Catalyst 10.12 performance review and comparison of three generations of AMD/Nvidia

[Xarick]

I like your article BFG, but i would love to see something to indicate what else has been added from then to now. Also would have liked to see a midway driver like the 10.4s added in, and finally would have liked to see 1920x1080 added as it is one of the most popular resolutions.

 

[Soleron]

Because the competition isn't a 8800GTX either.

The 6970 is a cut down chip to fit die size on the 40nm process. For a new major architecture, that was designed to 32nm, doesn't seem to have the major problems that sometimes happen.

S|A said it wasn't cut down, it just would have been much smaller on 32nm. They also said AMD hit its planned performance target of being faster than the 480 with it. I think AMD never wanted to make a chip larger than Cypress, instead favouring dual-chip cards.

Still, I think Barts-and-a-half (~1728 shaders) would have performed better than the 6970 for the same die size and power consumption. VLIW-4 was a poor choice, seemed to reduce efficiency vs. Barts.

 

[Troll Trolling]

S|A said it wasn't cut down, it just would have been much smaller on 32nm. They also said AMD hit its planned performance target of being faster than the 480 with it. I think AMD never wanted to make a chip larger than Cypress, instead favouring dual-chip cards.

Still, I think Barts-and-a-half (~1728 shaders) would have performed better than the 6970 for the same die size and power consumption. VLIW-4 was a poor choice, seemed to reduce efficiency vs. Barts.

This is why I think drivers are going to improve performance significantly: it is a lot slower per shader than Barts; said that VLIW-4 perform the same as a VLIW-5, I can see Cayman getting at least 15% faster.

 

[3DVagabond]

S|A said it wasn't cut down, it just would have been much smaller on 32nm. They also said AMD hit its planned performance target of being faster than the 480 with it. I think AMD never wanted to make a chip larger than Cypress, instead favouring dual-chip cards.

Still, I think Barts-and-a-half (~1728 shaders) would have performed better than the 6970 for the same die size and power consumption. VLIW-4 was a poor choice, seemed to reduce efficiency vs. Barts.

I'm thinking that there must be some teething pains for the VLIW-4 arch. If not, and it's as good as it's going to get, they'd be better off just scrapping it and upscaling Barts. Might be mem bandwidth too, that's holding it back? I don't think a 1920 barts, for example, would scale with current speed RAM on a 256 mem bus.

 

[BFG10K]

I like your article BFG, but i would love to see something to indicate what else has been added from then to now.

Some of this information is upcoming in a 6850 IQ analysis piece. I did however mention that SSAA has stopped working in OpenGL, which sucks. :thumbsdown:

Also would have liked to see a midway driver like the 10.4s added in, and finally would have liked to see 1920x1080 added as it is one of the most popular resolutions.

Understood, but the results will be reused in future performance tests, so I had to pick settings to make sure the faster cards don’t run too fast.

If you drop all of the resolutions one notch (e.g. go from 1920x1200 to 1680x1050), you’ll find the 5770 generally performs okay in the games I tested.

I always appreciate constructive feedback, so thanks for that. :awe:

 

[Xarick]

Any early advice? I am debating updating the 10.4s to the 10.12s

 

[BFG10K]

Any early advice? I am debating updating the 10.4s to the 10.12s

The changes that have occurred since the 5000's launch are both related to SSAA: it no longer works in OpenGL, and the LOD bias is adjusted automatically now when you use it.

On the 6000 series, additional changes to the control panel have also been made with regards to extra options for optimizations.

 

[Xarick]

The changes that have occurred since the 5000's launch are both related to SSAA: it no longer works in OpenGL, and the LOD bias is adjusted automatically now when you use it.

On the 6000 series, additional changes to the control panel have also been made with regards to extra options for optimizations.

Is this a good thing or a bad thing for us 5k series owners?

 

[Will Robinson]

Interesting results there,why by the way do you have HT turned off...is it significant?

 

[(sic)Klown12]

Is this a good thing or a bad thing for us 5k series owners?

Good. Before, if you didn't lower the LOD, SSAA would make the image a bit more blurry. Now that it does this automatically, you don't have to go the extra step. This also that much better since you could only change the LOD with a third-party tool, such as RadeonPro.

 

[evolucion8]

I think saying that the 5870 was a bigger leap over the 4870, than the 6970 is over the 5870 is sorta a no brainer conclusion. The 6970 and the 5870 are on the same process, obviously it won't have the performance increases that the 5870 had over the 4870. Also, isn't it fair to say that the GTX480 increased over the GTX280 much more than the GTX 580 did over the GTX480?

The HD 4870 was manufactured on the same node process as the HD 3870 and yet, brought 2.5 times more performance.

 

[Lonyo]

The HD 4870 was manufactured on the same node process as the HD 3870 and yet, brought 2.5 times more performance.

That's because the 3870 was just a fixed 2900. It was more like the 9800 to the 8800 than the 5870 to the 4870.

 

[evolucion8]

That's because the 3870 was just a fixed 2900. It was more like the 9800 to the 8800 than the 5870 to the 4870.

Also the GeForce FX to GeForce 6800 Ultra, both were manufactured on the same 130nm process, and yet, the latter was much faster than the former.

 

[Qbah]

That's because the 3870 was just a fixed 2900. It was more like the 9800 to the 8800 than the 5870 to the 4870.

So?

The HD4870 was on the same node (55nm)
The HD4870 was only a bit bigger (33%, 192 vs 256 mm2).
The HD4870 had 43% more transistors (666M vs 956M).
The HD4870 was more than twice as fast.

Apparently ATi really tried and delivered. Now we got a refresh at most (on a new architecture and not really that better with performance / mm2). The die is ~16% bigger on the HD6970 when comparing to a HD5870. So yes, they improved, just not really that much.

 

[Spikesoldier]

So they basically accomplished what AMD had with the 5xxx series, then AMD got similar gains going from 58xx to 69xx as NVidia got going 4xx to 5xx.

TSMC is the culprit.

agreeing that TSMC dropped the ball here. amd had a more robust plan for the 69xx series but had to salvage what they had and make a 32nm blueprint fit on a 40nm node.


1950XTX was on a completely different architecture than the 2900XT. along with all of the video features that was introduced with the new architecture and DX10, the 2900XT and derivative 3xxx series brought unified shaders that could easily be made more robust by adding more shaders later on in design.

the 38xx series was a refinement, a 'tock' process for those familiar with intel's strategy. the 48xx series saw a move from 320 to 800 shaders (2.5x performance increase? i wonder why? because i added 250% more shaders and more memory bandwidth)

you saw the same with the 5xxx series, doubling the shaders from 800 to 1600.

you would think that if amd wanted to increase performance with the 69xx series that they would have added more shaders. but instead, they could not fit all the shaders they wanted to put on it because they had an inferior node to work with than what they originally had planned.

amd also altered the composition of these shaders and they way they work. id say that this change allows it to be more efficient (per shader) than the old architecture found on 38xx 48xx and 58xx series. id consider the 6xxx series to be a 'tock', much like the 38xx series, considering TSMC's 32nm fab failure.

 

[Lonyo]

So?

The HD4870 was on the same node (55nm)
The HD4870 was only a bit bigger (33%, 192 vs 256 mm2).
The HD4870 had 43% more transistors (666M vs 956M).
The HD4870 was more than twice as fast.

Apparently ATi really tried and delivered. Now we got a refresh at most (on a new architecture and not really that better with performance / mm2). The die is ~16% bigger on the HD6970 when comparing to a HD5870. So yes, they improved, just not really that much.

And they managed to fit more than twice as many shaders into that die, 800 vs 320.
That means with a complete redesign they managed to save a lot of space they could dedicate to shaders, and well done to them. That just means it was a pretty wasteful design in terms of die space and transistors.
The HD4670 is 146mm^2, 514m transistors and also 320 shaders, and about as fast as the HD3870 (also 55nm). So they managed to save 150m transistors and 46mm^2 of die space while having the exact same performance, which tells you the HD3850 was pretty badly designed compared to the 4000 series.
With the 5000 series they built off the apparently well made 4000 series, so when it came to the somewhat hacked together 6900 series, it seems they weren't able to do all that much while on the same process.

 

[dguy6789]

edit: misread

 

[BFG10K]

Is this a good thing or a bad thing for us 5k series owners?

It’s a good thing providing (1) they use the right value and (2) you aren’t a hardcore purist that insists on using his own value.

Basically it’ll automatically reduce the blur that RGSS/SGSS can occasionally cause.

 

[BFG10K]

Interesting results there,why by the way do you have HT turned off...is it significant?

I leave HT off because it increases the CPU’s TDP, and it also causes flaky behavior in some games.