Nvidia Pascal Announcement

[CSI PC]

hmm maybe they are getting mixed up with a lower end card.... I don't think a 980ti replacement is coming this soon.

I tend to agree, but it is strange for Sweclockers to go with this one as I have not noticed them printing all rumours that come up.
Now HWBattle saying the same thing *shrug*.
It could be that they are just ceasing the 980ti while still manufacturing the Titan X, and some sources are reading too much into it.
That makes more sense to me.
Cheers

 

[Arun]

I think it's worth thinking outside the box and considering possible SM-level differences between GP100 and GP10x (ala GF100 vs GF104). I suspect the finer SM granularity's main benefit (besides maybe yields) is more local/shared memory capacity/bandwidth per ALU, and I get the impression that part of the GPGPU performance uplift for GP100 vs GM200 is the better co-issue of memory operations (including shared memory) and the higher bandwidth to make use of that co-issue. This is probably not as important for gaming workloads (but may still be included if it's cheap or too much work to modify, of course).

Also while NVIDIA doubled the register file size, there is nothing that prevents them from implementing a more gradual increase with a non-power-of-two register file size (the implication from GK210 vs Maxwell is that gaming needs fewer registers than HPC, at least for current workloads, but maybe the Maxwell ratio is getting a bit too low especially with higher clock speeds - hard to say without a lot of lower-level scheduling knowledge, the optimal amount of registers per ALU is *not* constant between architectures).

 

[CSI PC]

Arun,
they still need to improve on their operation more associated with "async compute" in gaming, which the P100 seems to also address to some degree.
Reducing some of what you mention would impact that?

Cheers

 

[CSI PC]

I notice multiple publications are referring to the 12Gb/s GDDR5X as operating at 12ghz.
Technically is it not really operating at 6ghz and utilising the QDR with PLL to double up the data (squeeze into existing format).
Reading specs the DDR and QDR use the same clocks, in fact it is fair to say the clock speed for QDR has to be lower when the GDDR5X is spec'd at 100 and 110 than its DDR compatible mode, only the 120 can operate both QDR and DDR at frequency 1.5ghz.

Only thing changing at its core from what I have read is the doubling of DQ,DBI_n.
CK_t, CK_c,Command, Address, WCK_t,WCK_c,EDC all remain unchanged.

Cheers

 

[Adored]

I have a problem with statements like this Adored. First off Polaris 10 was shown playing Hitman in DX12 at 1440p, at unknown settings, and it was stated it played at 60fp all the time (with frame cap), in a limited part of the game, we don't know what FuryX was doing in that part of the game with a similar settings. Fury X does get around 60 FPS at that resolution at the max settings for the entire game. So we can't draw any conclusions from that demonstration with Polaris 10 because we don't know the settings.

So use the Polaris 11 benchmark instead, that's what I'm basing most of my beliefs on. In your opinion, what is the worst-case scenario for what we saw there, in raw performance terms?

From my perspective this 850MHz Polaris 10 appears to be equal to whatever 950 they used, close to flat-out given what we know about the 950's power draw (around 90 to 100W). If that is the case then at 1.2GHz, Polaris 10 should be GTX 960 level plus another 10-20 percent. Using GTX 960+10%, then doubling resources for Polaris 11 should be another 75% or so performance on top, bringing the card to Fury X level at 1080p. The Hitman benchmark we saw is just another data point pointing to that probability.

Up till recently I would have said that SW:BF was friendly to AMD as well, but given the recent DX12 results it looks almost unfavourable now. I think this is likely to be the deciding factor more than anything - which titles are used for benchmarking.

 

[trinibwoy]

I maybe the Maxwell ratio is getting a bit too low especially with higher clock speeds - hard to say without a lot of lower-level scheduling knowledge, the optimal amount of registers per ALU is *not* constant between architectures).

While higher clocks speeds don't necessarily mean longer pipeline it would be a good reason to increase occupancy via the larger register file.

 

[Benetanegia]

960 + 85% barely scratches 980/390X level performance, let alone 980 Ti/Fury X which are like 40% faster...

And where exactly does that nearly 50% increase in clocks come from? If they could clock it at 1200Mhz, why wouldn't they do just that and compare it to 960 or even 970?

 

[Clukos]

Interesting to see the P100 variants jumping up in TDP from GK100/GM100. Wondering if that means a change in cooling solution as well. The reference Titan X/Titan were running @75-80 C under 100% load. Or maybe more aggressive voltage regulation with GPU Boost 3 or whatever they end up calling it

 

[CarstenS]

From my perspective this 850MHz Polaris 10 appears to be equal to whatever 950 they used, close to flat-out given what we know about the 950's power draw (around 90 to 100W). If that is the case then at 1.2GHz, Polaris 10 should be GTX 960 level plus another 10-20 percent. Using GTX 960+10%, then doubling resources for Polaris 11 should be another 75% or so performance on top, bringing the card to Fury X level at 1080p. The Hitman benchmark we saw is just another data point pointing to that probability.

From what they've shown at the Polaris Preview Day back in December, it was a 1080p setup with 60 Fps lock and medium detail settings. In a recreation of the scene (albeit of course not exaclty, but there's really not much where you can massively deviate there), it turned out that a GTX 950 would have maxed out at round 90 fps in that scene. Now, the comparability of the two entirely depends on how much the Polaris performance had to be reduced in order to reach (down to) 60 fps.

Working theory: The harder it had to step on it's brakes in order to get there, the better for power consumption because more of the power saving features would be able to kick in. But of course, Raja Koduri said that this particular prototype did not even have all it's power saving features enabled.

So, that benchmark tells us basically... that the Polaris-GPU can reach 60 fps in 1080p/medium and the system uses around 90 watt. Not much else.

 

[Adored]

960 + 85% barely scratches 980/390X level performance, let alone 980 Ti/Fury X which are like 40% faster...

http://tpucdn.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/images/perfrel_1920_1080.png

Fury X at 79% vs GTX 960 at 45% (41+10%) = 75% faster.

And where exactly does that nearly 50% increase in clocks come from? If they could clock it at 1200Mhz, why wouldn't they do just that and compare it to 960 or even 970?

If they had wanted to compare it in performance that's what they would have done. They wanted to compare it in perf/Watt. Do you really believe that the 35W Polaris 11 shown vs the 950 will be their fastest entry-level GPU?

 

[Adored]

From what they've shown at the Polaris Preview Day back in December, it was a 1080p setup with 60 Fps lock and medium detail settings. In a recreation of the scene (albeit of course not exaclty, but there's really not much where you can massively deviate there), it turned out that a GTX 950 would have maxed out at round 90 fps in that scene. Now, the comparability of the two entirely depends on how much the Polaris performance had to be reduced in order to reach (down to) 60 fps.

I read about the 90fps for the 950 in that scene but never saw any evidence of it. Did you benchmark it? If that is the case then it really doesn't add up at all given what we know about the 950's power draw being around the 90W+ mark.

But if the 950 had 50% extra FPS headroom then it surely was well under that 90W mark, meaning the system power draw must have been even higher in both cases...which just makes Polaris's Perf/Watt even more spectacular.

So, that benchmark tells us basically... that the Polaris-GPU can reach 60 fps in 1080p/medium and the system uses around 90 watt. Not much else.

It doesn't matter how much it uses in total though - what matters is how much it uses compared to the 950. There is no good result for the 950, regardless of how many Watts the rest of the system is using.

Now I can certainly believe that AMD is putting their very best show on - back then SW:BF was certainly one of their more favourable games - but I can't see any reasonable way to get those power numbers without a clear perf/Watt advantage.

 

[Razor1]

So use the Polaris 11 benchmark instead, that's what I'm basing most of my beliefs on. In your opinion, what is the worst-case scenario for what we saw there, in raw performance terms?

From my perspective this 850MHz Polaris 10 appears to be equal to whatever 950 they used, close to flat-out given what we know about the 950's power draw (around 90 to 100W). If that is the case then at 1.2GHz, Polaris 10 should be GTX 960 level plus another 10-20 percent. Using GTX 960+10%, then doubling resources for Polaris 11 should be another 75% or so performance on top, bringing the card to Fury X level at 1080p. The Hitman benchmark we saw is just another data point pointing to that probability.

Up till recently I would have said that SW:BF was friendly to AMD as well, but given the recent DX12 results it looks almost unfavourable now. I think this is likely to be the deciding factor more than anything - which titles are used for benchmarking.

Worst case scenario, lets take SW:BF which as you stated friendly to AMD at the time it was shown with Polaris, we know limiting frame rates, drops AMD's power consumption greatly even when its a small amount of frames lost at least on Fiji, a 5% loss or cap of frame rates gave as much as 15% in power consumption drop, but the frame rate cap does nothing for power consumption on nV hardware, as they don't have the ability to drop frames when not needed so they are working fill tilt even when frame rate lock is on. We can see this when Fiji is capped vs a 980ti, it does have a performance advantage by 15% and sometimes greater. Now in SW BF, the 950 depending on settings could be well above 60 fps what if its doing 75fps or 90 fps at 1080p with mid or low settings respectively, we don't know what Polaris was getting maybe its at 65 fps, we don't know the settings used, all we know at 60fps the frame rate cap helps AMD conserve power but that is what was AMD's marketing, it has much better perf/watt, based on settings and frame lock. That doesn't show absolute performance without a framelock it might use much more power and that 2 times the efficiency could actually be more like 1 to 1.4 vs a gtx950 and that does not bode well for Polaris as we know the gtx950 isn't the best performance per watt in the Maxwell 2 line up.

End result we really don't know what Polaris will bring to the table in absolute terms because of that frame lock.

 

[CSI PC]

OK this is interesting, more so because of the photo but is it a mock-up by them???.....
Invites also seem to be going on for Pascal consumer launch.
http://pctuning.tyden.cz/component/...p100-pujde-i-do-hernich-grafik-nejen-do-tesel
Photo in link anyway but main news is the launch invites going out:

NVM looking closely does not seem right, excited as it aligned with some early rumours doh
Cheers

 

[Razor1]

that is interesting lol

 

[CarstenS]

I read about the 90fps for the 950 in that scene but never saw any evidence of it. Did you benchmark it?

Yes.

If that is the case then it really doesn't add up at all given what we know about the 950's power draw being around the 90W+ mark.

I disagree for obvious reasons.

But if the 950 had 50% extra FPS headroom then it surely was well under that 90W mark, meaning the system power draw must have been even higher in both cases...which just makes Polaris's Perf/Watt even more spectacular.

Quite the contrary.

It doesn't matter how much it uses in total though - what matters is how much it uses compared to the 950. There is no good result for the 950, regardless of how many Watts the rest of the system is using.

Oh, but it does, really! I explained already why. And FWIW, I don't care about the GTX 950.

Now I can certainly believe that AMD is putting their very best show on - back then SW:BF was certainly one of their more favourable games - but I can't see any reasonable way to get those power numbers without a clear perf/Watt advantage.

They surely will have a Perf/Watt advantaged compared to GTX 950. Question is: How much. And here, the benchmark shown does not give us the answer. It would only do that, if both cards were at their max. performance with 60 fps - which clearly is not the case.
Oh - and AMD's was a prototype at the card level as well, so there should be room for additional improvements.

 

[Benetanegia]

http://tpucdn.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/images/perfrel_1920_1080.png

Fury X at 79% vs GTX 960 at 45% (41+10%) = 75% faster.

And here it's about 55% of a vanilla 390. And 50% of a 390X and 980.
http://tpucdn.com/reviews/Sapphire/R9_390_Nitro/images/perfrel_1920_1080.png

If anything what these charts show is Fury X's poor performance in 1080p (which was already known). The 980 is really close to the Fury X, the 390 is not too far behind, while the 980 Ti is significantly faster. I don't see how that helps your argumenting line tho. Nothing in that represents the real potential performance of the cards. It simply shows that bigger chips tend to scale worse on lower resolutions, a property which could very easily extend to Polaris and its higher ALU counts.

Personally I'd give much more weight to the fact that AMD has been presenting Polaris 10 as "bringing the minimum VR spec to affordable prices". Minimum VR spec is 390/970, so a little over that is what I'd would expect.

If they had wanted to compare it in performance that's what they would have done. They wanted to compare it in perf/Watt. Do you really believe that the 35W Polaris 11 shown vs the 950 will be their fastest entry-level GPU?

So then, that coparison is completely irrelevant. Well, as CarstenS pointed out, at the settings they used the GTX 950 is also faster than 60fps, so it seem no conclusion can be made from that comparison, other than they improved power consumption, which no one is putting in doubt.

 

[Benetanegia]

http://pctuning.tyden.cz/component/...p100-pujde-i-do-hernich-grafik-nejen-do-tesel

OK. My head hurts after trying to read Google's translation of that. Can a charitable soul, who hopefully knows czech, tell us what they're saying there?

 

[silent_guy]

OK. My head hurts after trying to read Google's translation of that. Can a charitable soul, who hopefully knows czech, tell us what they're saying there?

They say that GP100 will eventually make a move from Tesel (Google Translate) to GeForce.

 

[Adored]

And here it's about 55% of a vanilla 390. And 50% of a 390X and 980.
http://tpucdn.com/reviews/Sapphire/R9_390_Nitro/images/perfrel_1920_1080.png

If anything what these charts show is Fury X's poor performance in 1080p (which was already known). The 980 is really close to the Fury X, the 390 is not too far behind, while the 980 Ti is significantly faster. I don't see how that helps your argumenting line tho. Nothing in that represents the real potential performance of the cards. It simply shows that bigger chips tend to scale worse on lower resolutions, a property which could very easily extend to Polaris and its higher ALU counts.

Meanwhile over at Anandtech, The Fury X would only be 66% faster...http://www.anandtech.com/bench/product/1513?vs=1656

I do agree that at 1080p the 960 looks better, but in the context of the Polaris 10 it shouldn't be anything near as big a factor as we're talking about a true midrange chip with a 256-bit bus. Also AMD are benchmarking it at 1440p.

Personally I'd give much more weight to the fact that AMD has been presenting Polaris 10 as "bringing the minimum VR spec to affordable prices". Minimum VR spec is 390/970, so a little over that is what I'd would expect.

Roy's exact words 18:00 -

If we just take it as "cheaper and faster" than minspec VR as he says, he is very likely to be talking about the cut-down Polaris 10 as historically the full midrange GPU would cost more than the 970/390's current prices.

I guess really it comes down to if you believe small Polaris can beat the 960 by 10% or more. If not, for me AMD might as well pack it in and go home.

 

[Ailuros]

OK. My head hurts after trying to read Google's translation of that. Can a charitable soul, who hopefully knows czech, tell us what they're saying there?

I don't speack any czech, but they're rambling something about a chip with yet unknown SP resources (2500-3200), but they consider it for granted that the chip has 8Gbps GDDR5, the same TDP as the 980 and 30% higher performance with a core frequency over 1400MHz. It's extremely hard to understand that translated gibberish, but I have the feeling that even if I'd speak the language I wouldn't think the content makes sense....