Nvidia Pascal Reviews [1080XP, 1080ti, 1080, 1070ti, 1070, 1060, 1050, and 1030]

The old "CUDA has access to command queues which should be exposed as compute queues in DX12 rather than doing everything on the GPC" complaint appears to remain valid though. I've not seen any indicator that they've fixed this yet.

Fair enough. Hopefully that capability extends to DX12 as well.
Nobody said that it has to be the most efficient or fastest implementation in existence. Similarly, nobody said that enabling async compute has to be faster than not enabling it: if a particular implementation is such that it can't find inefficiencies to exploit, then so be it.

Yeah there seems to be a general sentiment that async is some sort of silver bullet that provides benefits in every situation. It's just another tool that should only be applied when needed. It's very likely that optimizations done with GCN in mind will not work as well for nVidia's hardware and vice versa.

image.png
.
 
Yeah there seems to be a general sentiment that async is some sort of silver bullet that provides benefits in every situation. It's just another tool that should only be applied when needed. It's very likely that optimizations done with GCN in mind will not work as well for nVidia's hardware and vice versa.

Not even just between vendors. Different generations of GCN, or even different implementations see different speedups or regressions. The DX12 performance thread showed varying behaviors between GCN generations.
 
Spec for spec it's very close to a Titan-X. I understand NV are saying it will be faster, although it will likely be very marginally so. Factory O/C versions though should easily cruise past the Titan-X.

Forgetting memory bandwidth, the 1070 needs about 2050Mhz to match the Titan X at 1300Mhz, which is quite easy for Titan X and if you push the Titan X to 1400Mhz, 1070 then needs 2240Mhz to match it, so I would say 1070 has it's hands full with Titan X.

Granted if Titan X is kept at stock it is easily surpassed. In any case 1070 is a nice boost in performance for dollar compared to the GM200 chips.
 
So the GTX1070 has 25% of the compute resources disabled and 4-6% lower clocks. This means the GTX 1070 will have 70 to 73% the performance of a GTX 1080 (EDIT: when compute power is the bottleneck, obviously).
Are people still expecting the 1070 to be a "Titan X killer"?


By the way, 150 watts TDP.Thats Polaris 10 XT territory.We'll see.

Says who?
All sources so far point to Polaris 10 having Hawaii performance at lower price and power consumption. AMD has claimed 2.5x better power efficiency for Polaris.
Hawaii's TDP is 275W. 275/2.5 = 110W.

I know people love shootouts, but maybe it's time to realize that Polaris 10 is not in the same performance range as the GP104.
The only way I see Polaris 10 XT performing in the same ballpark as a GTX 1070 is if the latter turns out to be a huge disappointment in performance.
 
Last edited by a moderator:
I don't think any one is expecting the gtx 1070 to be a titan x killer, it should be right around the titan X in performance though

Just for fun what is the cost of this new chip and how is AMD going to price it lower to current products and maintain margins....

I can't see them keeping margins close if they place it at $300 for their top end Polaris if they cut prices less that $300, I'm sure they would be cutting into current margins, if they have the performance so they can bump up their price, I think they would do it wouldn't you?
 
So the GTX1070 has 25% of the compute resources disabled and 4-6% lower clocks. This means the GTX 1070 will have 70 to 73% the performance of a GTX 1080.
Are people still expecting the 1070 to be a "Titan X killer"?

It also has 80% for the 1080's memory bandwidth and almost all of it's fill rate and geometry throughput. And even if it didn't, since when did performance scale perfectly linearly with CU's and clock speed?

Also, I think the claim is that it will be faster than Titan-X, not a Titan-X killer. However, in sales terms, it should be a Titan-X killer (and a 980Ti killer) since it will be about as fast, or faster, at a vastly lower price.
 
I know people love shootouts, but maybe it's time to realize that Polaris 10 is not in the same performance range as the GP104.
The only way I see Polaris 10 XT performing in the same ballpark as a GTX 1070 is if the latter turns out to be a huge disappointment in performance.

It's not too great for AMD if their 232mm2 chip is not in the same performance range as nVidia's 314mm2 X 0.75 = 235.5mm2. It's going to be hard for them going forwards if this is the case.
 
It's not too great for AMD if their 232mm2 chip is not in the same performance range as nVidia's 314mm2 X 0.75 = 235.5mm2. It's going to be hard for them going forwards if this is the case.

Are you saying the GTX 1070 only has a 192bit bus, 48 ROPs, 25% of its cache disabled and 25% of its geometry output disabled?
Well then maybe you should send an e-mail to nvidia, because they seem to have their GTX 1070 specs page wrong.
 
It also has 80% for the 1080's memory bandwidth and almost all of it's fill rate and geometry throughput. And even if it didn't, since when did performance scale perfectly linearly with CU's and clock speed?

Also, I think the claim is that it will be faster than Titan-X, not a Titan-X killer. However, in sales terms, it should be a Titan-X killer (and a 980Ti killer) since it will be about as fast, or faster, at a vastly lower price.
Seeing how often the 1080 is hitting it's self-imposed power wall, it seems interesting to see that in terms of Compute/Texturing/Whatever per watt allowed, the 1070 has more headroom.
 
the 1070 will have more overclocking head room because its base clocks start off lower, but its going to be limited by the same 215 watt limit as the 1080, so having less units, it might not be able to reach the 1080 overclocked performance, should get pretty close I think though.
 
I wonder how much the bandwidth is going to hold back the 1070 overclock capability. It might have more thermal and power headroom than the 1080, but it might be bandwidth constrained too much to see a similar performance increase.
 
Are you saying the GTX 1070 only has a 192bit bus, 48 ROPs, 25% of its cache disabled and 25% of its geometry output disabled?
Well then maybe you should send an e-mail to nvidia, because they seem to have their GTX 1070 specs page wrong.

Fair enough... It'd be nice to see a die shot and see how much everything takes space there. I assume the SMs take the most space by far, so I'd imagine the difference still being quite small and also I've seen it speculated that the GF's 14nm process is denser than TSMC's 16nm.
 
Fine. That's Maxwell. So with Pascal, they're able to avoid this flush and reassign the SMs dynamically? That's a major improvement, right? So why the complaints? It's not perfect, it doesn't have the granularity of AMD. It's not the first time that there have been features that worked better for one vendor than the other.

I guess the main complain is that the big market quote of Nvidia will slow down its adoption and utilization.

I was even hoping we could have had that on GL with the next major version
 
I wonder how much the bandwidth is going to hold back the 1070 overclock capability. It might have more thermal and power headroom than the 1080, but it might be bandwidth constrained too much to see a similar performance increase.
I'm more afraid what the 1.5(5) Volt GDDR5 will do in terms of power consumption compared to 1.35v GDDR5X. Maybe partners will be able to market 1070s with G5X as well after a while. Gonna be interesting.
 
Well just like Maxwell 2, seems the 1080 is the poor value card.
According to interview with pcgameshardware, the 1070 will have the same number of ROPs, L2 cache, Delta-C compression.
The only difference according to the NVIDIA representative is the use of GDDR5 memory instead of GDDR5X.
Ok so bandwidth limited at higher resolutions, but might still pull off good results at 1440p.
http://www.pcgameshardware.de/Nvidi...cials/Geforce-GTX-1080-GTX-1070-KFKA-1195567/

So just like Maxwell 2, seems the two best in terms of performance/value will be 1070 and probably a 1080ti/Titan type card.
And a positive at least the 1070 is not crippled on memory side this time, so making it probably even better perf value in comparison to 1080.
Cheers
 
Well this is the goal of nV, because they want people to buy the 1070 instead of the 1080, and this actually grows the performance market.

Also when the 1080ti's come out, some of those 1070 buyers will upgrade to the 1080ti ;)

They did exactly this with the 970 too.
 
Back
Top