Nvidia Pascal Announcement

http://www.guru3d.com/news-story/nvidia-announces-the-geforce-gtx-1070-and-1080.html

 

Well seeing as how he was talking about TAM and specifically mentioned $349 and up as the differentiator... he would be quite right.
He never said you couldn't compare the performance, he was just talking about target markets. Nvidia is targeting a smaller market(7.5million) with Pascal while they are specifically targeting a much larger market with Polaris.

 

How on earth do you get those numbers?
Pascal is ~20% faster than Maxwell at bit over 70 % of the power

 

Pascal from the gtx 980 to the gtx 1080, its more like 60% in increased performance with 10% increase in power consumption, which is 2.0x perf per watt.

From the 980ti its a 30% increase in performance with a 30% drop in power usage. That goes to 2.5 x the perf. watt. All of this with less ALU's.

 

How you get to that ?
180 Watt for 9 TFLOP vs 250 Watt for 7 TFLOP is 20 Watt/TFLOP vs 36 Watt/TFLOP or 1.8x better.
Pretty good actually.

 

Flops aren't the only thing that come into the performance metric

PS, for more increases in performance

http://deliddedtech.com/2016/05/07/hidden-information-found-in-source-code-of-nvidias-gtx-1080-page/

they measured out the graph lengths and calculated from them.

80% increased performance for ROTR with a 10% increase in power consumption. That it higher than 2.5x perf/watt

70% increased performance for Witcher 3 which will end up around 2.5x per per watt.

 

Isn't that around 1.7x perf/W?

 

I think you should check your math, seriously. Using your numbers, versus the 980 it's more like 1.5 times perf/w. Versus the 980Ti, it would be 1.85x perf/w. These are solid numbers, but were pretty much expected (due to FinFet). We'll see if actual power consumption figures will mirror those closely (though I think they should).

 

Nvidias own "average gaming performance" graph shows around ~20% performance advantage over the Titan x and ~60% over the gtx980, the ROTR and Witcher 3 scores are clearly outliers.

 

Sure, there is also for example bandwidth which is at 320 GB/s vs 336 GB/s.
Assuming there will be 64 vs 96 ROPs, I can imagine worst case UHD scenarios with 4xmulti-sampling where performance might not even reach GTX980Ti.

 

Its definitely higher than 20% over the titan X, I can say that for sure.

http://deliddedtech.com/2016/05/07/hidden-information-found-in-source-code-of-nvidias-gtx-1080-page/

They are saying 25%

 

OK 25% is most likely accurate. I admit I did not do any pixel counting. A 25% performance increase would result in just under 75% increase in performance per watt.

 

How are you working out those numbers?
According to their site (yeah short on specifics such as resolution,settings but still): http://www.geforce.com/hardware/10series/geforce-gtx-1080
It shows 67% performance gain in Witcher 3, 75% performance gain in Rise Tomb Raider when compared to a 980.
VR results are sillier in a good way.

I would say that is comparing reference to reference; as we know it OCs well this is a fair comparison as it has headroom for AIBs when released in OC models that will be interesting to compare to the OC 980s.
I guess you could be approximating for apples-to-apples comparison but not sure how easy/relevant that is with changes in characteristics between frabrication dies/ideal clock speeds/libraries and allowing for transisters/cores/TDP.
IMO only real apple-to-apple comparison would be matching TDP/watts and seeing the performance difference, and separately benching its OC overhead capabilities to the reference 980.
Cheers

 

OK we will disagree,
but that is not entirely correct in context....
He said about Pascal:

Your now putting this into a context that may be beyond what the whole article emphasised, as again mentioned throughout the article with mainstream and high end.
Mainstream has greater footprint than high end, so of course NVIDIA target is smaller market.
But again, he said high end part,,,,,,
He is differentiating between a high end part and Polaris in the first set of releases from both companies (this is probably for the next 2-4 months as he deliberately ignores the Pascal GP106 product coming out).
This is digressing the NVIDIA Pascal thread so maybe take it to the Polaris one, where it is more relevant for those that want to disagree with an article all about Polaris being good mainstream performance with great efficiency.

Cheers

 

Lol, did you just break NDA?

 

got a link?

 

you're scaring me sebbbi.

 

GCN can run compute and graphics on the same CU (wave granularity). This for example allows you to use idle CU cycles for compute when the graphics shader is sampler bound. And use remaining CU GPRs and LDS for compute tasks when the graphics shader is simple.

However per CU / SM granularity hardware scheduling is MUCH better than static allocation. If Maxwell is indeed only using static allocation, this explains why async compute doesn't bring much gains (and even hurts sometimes).

 

sebbbi - is it just one compute shader and one graphics shader per CU, or can each wave effectively be doing its own thing?

 

The heuristics on wave level scheduling must be insane. How does AMD decide on the number of slots to allocate to graphics or compute?