[News/Rumor] NVIDIA GeForce GTX 1080 Ti Specifications Leaked, Inbound for Holiday

[StyM]

https://www.techpowerup.com/225849/n...r-holiday-2016

NVIDIA is giving finishing touches to its next enthusiast-segment graphics card based on the "Pascal" architecture, the GeForce GTX 1080 Ti. Its specifications were allegedly screengrabbed by a keen-eyed enthusiast snooping around NVIDIA website, before being redacted. The specs-sheet reveals that the GTX 1080 Ti is based on the same GP102 silicon as the TITAN X Pascal, but is further cut-down from it. Given that the GTX 1080 is unflinching from its $599-$699 price-point, with some custom-design cards even being sold at over $800, the GTX 1080 Ti could either be positioned around the $850-mark, or be priced lower, disrupting currently overpriced custom GTX 1080 offerings. By pricing the TITAN X Pascal at $1200, NVIDIA appears to have given itself headroom to price the GTX 1080 Ti in a way that doesn't cannibalize premium GTX 1080 offerings.

The GTX 1080 Ti is carved out of the GP102 silicon by disabling 4 out of 30 streaming multiprocessors, resulting in 3,328 CUDA cores. The resulting TMU count is 208. The card could retain its ROP count of 96. The card will be endowed with 12 GB of GDDR5 memory across the chip's 384-bit wide memory interface, instead of GDDR5X on the TITAN X Pascal. This should yield 384 GB/s of memory bandwidth, significantly lesser than the 480 GB/s bandwidth the TITAN X Pascal enjoys, with its 10 Gbps memory chips. The GPU is clocked at 1503 MHz, with 1623 MHz GPU Boost. The card's TDP is rated at 250W, same as the TITAN X Pascal.

• 16 nm GP102 silicon
• 3,328 CUDA cores
• 208 TMUs
• 96 ROPs
• 12 GB GDDR5 memory
• 384-bit GDDR5 memory interface
• 1503 MHz core, 1623 MHz GPU Boost
• 8 GHz (GDDR5-effective) memory
• 384 GB/s memory bandwidth
• 250W TDP

[AliG]

Very interesting, I guess this means that P100 is still far away? I can't see why else they would risk castrating their own sales unless if they anticipate Vega will launch sooner than they can bring out their new flagship.

[StyM]

probably to be priced @ 999....

[AliG]

Maybe, but wouldn't that inherently still steal Titan XP sales? Or conversely, if you're a high end user, wouldn't you just fork over the extra $200 to get the absolute best product on the market?

I dunno, just seems interesting to me.

[StyM]

nvidia would probably price it between 699 - 1200.
somewhere around 799 - 899 or 899 - 999 .

[StAndrew]

These prices are getting a little ridiculous. I hope AMD can drive prices down with the RX 490.

[zanzabar]

that is missing a lot of shaders.

Very interesting, I guess this means that P100 is still far away? I can't see why else they would risk castrating their own sales unless if they anticipate Vega will launch sooner than they can bring out their new flagship.

p100 and the amd high end cards are way away. 16GB of HMB2 is around $700 right now since it only ships as "samples" unless you were NV and bought the whole test run from samsung. micron is not shipping yet and hynix has not validated. i also assumed that the titan x pascal was cut down so much to make room for the p100 when they can source reasonably priced HMB2.

[DashTrash]

These prices are getting a little ridiculous. I hope AMD can drive prices down with the RX 490.

I agree. It seems several years ago it seemed you could get a really high end card for $600ish. Now they are $1000+. I know you can get a good card for much less, but these prices at the high end are getting a bit out of control.

[AliG]

I'm curious if Vega 11 (I think that's the cutdown?) could make an earlier presence with GDDR5X. I personally think HBM is completely overhyped when it comes to high end gaming, and would be very content if they made the RX 490 a supersized RX 480, and then the next gen Fury X was the full Vega 10 with HBM2.

In either case, Vega needs to show big improvement in pixel fill rate. GCN's biggest downfall has been the lack of ROPs, and you can see how even with GCN 4 it requires significantly more float operations to accomplish the same work that a Pascal card can do (which clock higher as well). If they literally even out their texture:pixel fillrate ratio, they should perform much better in symmetric compute situations.

[DashTrash]

In either case, Vega needs to show big improvement in pixel fill rate. GCN's biggest downfall has been the lack of ROPs, and you can see how even with GCN 4 it requires significantly more float operations to accomplish the same work that a Pascal card can do (which clock higher as well). If they literally even out their texture:pixel fillrate ratio, they should perform much better in symmetric compute situations.[/QUOTE]

I don't know much about design but from am enthusiasts point of view, it does seem the last few GCN architectures are unbalanced. What I mean is Fiji has a lot of shaders, and in the right workload is really powerful. But in an average game, it doesn't stand out. Meaning it seems it should be faster than it is. Anyone know why they wouldn't have balanced the ROP/GCN core count more from Fiji to Polaris? I'm assuming it isn't real easy, otherwise they would have. Is it a cost related issue? Or is it a complication in manufacturing? I guess my main question is what can AMD do with GCN to make it faster besides clock speed? I'd be curious what your guy's thoughts on this are.