RS adamantly declared Kepler's architectural superiority over GCN, even in the face of real-life data that contradicted his claims re: tessellation (actual game runs at HardOCP as well as synthetics), and re: anti-aliasing (even though the evidence is mixed on that front). It wasn't so much about compute..
I didn't mock you and if you feel that I did, I apologize. I still stand by my comments regarding Kepler's superiority in Tessellation and FP16 textures. Kepler is
still ahead of GCN in Tessellation, which is why GTX660Ti did well in reviews. That is the whole controversy pointed out by the author of Tom's Hardware in the comments section (see
Cleeve's post on Page 7, about 8th post from the top):
"First off, I don't use maximum tessellation unless it provides a visual benefit, and doesn't cause too much of a frame rate hit. In the case of the 660 ti and the games we tested, I found it fails on both counts. Nvidia and other sites crank up tessellation, and I believe this gives the GeForce a sizable advantage in a number of tests. Also, I use 8x MSAA if the hardware can handle it. in the case of the 660 Ti, it can in many cases, so we enabled it. Our results suggest that this gives it a bit of a disadvantage compared to the radeons."
His explanation is sound for why he thinks his testing showed HD7950 being ahead of GTX660Ti. If he had instead reverted to using extreme tessellation and relying on FXAA or lower levels of AA, then GTX660Ti would have done rather well since it would have focused on the stronger aspects of Kepler's architecture, while minimizing its memory bandwidth bottleneck.
If a test is done to isolate for extreme tessellation in a tessellation-specific benchmark, Kepler will show its strength against GCN. Kepler architecture is also way faster than GCN is for FP16 integer textures. Both of these areas have been tested independently by TechReport:
However, games would need to use these features to allow Kepler to shine. A benchmark such as Unigine Heaven isn't just isolating for Tessellation or FP16 textures since Unigine Heaven uses various combinations of next generation gaming features, including lighting, bump mapping, etc. So it actually tests the GPU from many different aspects at the same time but doesn't specifically isolate for FP16 or for Extreme tessellation. That allows GCN to perform well under Normal tessellation setting which is what most reviewers use. If you toggle on the Extreme setting, GTX680 shows that it can handle added tessellation better:
If more games use FP16 textures or Extreme tessellation then Kepler would pull away from GCN without AMD's tessellation optimizations in CCC. However, there haven't been any recent games with excessive tessellation and conversely AMD is putting $ behind
DirectCompute.
This is the area where I was wrong and you were right. At that time only Dirt Showdown had DirectCompute and I thought NV would continue to work with developers to put more tessellation into games. Instead it was AMD that pushed DirectCompute harder than NV pushed for extreme tessellation. I couldn't have predicted that at the time, but that doesn't disprove that Kepler topples GCN in Tessellation and FP16 textures. You won't find a real-world game where the gaming world is made of 100% tessellated geometry (like NV's tessellated City Demo), but if you could, I am confident that Kepler would come out on top of GCN.
Many GTX660Ti reviews cranked Tessellation to extreme and left AA at FXAA / 4xMSAA, which is why GTX660Ti with neutured ROPs / memory bandwidth kept up with 7950. The thing is games now are using more than 1 of these next generation graphical features -- AMD benefits from global lighting model, 8xMSAA (due to higher memory bandwidth feeding 32 ROPs), contact hardening shadows and NV benefits from tessellation and lighter levels AA. Because NV is a lot more crippled on the DirectCompute side, it incurs a larger hit than GCN does in tessellation in the same game. This is why GTX680 is slower in Dirt Showdown/Sniper Elite V2. Not sure why you are digging up that thread but I specifically said Kepler has superior performance with MSAA in
deferred game engines and I also specifically sighted BF3. I never said that Kepler is better with MSAA overall (for example I didn't touch SSAA in DX10 games, etc.). Kepler still beats GCN in BF3 at 1080P and GTX670 is faster than HD7950 in that game easily.
GTX660Ti vs. HD7950 is entirely different. That's not Kepler vs. GCN, but a crippled GK104 vs. a barely cut-down Tahiti XT.
If we only look at Kepler vs. GCN architectures, then it's fair thatwe look at the top GK104 card -- GTX680 Lightning -- it is still the fastest single-GPU this round with volt mods. So it is about context. If we are ONLY focusing on architectures and not GPU SKUs (i.e., 660Ti vs. 7950 or 7850 vs. GTX670), then
Kepler architecture is way more efficient than GCN is for games. GTX680, despite only 1536 SPs and 192 GB/sec bus, is able to accomplish what takes GCN 2048 SPs and 264GB/sec+. There is no contest here at all. HD7970 just makes up for its inefficiency with brawn and die size. It's not magic. The only difference is NV bet on FP16 textures and tessellation and AMD bet on DirectCompute. So far AMD is winning in that regard but we are yet to see Medal of Honor Warfighter and Crysis 3. Both of those games can still bring out Kepler's advantage with Frostbite 2.0 deferred MSAA and Tessellation in Crysis 3. Can't write off GTX680 just yet.
In that thread, the OP was concerned for buying a card and keeping it for 3 years. There was no way I could recommend an HD7850 as opposed to the 670 for achieving such a goal, unless you actually truly believe that HD7850 will outlive GTX670 because you think DirectCompute in games will really take off in the next 2-3 years?
Right now the killer feature of GCN is DirectCompute for games. Otherwise,Kepler is still the superior architecture for tessellation and its performance in games on a per mm^2 and on a per shader + memory bandwidth basis is unmatched by GCN HD7000 series in the flagship space.
However, like I said GTX660Ti is not a full-fledged Kepler chip and 25% lower ROP and memory bandwidth hurt it a lot. You can't really use GTX660Ti as a basis to prove that Kepler is weaker than GCN since it a neutered SKU.
blastingcap, what I am most happy about is that through these types of discussions we can talk about certain gaps that exist for both GPU brands and you point out a perfectly valid opposing view to mine (which I think is great since it gives both sides of the coin). Often times these things aren't as covered in detail by the reviewers and someone who wants to "future-proof" (which I think neither you nor I really believe in long-term) their GPU purchase can read those details on the forums, can weigh in what's most important to them (tessellation, frostbite 2.0 performance for NV or DirectCompute / advanced lightning for AMD) and make a better guess regarding which GPU is better for purchase.
Ironically, the choice for what architecture to go for has really had an impact only 3 times in the last 10-12 years of GPU selection I think:
- 9700/9800 series for DX9 vs. FX5800/5900, but one could argue that 9700/9800 weren't fast enough for DX9 games anyway
* maybe GeForce 6 since it could run SM3.0 and X800 series couldn't which I believe blocked X800 series from playing BF2 or BF:BC2 (don't remember which one exactly).
- X1900 vs. 7900 series where 7900 series really struggled in shader intensive games, but then NV responded with GeForce 8 which was a seriously fast architecture for next generation DX9/10 games
- Fermi vs. Cypress where the latter really couldn't handle tessellation in future games well
However, for the most part as each generation passes, the top cards become "equally too slow". I've said it before a couple times
We are seeing that happen with GTX580 vs. HD6970. For example, Sleeping Dogs:
I guess the really smart way to future proof is to buy a GPU just enough for this generation of games and put aside the savings for the next generation of GPUs. Chances are HD7950 OC/GTX670/GTX680/HD7970 will not be fast enough for next generation games, especially once next generation of consoles launch and we may experience a large increase in graphics quality. Most people will probably be happy they saved $100-150 towards a next generation GPU and not have gone for the top $450-500 cards. From that point of view, GTX680/HD7970 GE may become just as slow as GTX670/HD7950 OC. Observing performance of HD6950/6970 vs. GTX570/580, it looks like that exact scenario is starting to happen. It also happened with HD4870/4890 vs. GTX275/GTX280/285.