Geforce Chipset and CUDA

[Mintmaster]

Only if you go with 790GX and a Phenom or that new GF9300 IGP. Maybe then it would be like a 9700 Pro. I'm not entirely sure about that though. My 780G + A64X2 isn't there. It has an unstable framerate due to issues with CnQ, and the reduced HT bandwidth with a A64X2. Even games like Elite Force 2 and NFS Hot Pursuit 2 are too much for it at 1680x1050.

But are you sure that R300 is significantly faster? Some googling shows me that a Ti4600 can only comfortably manage 1024x768 @ 16bit, and in most games it wasn't much slower than R300 w/o AA. From the limited info I have, I doubt R300 can manage 1650x1080.

Anyway, I'm quite excited about what ATI will have in store for us with its next IGP. Chipsets are usually pad limited, right? 780G is 200M transistors on 55nm, so ATI will be able to pack some serious power in there at 40nm. BW will be limited, but even 640x480 with efficient 4000 series AA/AF and 80-120 SPs to handle shaders at the highest setting...

That'll be awesome in a low power 12" notebook. Who cares if I have to settle for Phenom instead of Core 2, or low voltage AthlonX2 instead of Atom Duo or whatever. I'm buying that as soon as it comes out.

On a related matter, I wish people would test AA/AF more often in IGP reviews. In low res it's more important, not less. 640x480 w/ 4xAA/8xAF looks waaaaay better than plain 800x600, especially when upscaled on a LCD screen.

 

[Pressure]

But are you sure that R300 is significantly faster? Some googling shows me that a Ti4600 can only comfortably manage 1024x768 @ 16bit, and in most games it wasn't much slower than R300 w/o AA. From the limited info I have, I doubt R300 can manage 1650x1080.

I remember gaming at 1600 x 1200 @ 32-bit with my Radeon 9700 Pro in titles like Dark Age of Camelot.

 

[swaaye]

I think that "HD 3200" has 9700 Pro beaten on shader code, but when it comes to texturing, it's not so hot. Partly due to serious bandwidth constraints, partly because it has had its caches gimped. It has a lot less bandwidth than 9700 Pro, even if you run it with a top Phenom that has the 2GHz HT bus. In particular, I've found that it has trouble with layers of alpha textures. For example, the tree leaf texturing in Guild Wars drags it way down noticeably. I also tried Doom3 on it and found that it really couldn't play it adequately (very unstable framerate). A 9700 Pro absolutely plays Doom 3 better.

The magically "high" 3Dmark06 score that they brag about sure doesn't tell you all that much about what these IGPs are actually capable of. I wouldn't be surprised if an X800 could be beaten in that bench by 780G, but any X800 would dust 780G in every game.

I think it would be safe to think of 780G as a 9600 Pro/XT for games. I have a Mobility 9600 in one of my notebooks and 780G is faster than that. But it's not a 9700 Pro. One day I may get a 45nm Phenom for the board and see what that does for it. I think it could make a decent difference. I just have a hard time seeing that as a viable option for a "budget" game rig. You could do unbelievably better with a cheap mobo/CPU + almost any discrete GPU.

 

[ben6]

Rage 128 was as really as good as its contemporaries thoug, aside from drivers. Maybe it had similar 2D compared to Mach 64, but I honestly rather doubt that. It had DVD acceleration, for example, and Mach64 sure did not have that. Neither did G400 for that matter! (And maybe TNT2 too?)

IGPs today are not all that different from a performance perspective relative to their ancient ancestors. They still suck for games. Nearly useless for anything that's newer than, say, 2005. (780G tweaker/overclocker here) For example, the i810 IGP was not vastly different than a Voodoo2 released only ~two years earlier. There were boards with onboard TNT2s. I'm not sure I recall onboard Rage128s.

If you want to play any semi-recent 3D games on your machine at a res above 800x600 with low detail levels, don't go with an IGP. It's as simple as that. There are gobs of ex-hotstuff 3D cards out there that will dust them and can be had for cheap. Even brand new stuff is cheap now, with how crowed the sub $100 range is.

On the other hand, my 780G is great for everything that's not 3D. It eats up H264 and VC1, for example.

CUDA isn't going to magically run better on an IGP than games do. You still have hardware equivalent to the lowest of the low in recent GPUs, strapped to slow, shared RAM interfaces.

Err I didn't mean to downplay the Rage 128 as the discrete version had excellent 32-bit color and performance but the fact was the Rage II/Pro Turbo lasted well into that era as a integrated graphics solution and that wasn't fun I got to tell you. Ya know in this day and age where you can buy a HD 4870 512MB for $205 (including tax) or a HD 4850+2GB DDR2 memory for $118 (after rebates) there's never been a better time to buy highend video cards.

 

[Mintmaster]

I remember gaming at 1600 x 1200 @ 32-bit with my Radeon 9700 Pro in titles like Dark Age of Camelot.

Oh yeah, me too, but I'm talking about NFS HP2, which swaaye brought up. That's where a Ti4600 can only manage 1024x768, according to firingsquad.

swaaye, in texturing 780G is right on the 9700's heels. It's only alpha blending that it lags in, and that's due to BW. I also read an article of 780G being overclocked at 950 MHz instead the stock 500 MHz, which is pretty impressive. You're not going to get a 90% overclock on R300.

I know that 780G isn't really a gaming solution, but if you recall I was just pointing out how 6 years is all it took for a particular performance level to go from enthusiast to IGP. As long as scaling doesn't stop at 32nm, I have no doubt we'll see 4850 performance in an IGP in a similar amount of time.

 

[swaaye]

I've run my 780G at 950 MHz. It's not quite stable there though, but is at 930 MHz. The problem is, as you know, bandwidth. I'm sure it would be better than R300 if it had 15GB/s RAM bandwidth. But it sure doesn't so it definitely isn't.

Also annoying for me is that 780G and A64X2 don't get along very well. For best performance, you have to disable CnQ because otherwise framerate will vary dramatically, presumably due to the CPU changing speed even in the game, affecting RAM bandwidth. It causes game speed to literally slowly pulsate slower/faster. And I have problems with CnQ locking at full speed after video playback. This is fixed by disabling the hotkey service though. This nonsense has really soured me to this hyped up IGP. One would think that a board like 780G would be home to many more cheap K8 CPUs than Phenoms and expect AMD to be sure they worked perfectly together. Nope.

 

[swaaye]

I actually have some 3DMark2001SE results for both my 780G when it was super overclocked and of my old 9700 Pro on a Athlon XP-M. Both CPUs at 2300MHz. I realize it is 3DMark, but the results still tell you something about 780G's capabilities and bandwidth issues.

780G IGP at 930 MHz. I believe the FSB on my 780G board was 328 MHz for this test. So HT was a crazy 1640MHz. That's the highest I've ever managed to go with the board. CPU clock is so low because the thing won't POST at FSB > 300 Mhz if the multiplier is above 7x. Probably due to the RAM divisor changing and causing unhappyness. I don't recall RAM clock, but it was way above 800 MHz for certain. Perhaps ~1000 MHz because I usually need to leave RAM speed on auto to get it to POST at uber high FSB.

Especially note how the Nature test implodes 780G. I blame the layered alpha texturing. This happens in real games too and is why 9700 Pro generally outclasses 780G, in my experience.

Fun w/ unified shaders in those synthetic geometry tests.

Green = 780G
Red = 9700 Pro at default clocks
Benchmark settings all at defaults. 1024x768x32, compressed textures, no FSAA, 24bit Z, double buffered, Pure HW T&L

What does this have to do with CUDA? I dunno. It's about IGPs being exaggerated by 3DMark06 scores, mainly. Maybe if CUDA apps are not bandwidth sensitive, it won't matter and GF IGPs etc will perform like their little card-based friends.

 

[crystall]

Also annoying for me is that 780G and A64X2 don't get along very well. For best performance, you have to disable CnQ because otherwise framerate will vary dramatically, presumably due to the CPU changing speed even in the game, affecting RAM bandwidth. It causes game speed to literally slowly pulsate slower/faster. And I have problems with CnQ locking at full speed after video playback.

Out of curiosity: are those CnQ-related performance problem Windows-specific or do they happen also under Linux on X2/780G setups? (using AMD closed-source drivers for the IGP)

 

[swaaye]

Out of curiosity: are those CnQ-related performance problem Windows-specific or do they happen also under Linux on X2/780G setups? (using AMD closed-source drivers for the IGP)

Sorry, I haven't run Linux on it. You can try looking through this forum thread:
http://www.avsforum.com/avs-vb/showthread.php?t=992503

That's the biggest resource of experiences with 780G that I know of.

I can say that the CnQ probs occur in both XP and Vista though. Actually, there's another XP issue I forgot to mention earlier. If I have only one display enabled, 2D performance is noticeably diminished. Resizing windows, scrolling webpages, etc becomes visibly slower. Enable 2 monitors, problem goes away. I almost always run 2 monitors so it's not a big deal, but why this hasn't been fixed in many months is beyond me. I have submitted bug reports.

 

[Mintmaster]

I actually have some 3DMark2001SE results for both my 780G when it was super overclocked and of my old 9700 Pro on a Athlon XP-M. Both CPUs at 2300MHz. I realize it is 3DMark, but the results still tell you something about 780G's capabilities and bandwidth issues.

Well, 3DMark2001 completely eliminates the shader advantage of 780G. The Nature test needs some DX8 shaders, but it's a fraction of the workload because it is, as you said, dominated by alpha fillrate. BTW, your bench did show that 780G can hang with R300 in terms of texturing.

Anyway, in more modern games I'm pretty sure they'll be more even.

BTW, did you even try NFS:HP2 on R300 at 1680x1050 before faulting 780G for not being playable there? I know it's not a modern game, but I'm just curious.

 

[swaaye]

I don't have a 9700 Pro anymore but I did run one for several years so I have a good idea where it sits in various games. However, I've run a FX 5950 Ultra rather recently and still have that card. It actually can handle Guild Wars' DX9 effects in Eye of the North fairly well, believe it or not. Probably as well as a 9600 Pro, I'd guess considering my Mobility 9600's performance level. But it can handle 4X AA much better than the M9600 (gobs of bandwidth.)

My experience with 780G has been rather tainted because I paired it with a A64X2. It just isn't getting enough bandwidth and the various quirks with the pair cause performance issues on top of things. One day I hope to drop a 45nm Phenom into the mobo and see what that does combined with a 930 MHz IGP clock. Should be quite the difference I think. But will it make up for a ~208% difference in that nature test? I kinda doubt it.

However, it's worth noting that I never intended to play modern shader-heavy games on this hardware. I have better equipped machines for that so I do play mostly older games on it if anything at all. The machine is my always on file server/browser rig with the convenient ability to run retro stuff if I so desire. That includes Q3-powered games, UE1/2 games, old racers, and Guild Wars (reminds me of nature test).