If it OCes to around 1500MHz, then it will scratch GTX 1070 in some games. But in some games limited mainly by ROP performance, like Farcry Primal, where entire AMD's spectrum from r9-290 to Fury X scores very close to each other while having vastly different TMU count, RX-480 with boosted ROP performance out of box may actually show some teeth. But in general I think it will be mostly scratching back of stock GTX 1070. But at great price difference.
That's a geometry limitation, not rop This is why people are disappointed. Hoping in overclocking headroom as the saving grace for this product is unwise; the AotS benchmark AMD provided are damning; 34 fps single card, 1080p crazy 8xmsaa wasn't it ? A 390X does well above that, around 40-42fps. Either the 480 has lost shader efficiency relative to Hawaii, or the memory bandwidth is holding it back. In both cases overclocking will not "fix it". If the memory bandwidth is holding performance back at 5800gflops, overclocking will simply present rapidly diminishing returns, if any at all. Judging by the AotS bench and comparing with where it lands in the FireStrike ladder, the relieved geometry limitation is the main reason why it outperforms the 390x in it. IF the GCN cards were ROP limited in FC Primal, you would not be seeing a 980, with 64 rops and 2/3 the bandwidth of a 390x, and lower compute performance to boot, outperforming a 390x Spoiler The "390/390x/Fury/Fury X are ROP limited" argument has no legs to stand on at all. Geometry is the bottleneck ! FC Primal doesn't use outrageous amount of tessellation, but polygon counts in games are going up. If there were a ROP limitation you would see the 390X, Fury etc performing worse relative to 980Ti as resolution goes up, instead the delta remains fairly consistent, actually thinning to a mere 4% at 4K.
Does anyone know how does geometry throughput requirements increase with higher resolution which does not change number of polygons? Looking at Fury X vs r9-290x. Fury X has 50% higher Texture processing power, same rasterization power and same geometry throughput. Anyone can guess what's requirement for different resolutions? Dummy can guess that one needs more TMU power at high resolution. Therefore if TMU was limitation in game, Fury X would potentially gain 50% advantage over r9-290x. If ROPs were main limitation, then cards with same ROP count and clock would deliver very similar performance. Which would decline with higher resolution as more pixels on screen requires more rasterization. But what if it was Geometry throughput which decreases performance and is main limiting factor? If game limited by geometry at 1080p delivers 63fps. Going to 1440p or 4k which does not change amount of polygons. And limit fps would be still 63. - - - - So, we are presented with proof that Fury X has damn low polygon throughput compared to Titan X. (No question here, it is true.) But right next to it accompanied benchmark proof from game shows it matching Titan X in all resolutions? - - - - People should think about bottlenecks. At extremely low resolutions, ROPs and TMUs are definitely not limiting factor as geometry is going to limit maximum fps. But as resolution goes up there is point where TMU or ROP becomes limiting factor. How to translate it into ability to see geometry limitation? Having several resolutions with same geometry to be processed. Should not put higher load to geometry processing as it is vector calculation. So, if resolution is decreased (lower load on TMUs&ROPs) and fps stays same, it is point where geometry may be real limit. (Unless it is CPU.) In reverse, if resolution goes up and fps drops greatly, it is not caused by geometry throughput (unless game is made to increase geometry with resolution).
Precisely because polygon amounts are unchanging with resolution, it is apparent that geometry is the limiting factor, just because Titan X has virtually 3x the geometry throughput it does not mean it will necessarily outperform the Fury X, it depends on where the geometry limit falls. The geometry bottleneck could be limiting shader array utilization to 78.5%; incidentally the ratio of TX:FX compute throughput. Odd that TX is claimed to be running at 1100mhz though, I was under the impression stock boost on is around ~1190mhz but w/e. Also, rasterization is the process of calculating pixel coverage from geometry, the number of raster operations increase with resolution, not the amount of rasterization; at high resolution you have will have more pixels being covered by the same geometry, but the actual rasterization work is the same. ROP: Raster Operation Pipeline, not the rasterizer. Rasterizer is on it's own, towards the start of the pipeline, not the bottom. The rasterizer, or raster engine, is NOT the ROPs. Common misconception. Crysis 3 is heavily geometry bound yet Fury X performance decreases with increasing resolution, it's not that simple. Spoiler
^ Yeah I noticed Crysis1 really likes ROP and its pixel fillrate output. 8800>250>570>780.. each gpu had around 2x the old and it scaled almost perfectly by fps diff.
I don't know about Crysis 1 tbh, other than the fact that it's still a better looking game than many games out today Performance wise, I haven't seen a Crysis 1 benchmark in years lol. I'd imagine CPU is the main limitation nowadays
Pixel fillrate was the only thing that progressed slowly with each generation.. I had a 8800 to 250 to 570 with c2q and saw a difference, ok then another with 4770k , but still not enough for my liking, 780gtx finally managed it 1080p v.high and 2-4msaa (2x 570gtx),.. now 980ti 1.4ghz gets great fps at Dsr 1620p and 2xmsaa, still 60fps+ in worst case with lots of foliage or heavy ice particles.
When scientist makes theory. And puts it in test, and it does not survive test. Scientist returns to drawing board. When clergy reads their book of truths, anything even contradiction within single sentence is still proof of God. So, Titan X has 3 times higher geometry throughput, but does not perform better in supposedly geometry limited scenario than much weaker Fury X where impact should be much more severe? And polygon amount is unchanged with resolution, but still is source of same decrease in fps on higher resolution for geometry 'starved' and geometry abundant GPU? "Not that simple." You say? Lets face it: "God works in mysterious ways." Is what you intended to say. And btw. back to topic, it still stands that RX-480 without any improvements to architecture would have same TMU performance as r9-290x, But anything else cloned 1:1 from r9-290x should perform 20% faster due to 20% higher clock. ROPs, rasterize, VCE, ... whatever comes in mind. So, baseline performance of r9-290x and up to additional 20% in games which benefit from other parts of GPU than TMUs(SP) + additional performance from improvements of each block. As -Tj- said, if card doubles in certain type of throughput and gaming performance goes up considerably, then it was limiting factor. If card doubles different type of throughput and performance is practically unchanged, then it was not limiting factor. If you double workload to certain part of architecture and performance drops accordingly, then it is limiting factor. If you keep workload for certain part of architecture unchanged, but workload for other changes up/down and performance changes too. Then this one unchanged type of workload was not limiting factor, those other factors affected performance.
If that RX470 is actually able to keep up with an R9 290(390) in the 150-200 price range; howly **** !! Team green is going to have a big problem ...
You guys dont forget that HH didnt update scores for 280X and 290/290X for a very long time. If you notice in those charts 280X have 7.3k (im getting 8.3k now) thats how much AMD has improved. So those 9.1k for 290 is an old news. I highly doubt 470X will be trading blows with 290, no way.
You have not read the end notes.. the test was done on 16.1.1 driver meaning the 470 also could have improved greatly
These rumours are crazy. Now some seem to think that this will match GTX 1070. Some probably are going to be extremely disappointed when they see that it's (480, 470 obviously slower) just a much more efficient R9 390.
For $200? Sounds great. A 390 is slightly worse then a 970 right? A 970 for $200 is value. I do think all these benchmarks are hype though. Why haven't AMD figured out that people aren't going to commit until they see 3rd party benchies?
Which honestly isn't too bad. If it runs inaudible and sub 70C max temp and overclocks better than the last couple of generations it could be a nice little card. To be fair, if I was on a 680/770/780 or 7950/7970/280x and similar cards I would definitely consider this as a cheap boost. But that guy shouting about snapping at the heels of 1070 is a bit deluded. Not even close.
Since there is currently no RX-470X, You are quite right But as RX-480 should match r9-390x from paper perspective, RX-470 which is cut down of RX-480 may quite often compete with r9-390. It all depends how much it is cut down. As r9-290/390 is not cut down much. I would love to see RX-470 cut down to 2048SP(32CU), so we can do direct clock to clock comparison to r9-380x. That would be lovely exploration of changes AMD made.
I think he mentioned family member who had/has 1st hand experience. It is quite possible. Here, if you want it badly and know someone at some store where they do not care much about official date, you can have card already. But I think he is bit too enthusiastic with OC potential.
It would be pretty good value for the money indeed. It's just hilarious when some expect 200 € card beat Fury X or similar NVIDIA card.
No games yet though, there are benchmarks though You won't like em. 390 DX11 vs DX12 https://www.youtube.com/watch?v=JZUsAmXb3Jo 780Ti DX11 vs DX12 https://www.youtube.com/watch?v=4dOe0zxC7ZQ
