AMD Radeon R9 Fury X Reviews
@pjbliverpool said:
So will we see a Titan X overclock from NV now billed as a 990Ti/Titan X Black at a more competitive price? It would seem to be within their capability to do so and would have a reasonably clear lead at the top of the performance table from what I'm seeing.
So will we see a Titan X overclock from NV now billed as a 990Ti/Titan X Black at a more competitive price? It would seem to be within their capability to do so and would have a reasonably clear lead at the top of the performance table from what I'm seeing.
@Rurouni said:
The card balance is way off for current games. CU and bandwidth looks like underutilized. Nvidia seems to hit the best balance, at least for current games. Maybe a true DX12 games could change the balance, but right now it's kinda underwhelming.
What I meant by true DX12 games is that a game that built only for DX12 without thinking of compatibility with DX11 and below (unlike current Mantle games).
The card balance is way off for current games. CU and bandwidth looks like underutilized. Nvidia seems to hit the best balance, at least for current games. Maybe a true DX12 games could change the balance, but right now it's kinda underwhelming.
What I meant by true DX12 games is that a game that built only for DX12 without thinking of compatibility with DX11 and below (unlike current Mantle games).
@CarstenS said:
Unfortunately, there are already factory overclocked 980 TI's in the market, that have a base clock of 1.253 MHz, i.e. 25% over reference speed. :|
@silent_guy said:
The only way I can explain Fiji's existence in its current form is by assuming that AMD got complacent by the performance of Kepler, more or less the same as GCN but already touching reticle limits with GK110, and being massively blindsided by Maxwell. They probably thought that just touching the memory system would be sufficient to soar about whatever Nvidia could come up with and left the efficiency of GCN what it was.
While interesting as an HBM marketing demo, and good enough to convince a bunch of AMD enthusiasts, the most appropriate slogan for Fiji is probably "It sucks less than R600."
The only way I can explain Fiji's existence in its current form is by assuming that AMD got complacent by the performance of Kepler, more or less the same as GCN but already touching reticle limits with GK110, and being massively blindsided by Maxwell. They probably thought that just touching the memory system would be sufficient to soar about whatever Nvidia could come up with and left the efficiency of GCN what it was.
While interesting as an HBM marketing demo, and good enough to convince a bunch of AMD enthusiasts, the most appropriate slogan for Fiji is probably "It sucks less than R600."
@fellix said:
For a first generation of a brand new memory implementation, this thing performs just fine, considering the huge risk AMD is investing itself in. It's probably a better decision to move now with what they have, than to wait for the more robust and mature HBM2 that is cooking right now, since apparently AMD is not yet ready with a new architecture and much less about new node process to enable it.
For a first generation of a brand new memory implementation, this thing performs just fine, considering the huge risk AMD is investing itself in. It's probably a better decision to move now with what they have, than to wait for the more robust and mature HBM2 that is cooking right now, since apparently AMD is not yet ready with a new architecture and much less about new node process to enable it.
@pjbliverpool said:
So according to TPU which usually has the most comprehensive reviews IMO, the Fury X is slower than the 980Ti at every resolution below 4K and only matches it at 4K (kind of). It's less power effiient, but quieter, it's smaller but has less VRAM, it's cooler but potentially doesn't overclock as well. When you throw in the general advantage NV as in drivers and game support, plus the full FL12_1 implementation, I really don't see how this can justify the same price point as the 980Ti.
Still, I hope it sells well. God knows we all need AMD to prosper.
So according to TPU which usually has the most comprehensive reviews IMO, the Fury X is slower than the 980Ti at every resolution below 4K and only matches it at 4K (kind of). It's less power effiient, but quieter, it's smaller but has less VRAM, it's cooler but potentially doesn't overclock as well. When you throw in the general advantage NV as in drivers and game support, plus the full FL12_1 implementation, I really don't see how this can justify the same price point as the 980Ti.
Still, I hope it sells well. God knows we all need AMD to prosper.
@kalelovil said:
Either the GPU packages AMD received back from manufacturing exhibited potential clock rates significantly below their expectation and that information was not transmitted to the marketing department in time, or there was deliberate misrepresentation on the 16th.
It is somewhat bizarre. Reviewers seem to be lucky to eek out 5%more performance from overclocking.
Either the GPU packages AMD received back from manufacturing exhibited potential clock rates significantly below their expectation and that information was not transmitted to the marketing department in time, or there was deliberate misrepresentation on the 16th.
@silent_guy said:
Right now, what HBM bought them was lower power and more compactness. (And higher cost.) Yet they are still 25% behind gm200 in perf/W, despite using water cooling which definitely helps keeping the temperature dependent leakage current in check. HBM is only used as a band-aid to hide the deficiencies of the rest of the architecture.
I don't want to diss AMD HBM engineers: they deserve huge kudos for making HBM happen. But I can imagine them being just a little bit upset for the rest of chip not pulling its weight.
It only performs just fine if you restrict that to "it's able to read and write bytes to memory." I have yet to see a single convincing result that shows that the BW of HBM provides a real performance advantage and I'm wondering if a hypothetical Fiji with a 512 bit GDDR5 interface would have performed any worse.
Right now, what HBM bought them was lower power and more compactness. (And higher cost.) Yet they are still 25% behind gm200 in perf/W, despite using water cooling which definitely helps keeping the temperature dependent leakage current in check. HBM is only used as a band-aid to hide the deficiencies of the rest of the architecture.
I don't want to diss AMD HBM engineers: they deserve huge kudos for making HBM happen. But I can imagine them being just a little bit upset for the rest of chip not pulling its weight.
@pjbliverpool said:
Considering the reviews seem to only be getting around 10% more clock speed out of Fury X, that doesn't bode well.
Considering the reviews seem to only be getting around 10% more clock speed out of Fury X, that doesn't bode well.
@silent_guy said:
That's an interesting graph. I think what we're looking at here is not worse color compression, but a different bottlenecks coming into play. When using color compression, it's hitting some non-memory BW bottleneck somewhere (one that's very similar than TitanX), while without compression, it gets much better results than TitanX due to the HBM that's provided by the extra BW.May be AMD should have gone with more pixel throughput for Fiji this time. There's still leftover BW for that to burn in.
p.s.:
Looks like Fiji's frame-buffer compression is only a bit better than Kepler's:
Now I revise my suggestion -- AMD should have put a more aggressive colour compression instead.
@Alexko said:
R600 had a process advantage over G80 (80nm vs. 90nm) but was more power-hungry and much slower. The 8800 Ultra was about 40% faster with anti-aliasing.
Here Fiji is a bit slower at 2560×1440 and roughly comparable at 4K. I suspect it would be faster at 5K or multi-display 4K, provided it didn't run into capacity constraints. All in all it's not great but a very acceptable product.
That said, since there was apparently talk of a KingoftheHill GPU at AMD for a while, I think it's fair to say that Maxwell was somewhat unexpected.
R600 had a process advantage over G80 (80nm vs. 90nm) but was more power-hungry and much slower. The 8800 Ultra was about 40% faster with anti-aliasing.
Here Fiji is a bit slower at 2560×1440 and roughly comparable at 4K. I suspect it would be faster at 5K or multi-display 4K, provided it didn't run into capacity constraints. All in all it's not great but a very acceptable product.
That said, since there was apparently talk of a KingoftheHill GPU at AMD for a while, I think it's fair to say that Maxwell was somewhat unexpected.
@Razor1 said:
Well after the 750 ti was launched, they should have had a good idea of what was coming up, but at that point it might have been already too late.
Well after the 750 ti was launched, they should have had a good idea of what was coming up, but at that point it might have been already too late.
@kalelovil said:
It's come a long way from the 290X/390X (and yes, its largely due to HBM), and is within a margin now where I doubt it will affect potential purchasers.
It's not that bad. The average power consumption from the reviews in the OP is only 10% above the 980 Ti and performance is roughly the same.
It's come a long way from the 290X/390X (and yes, its largely due to HBM), and is within a margin now where I doubt it will affect potential purchasers.
@Lightman said:
So far read TR and TPU reviews. Not bad for a 99 (£400) card, but a bit weak for 49 card, even considering AIO and compactness and HBM.
So this is my first AMD top end single GPU card in a while (famous R600
) I will not rush to buy, maybe even skip completely. Some may say, it's close to 980Ti, but not that close at 1440p I game at, and too close to Hawaii for my liking! Sometimes few % can make or break deals!
So far read TR and TPU reviews. Not bad for a 99 (£400) card, but a bit weak for 49 card, even considering AIO and compactness and HBM.
So this is my first AMD top end single GPU card in a while (famous R600
) I will not rush to buy, maybe even skip completely. Some may say, it's close to 980Ti, but not that close at 1440p I game at, and too close to Hawaii for my liking! Sometimes few % can make or break deals!
@silent_guy said:
I assume so. But in terms of pure BW, 64 ROPs should be sufficient to cover the full BW.
