It is quite funny because you have posted that table in this very thread
http://forums.anandtech.com/showpost.php?p=38146747&postcount=1065
64 Cores from Pascal Arch have exactly the same performance clock to clock as 128 Cores from Maxwell and 192 from Kepler.
Yes, you are correct. But IPC performance of Maxwell Cores was bigger than Kepler due to the increased amount of work they were able to do. The same thing is for Pascal, but in much higher ratio than Kepler.
64 cores can do exactly the same amount of work that 192 Kepler cores, and 128 Maxwell Cores. It is in that table .
Edit: To clarify, because the scheduler & warp scheduler remains the same, the max thread per SM (which has 2x Warp Schedulers each SM) remains the same. You can queue up the threads to that limit and the warp scheduler will send it to the CC to process over time. It's not an instant "send 2048 threads" thing and have the CC magically finish all those threads at the same time with 64 CC vs 128 CC in Maxwell.
Thats because, what Mahigan pointed already out: Maxwell Cores are starved for Bandwidth. Secondly - scale.
2880 CUDA core GTX 780 Ti is 50%(sometimes less, sometimes more, depending on the game and... drivers) slower than reference GTX 980 Ti which has 2816 CUDA cores. Both have identical bandwidth, but core clocks are different and they differ by the amount of performance cores can output. Cores of different architectures are bringing advancements. Jump from 128 Cores of Maxwell to 128 Cores Pascal is exactly 2X. The question is, will they be fed enough. Core clocks will be high, very high. But the ROP amount in GP104 is worrying: 64 ROPs, and GDDR5 and GDDR5X memory. And we already know that this may not be enough, by looking at Maxwell. Thats why Nvidia went with HBM2 for their GP100.
Overall performance increase between 2048 Maxwell CUDA GPU and 2048 CUDA Pascal GPU will not be 2X. But it may end up between 30 and 50%. And that is quite a feat, to be honest.
In your example, 780Ti vs 980Ti, there is again a huge clock speed deficit on the 780Ti.
~25-30% clock speed gains on the 980Ti and extra vram which has an effect in many games from 2015 onwards. When you work it out, each Maxwell CC is effectively x1.2 to x1.3 Kepler CC only at the same clocks. Hence the 20-30% IPC.
If you want to claim Pascal has 30-50% IPC, with its ~20% clock speed advantage, GP104 will be 50-70% faster than Titan X.
Now, compare the last node shrink and uarch change, 580 -> 680. The 680 was ~25-30% faster. But the 580 chip itself was compute heavy and so it suffered perf/mm2 and perf/w, thus the 680 being a gaming focused chip, has even a better handicap.
This time, Titan X is already beast mode for gaming with gimped compute. It's already a lean-mean chip.
There's zero chance of GP104 being 50-70% faster. That's more for GP100. Because GM200 is already a gaming focused chip, I would say this time around, the delta will be potentially less than the 680 vs 580 comparison.
GP104 full ~Titan X + 20% is a good result on a small chip with the power savings.
We received independent confirmation of the invites and they were actually handed out at the time before the GTC Keynote. The Nvidia Editor’s event will be the official press launch of for the new lineup of 16nm Geforce GPUs by Nvidia. It is not yet known at this point whether Nvidia will allow the event to be covered live or it will have an NDA to be revoked at the time the reviews hit. At any rate, we expect reviews of the new GPUs to be here by mid-May.
2880 CUDA core GTX 780 Ti is 50%(sometimes less, sometimes more, depending on the game and... drivers) slower than reference GTX 980 Ti which has 2816 CUDA cores.
I did not said it was 50% slower . I said that GTX 980 Ti is 50% faster than GTX 780 Ti, which exactly relates to 25-30% slower GTX 780 Ti . The matter of point of view .
Performance gains will be bigger however in IPC on Pascal, because of the scale. 192 vs 128 Cores. 128 vs 64 Cores. See my point?
No, you said the 780Ti was 50% slower than the 980 Ti. antihelten even quoted where you used those exact words.
I did not said it was 50% slower . I said that GTX 980 Ti is 50% faster than GTX 780 Ti, which exactly relates to 25-30% slower GTX 780 Ti . The matter of point of view .
Performance gains will be bigger however in IPC on Pascal, because of the scale. 192 vs 128 Cores. 128 vs 64 Cores. See my point?
Eerm no you did not say 50% faster, you said 50% slower, try going back and reading your own post.
Yes I see your point, but there is no logic to it, I still don't think you understand why Nvidia went from 192 cores in Kepler to 128 cores in Maxwell.
Reducing the number of cores per SM is not some magical method to increase IPC, it was done to achieve a better balance of functional unit and thus throughput since Kepler had poor throughput. It is basically the same reason why AMD switched from a VLIW5 design to a VLIW4 design back in the days. If Nvidia did their job properly there is no reason to believe that Maxwell requires additional balancing of functional units to achieve proper throughput.
Most likely Pascal cut the SM of Maxwell in half whilst keeping the registry file the same size because that was the easiest way of ensuring that each SM had access to a registry file twice the size (the alternative way would have been to keep the SM the same size as in Maxwell and then simply doubling the size of the registry file). The reason why Nvidia wanted double the registry file size is probably related to compute, not gaming.
After thinking about it - you are right. And you know with what is it related?
Unified Memory.
Nvidia launching Pascal Geforce lineup at editors event before Computex 2016 (mid-May)
Nvidia launching Pascal Geforce lineup at editors event before Computex 2016 (mid-May)
http://wccftech.com/nvidia-launch-pascal-editors-event-may
Reviews in a month?
Are they "launching" the Pascal lineup or simply "announcing" the Pascal lineup.
Big difference.
Are they "launching" the Pascal lineup or simply "announcing" the Pascal lineup.
Big difference.
Hey, it could happen, but you seem to be avoiding answering me. Probably because you know how stupid the answer sounds.
I'll make it easier for ya, since you have a hard time understanding a simplistic example...
Say a company develops a GPU that is 150% the performance of anything out there today, but can do so at the same cost. Why, exactly, do you feel they should charge the same price?
Ready... GO!
That scenario isn't just realistic, it's downright inevitable. It's just a question of time, wont happen on 16nm obviously, but 10nm should get close.
Great! You're calling me stupid. lol I'm crushed. How will I sleep at night?