28-Way NVIDIA GeForce GPU Comparison On Ubuntu: From GeForce 8 To GeForce 1080

[andrei_me]

I love this kind of whole decade of graphics tests, this way I can imagine which kind of jump in performance I would get replacing my +7 years old computer

[leipero]

What i was suprized about were those "retro" GPU's, mainly 8800GT (256MB) vs. 9800GT (512MB, with semi-significant clock difference), since i owned both in the past, and there was literally no difference between them even with double of VRAM at that time, even more suprizing is that GT 520 often outpreformed 8800GT. I know about the thing "green team doesn't age well", but this was suprizing even with that on mind. I know you mentioned similar with pre-GCN included, looking forward for that.

[schmidtbag]

What i was suprized about were those "retro" GPU's, mainly 8800GT (256MB) vs. 9800GT (512MB, with semi-significant clock difference), since i owned both in the past, and there was literally no difference between them even with double of VRAM at that time, even more suprizing is that GT 520 often outpreformed 8800GT. I know about the thing "green team doesn't age well", but this was suprizing even with that on mind. I know you mentioned similar with pre-GCN included, looking forward for that.

Despite the VRAM differences, the performance levels of those GPUs seems suspiciously bad. The 240, for example, is only a couple years newer than the 8800 and it outperforms that in almost every test. The 8800 is pretty much the reason why Nvidia took the market from ATI and kept it ever since; I don't see how that product could perform so horribly.

[leipero]

Despite the VRAM differences, the performance levels of those GPUs seems suspiciously bad. The 240, for example, is only a couple years newer than the 8800 and it outperforms that in almost every test. The 8800 is pretty much the reason why Nvidia took the market from ATI and kept it ever since; I don't see how that product could perform so horribly.

Yeah that was my point, 8800GT and 9800GT are the same GPU, same core, same number of shader units, even same stock clocks (review samples have different clocks, 9800GT have advantage). It's most likely due to the RAM, it wasn't big difference back then, but it seems it is now, or simply drivers are cripled for 8800GT somehow.

[khnazile]

Yeah that was my point, 8800GT and 9800GT are the same GPU, same core, same number of shader units, even same stock clocks (review samples have different clocks, 9800GT have advantage). It's most likely due to the RAM, it wasn't big difference back then, but it seems it is now, or simply drivers are cripled for 8800GT somehow.

It's not only VRAM size, but also bandwidth, older GPUs used slow GDDR3 VRAM. The tests that would show real retro-GPU limits should probably made with lower resolution (1366x768, 1280x1024) and reasonable texture sizes. And no anti-aliasing, of course.

[leipero]

It's not only VRAM size, but also bandwidth, older GPUs used slow GDDR3 VRAM. The tests that would show real retro-GPU limits should probably made with lower resolution (1366x768, 1280x1024) and reasonable texture sizes. And no anti-aliasing, of course.

Both 8800GT and 9800GT use GDDR3, while it might be true that GT 520 (I've mentioned earlier) for example use GDDR5, bus width of 8800/9800GT is 256-bit, while 520 most likely is 64-bit, so it's effectively less memory bandwidth. But yes, you are right, lower resolutions might play some small role, but the problem is nature of games, games are usually optimized to use X amount of memory on specific settings (low, medium, high etc.) regardless of resolution (it makes insignificant difference in most cases). So it's really interesting to see such difference between same GPU's with similar memory bandwidth and clocks but double the difference of VRAM preform so drastically different, it was just interesting, it's not really relevant for the whole testing and in general.

[wdb974]

Both 8800GT and 9800GT use GDDR3, while it might be true that GT 520 (I've mentioned earlier) for example use GDDR5, bus width of 8800/9800GT is 256-bit, while 520 most likely is 64-bit, so it's effectively less memory bandwidth. But yes, you are right, lower resolutions might play some small role, but the problem is nature of games, games are usually optimized to use X amount of memory on specific settings (low, medium, high etc.) regardless of resolution (it makes insignificant difference in most cases). So it's really interesting to see such difference between same GPU's with similar memory bandwidth and clocks but double the difference of VRAM preform so drastically different, it was just interesting, it's not really relevant for the whole testing and in general.

Yeah, this sounds like a RAM issue to me as well. I had a 8800 GTS (G80) back then, and it only had 640 MB of RAM, so I couldn't play Doom III on max settings. BF3 would be playable on medium settings (though I'd set them to 'low' because I prefer smoothness over eye candy), but flying a chopper would reveal that far away terrain was totally black.

That was in 1680 x 1050 screen resolution, by the way.

[DanL]

even more surprising is that GT 520 often outperformed 8800GT.

Keep in mind that the 8800GT is using the legacy 340.xx driver, which may not have optimized profiles for newer games.

[kaseki]

As always, thanks for this testing program, in both senses of the word.

I decided to take you up on duplicating the test on my new HTPC build. Lets say it was an interesting experience, in the Chinese proverbial sense. As a noob I didn't realize that many tests required a Steam account, and once that account has been established, it seems the tests still won't run unless one has bought(?)/rented(?) the relevant games. Does this imply that every time you run a battery of tests over many cards another small country's GDP moves from your wallet to Valve's wallet? Or is there some trick that causes the cost to be trivial due to the short time that each program is utilized. I feel that I am missing some set-up process that I didn't see described when I looked over the documentation for PTS.

[Michael]

As always, thanks for this testing program, in both senses of the word.

I decided to take you up on duplicating the test on my new HTPC build. Lets say it was an interesting experience, in the Chinese proverbial sense. As a noob I didn't realize that many tests required a Steam account, and once that account has been established, it seems the tests still won't run unless one has bought(?)/rented(?) the relevant games. Does this imply that every time you run a battery of tests over many cards another small country's GDP moves from your wallet to Valve's wallet? Or is there some trick that causes the cost to be trivial due to the short time that each program is utilized. I feel that I am missing some set-up process that I didn't see described when I looked over the documentation for PTS.

Unfortunately, not much of a viable solution for Steam game benchmarks... You must own the game in nearly all of the cases except for like TF2 where it's free-to-play. In the "good old days" it was much better for benchmarking when not having to deal with Steam / the Steam runtime and it was also that period when many games shipped binary demos, which also tended to be good enough for benchmarking. But these days pretty much all modern Linux games require Steam and will only work for benchmarking if owning the full game.

I only use one Steam account for all of my test systems, just store all the games mostly on one SSD and if running many tests concurrently, just needing to run in offline mode so then it will let me run the Steam game tests multiple times.

Anyhow, for your purposes, if you just ignore the Steam game tests or those that requriing you own the games, this result file still has game tests like ET Legacy, Xonotic, Tesseract, and others that don't require Steam or a game license.