match a Titan X, much less be faster.
$370 does sound great for something (that might be) close to 980 Ti's performance at a smaller TDP, though. With this, Polaris 10 might get pushed to lower than $299.
 
Back
Top