You must have read a different post than I did. The post you quoted, and I read, he didn't make any statement against GM. If you are inferring that, well that's your statement to argue (which you clearly are).
He even acknowledge in his very post that it isn't linear (look at Parvo's post telling you that Fiji has more than double the specs of Tonga) that it would still be at least 80%.
The Techreport conclusion was that Fiji would be bottlenecked by ROPs using the Hawaii data set. RS showed that TOnga (closer to Fiji than Hawaii) aready proves this wrong. The data was in Techreport's own articles.
At this point there is nothing to say. You are creating your own argument to fight against, if you can show me exactly where in that post RS made the following claims:
1) it will be faster than GM200
2) it scales asbolutly linearly
Then you got something to stand on. As of now, you created those arguments and are now fighting them.
I provided a data set that shows 980 Ti results to give a point of comparison to Fiji's primary competition. The target is > 36GPixels/sec here. He's even admitting less than perfect scaling (albeit not nearly as conservative as they need to be, as demonstrated by the comparison of 980 Ti to 980 i.e. GM200 to GM204).
I edited my post
I've been responding too quickly here. I'm going to slow down and take time to say what I mean.
Just gonna make a new post, to your edit:
You're the one making that argument. Not RS. You are trying to use a completely different uarch to dictate what happens for another.
Fill rate isn't as simple anymore as clock x rops, you even see that with the results in 980 Ti. So just throw that meaningless calculation out. There are clearly other factors at play.
But you are ignoring the basis of this whole thing by focusing on Nvidia.
Let me try to re-write this
"Product A has 64x and it performs at 64, X=1" TechReport said in 2013
"Product B has 32x and it performs at 66, X=2.0625" TechReport said in 2014
"Product C has 64x, clearly it is going to perform like Product A where X=1, this is not good." TechReport said in 2015
RS statement:
"Why are they ignoring their Product B numbers?"
Let me try again.
When you break it down to its simplest level, RS' claim is that TR's argument about GM200 having a fill rate advantage relative to Fiji is inaccurate. I've provided some results for GM200 from TR in the same test. That 36GPixel/sec number is not a calculation, that's a result for 980 Ti from the table. That is the actual target for Fiji. All I'm saying is Fiji will not reach that number because the actual fill rate for GM200 can exceed 100GPixel/sec depending on clock speed yet its results are a mere 36GPixels/sec, demonstrating some inherent bottleneck which all GPUs throughout history have exhibited in this test.
Therefore, GM200 WILL have a fill rate advantage relative to Fiji.
Let me try again.
When you break it down to its simplest level, RS' claim is that TR's argument about GM200 having a fill rate advantage relative to Fiji is inaccurate. I've provided some results for GM200 from TR in the same test. That 36GPixel/sec number is not a calculation, that's a result for 980 Ti from the table. That is the actual target for Fiji. All I'm saying is Fiji will not reach that number because the actual fill rate for GM200 can exceed 100GPixel/sec depending on clock speed yet its results are a mere 36GPixels/sec, demonstrating some inherent bottleneck which all GPUs throughout history have exhibited in this test.
Therefore, GM200 WILL have a fill rate advantage relative to Fiji.
At this point you are making your own argument, which is fine. If the bold is what you are saying - that's fine. We can't disprove that or anything, yet. We'll have to wait.
However, the Red - where did he make that statement? Please quote me in this thread where RS made that claim.
He made that statement by inference. TR claims that GM200 will have a fill rate advantage over Fiji. He is arguing against that claim.
Dude, where are you getting GM200 references from? This from TR's article:
"In other respects, including peak triangle throughput for rasterization and pixel fill rates, Fiji is simply no more capable in theory than Hawaii. As a result, Fiji offers a very different mix of resources than its predecessor. There's tons more shader and computing power on tap, and the Fury X can access memory via its texturing units and HBM interfaces at much higher rates than the R9 290X.
In situations where a game's performance is limited primarily by shader effects processing, texturing, or memory bandwidth, the Fury X should easily outpace the 290X. On the other hand, if gaming performance is gated by any sort of ROP throughputincluding raw pixel-pushing power, blending rates for multisampled anti-aliasing, or effects based on depth and stencil like shadowingthe Fury X has little to offer beyond the R9 290X. The same is true for geometry throughput." - TR
Which is total BS. Unless they are claiming Fury/Fiji is based on GCN 1.1 and not GCN 1.2 or even GCN 1.3.
I agree TR is wrong on the capability of Fiji's ROPs.
However, that doesn't mean Fiji is going to have more fill rate than GM200.
That's all I'm saying.
I agree TR is wrong on the capability of Fiji's ROPs.
However, that doesn't mean Fiji is going to have more fill rate than GM200.
That's all I'm saying.
And no one was really disputing that.
So the contention is that Tonga ROPs are significantly more powerful than Hawaii ROPs and that reviewers aren't factoring this into their assessment of Fiji. Ok, that would seem to be the case based on the addition of Delta Color Compression, I'll grant that. However, the thing that is lacking from those 3dmark Vantage pixel fill rate tests (an almost 10 year old application, by the way) are results from Maxwell architecture GPUs. So the argument that AMD's ROPs are being under-reported by review sites is an argument that can also be made for NV's modern cards with the data *you've* presented. Now, you can cry wolf all you like about how unfairly AMD is being treated but when you use out-of-date data in order to represent one side of an argument, you seem to be cherry-picking your results to make that argument appear stronger.
Here's the proof:
The only way Fury X can produce a higher result that 980 Ti in this case is if it scales linearly with its functional unit count increase compared to its next closest relative, the R9 285. In this test 980 Ti is more than twice as fast as 290x there, but only about 20% faster than GTX 980, its little brother with 50% fewer ROPS. That tells us there are other bottlenecks here, whether they be within the GPU design, some other system bottleneck (CPU), the application being used for testing this feature, or the API utilized by the application. The bottom line is that we don't see linear scaling on this test, which is *needed* in order for this argument to hold up. Even 980 Ti's result, while impressive, is still only about 35% of its theoretical maximum rate (1GHz * 96 ROPS = 96 Gpixels/sec).
Speculate all you like, but its kind of useless without card in hand to test.
I think it is on purpose.They really cant be that bad.But if they really are that bad they should just quit from GPU review.From your link:
"The most impressive part about the 12K AMD Eyefinity demo shown this week by AMD running the game title Dirt Rally is that they needed just one video card in the PC to push all those pixels at a playable frame rate, the AMD Radeon R9 Fury X. The frame rate was pushing 60FPS, which isn’t too bad considering what the single $649 graphics card is doing behind the scenes to push out all those pixels at an acceptable rate. AMD informed us that two Radeon R9 290X’s or a Radeon R9 295×2 get around 45-50 FPS on this exact setup."
WOW.
Also, I wouldn't worry about ROPs as it's double Tonga's and Tonga's 32 ROPs are > Hawaii's 64. TechReport once again made a giant mistake of comparing paper specs, that is despite themselves being fully aware that Tonga's pixel fill-rate is higher than Hawaii's. They should have known better from HD6970 vs. 7970 or from R9 280X vs. 285 that you cannot compare ROPs on paper in theory from different GPU architectures. Since Fury X is not based on a Hawaii architecture, TechReport's projection about Fury X's ROP bottlenecks is unfounded.
"In other respects, including peak triangle throughput for rasterization and pixel fill rates, Fiji is simply no more capable in theory than Hawaii. As a result, Fiji offers a very different mix of resources than its predecessor. There's tons more shader and computing power on tap, and the Fury X can access memory via its texturing units and HBM interfaces at much higher rates than the R9 290X.
In situations where a game's performance is limited primarily by shader effects processing, texturing, or memory bandwidth, the Fury X should easily outpace the 290X. On the other hand, if gaming performance is gated by any sort of ROP throughput—including raw pixel-pushing power, blending rates for multisampled anti-aliasing, or effects based on depth and stencil like shadowing—the Fury X has little to offer beyond the R9 290X. The same is true for geometry throughput." ~ AMD's Radeon Fury X architecture revealed
vs.
I really don't know what to say -- how do you make such MAJOR mistakes as a professional reviewer?! Wow, just wow, it's like so many of these "professional sites" can't even keep track of their own reviews. If 32 Tonga ROPs > 64 Hawaii ROPs, what happens if you have 64 Tonga ROPs vs. 64 Hawaii ROPs? According to TechReport, the same pixel fill-rate throughput. *Shakes head*.
So according to TechReport, Fury X also has little to offer beyond R9 290X when it comes to geometry throughput. Did Scott Wasson literally forget his R9 285 review or is he assuming Fury X is GCN 1.1 not 1.2? I mean come on, the data in his own reviews proves his own hypothesis wrong and if he wants to go on paper, Fury X should have close 2X the pixel fill-rate and geometry performance of a 290X (even if we account for non-linear scaling, surely 70-80% higher than a 290X).
PCPer, Legit Reviews, KitGuru, HardOCP, TechReport - so many gross mistakes and exaggerations/misconstrued data on the AMD side as of late. It's getting ridiculous.
The point of using NV results is two-fold
1) to offer a more complete data set
2) to demonstrate that NO GPU ARCHITECTURE SCALES PERFECTLY ON THIS TEST
The fact that the data has more up-to-date NV cards doesn't somehow magically become invalid when discussing AMD GPUs results, they're not going to scale perfectly either so even with more powerful ROPs it's not going to smash GM200's results.
PS. That is with the assumption there was no improvements made from tonga to fiji, and HBM doesn't give any advantage in pixel fillrate
Just to go back to Fiji overall performance estimation:
If we take 99% as base, counting only shader count, 4096 vs 2816 (45% larger). Then performance should be over 10% faster than Titan X.
So the contention is that Tonga ROPs are significantly more powerful than Hawaii ROPs and that reviewers aren't factoring this into their assessment of Fiji. Ok, that would seem to be the case based on the addition of Delta Color Compression, I'll grant that. However, the thing that is lacking from those 3dmark Vantage pixel fill rate tests (an almost 10 year old application, by the way) are results from Maxwell architecture GPUs. So the argument that AMD's ROPs are being under-reported by review sites is an argument that can also be made for NV's modern cards with the data *you've* presented. Now, you can cry wolf all you like about how unfairly AMD is being treated but when you use out-of-date data in order to represent one side of an argument, you seem to be cherry-picking your results to make that argument appear stronger.
Here's the proof:
The only way Fury X can produce a higher result that 980 Ti in this case is if it scales linearly with its functional unit count increase compared to its next closest relative, the R9 285. In this test 980 Ti is more than twice as fast as 290x there, but only about 20% faster than GTX 980, its little brother with 50% fewer ROPS. That tells us there are other bottlenecks here, whether they be within the GPU design, some other system bottleneck (CPU), the application being used for testing this feature, or the API utilized by the application. The bottom line is that we don't see linear scaling on this test, which is *needed* in order for this argument to hold up. Even 980 Ti's result, while impressive, is still only about 35% of its theoretical maximum rate (1GHz * 96 ROPS = 96 Gpixels/sec).
Speculate all you like, but its kind of useless without card in hand to test.