Sir Eric Demers on AMD R600
- Thread starter B3D News
- Start date
- Status
Umh? sw rasterization?
Really, ouch. But it needs to get almost 100% better to compete with the 8800GTX.
What?Actually, that`s not accurate. Since the R600 supports PCF and DST(a first for ATi), Stalker uses its best algorithm for shadow rendering on it, unlike the prior ATi HW that had no support for those features. This can be seen in Stalker's log/in the console. All of this software shadows buzz was started by the Float32 guy(who did great work with the mod, BTW), and I don`t think it`s been actually detailed what he meant with saying that. Razor, you should stop going to Rage3D so often, it`s rubbing off on you
.What?Actually, that`s not accurate. Since the R600 supports PCF and DST(a first for ATi), Stalker uses its best algorithm for shadow rendering on it, unlike the prior ATi HW that had no support for those features. This can be seen in Stalker's log/in the console. All of this software shadows buzz was started by the Float32 guy(who did great work with the mod, BTW), and I don`t think it`s been actually detailed what he meant with saying that. Razor, you should stop going to Rage3D so often, it`s rubbing off on you
LOL its not from Rage, just a sec let me find the link, I was looking into this before, a guy is working on a Mod to correct the problem, but also contacted AMD and THQ about it. Ok you mentioned the mod lol. So thats not accurate at all? I mean thats alot of performance issues in Stalker if there isn't a real problem.
It`s accurate for HW prior to R600, where ATi had no support for PCF. The R600 supports this. The mod author enabled Fetch4(which was, more or less, the ATi alternative to HW PCF, and is in DX10 AFAIR) for R5xx family cards...it seems. Wheter or not that improves performance I dunno, because he`s changed a lot of things alltogether and I don`t have R5xx hardware around to test anyhow.
What I`ve noted is that ATi's PCF seems to be a tad bit slow ATM, with the 8.37.4.3 drivers-at least in AndyTX's VSM demo, and in the DX SDK examples and in nVs DX10 VSM and PCF demos, so it may be a driver bug that has an indirect effect on Stalker performance. That being said, my experience with that game was odd, as I feel that the move from Crossfired x1900s to a single 2900 brought a performance increase...but I haven`t really benched it, so take that with a grain of salt.
That being said, I guess that Stalker wasn`t really top of the list in terms of driver optimisations, as it`s not yet in all benchmark suites, and the aim for the launch was(IMHO) to have good 3DMarks, good Oblivion, Prey/Q4 and whatever else is used by every reviewer everywhere, so that may be another thing holding performance back. Or it may simply suck with the X-Ray engine, and that`s that...although I really hope that`s not the case. We'll see.
What I`ve noted is that ATi's PCF seems to be a tad bit slow ATM, with the 8.37.4.3 drivers-at least in AndyTX's VSM demo, and in the DX SDK examples and in nVs DX10 VSM and PCF demos, so it may be a driver bug that has an indirect effect on Stalker performance. That being said, my experience with that game was odd, as I feel that the move from Crossfired x1900s to a single 2900 brought a performance increase...but I haven`t really benched it, so take that with a grain of salt.
That being said, I guess that Stalker wasn`t really top of the list in terms of driver optimisations, as it`s not yet in all benchmark suites, and the aim for the launch was(IMHO) to have good 3DMarks, good Oblivion, Prey/Q4 and whatever else is used by every reviewer everywhere, so that may be another thing holding performance back. Or it may simply suck with the X-Ray engine, and that`s that...although I really hope that`s not the case. We'll see.
Is a listing of the latencies for various register and cache reads available?
How about what can be expected for trips over the ring bus?
Is that not for public consumption?
How about what can be expected for trips over the ring bus?
Is that not for public consumption?
OpenGL guy
Veteran
Slow compared to what? We've seen some nice performance gains from exposing PCF.
Umm, slow compared to what I expected, but again, this is subjective, haven`t had the time to test it properly, I worded it poorly probably. And, BTW, kudos on the PCF thing...I`ve discovered that quite a few titles were using this, and it resulted in an improvement in shadowing quality in them.
OpenGL guy
Veteran
There shouldn't be any difference in shadow quality if the application coded things identically. If you lack PCF, then you do the following:
- bind color (R32F format if you want) and depth buffers
- render your shadows, while outputting depth to the color buffer
- to shadow objects, bind color buffer as texture, sample and compare to incoming depth value
These should give identical results. Only problem is performance as with PCF you can take four samples at once (a la bilinear) but with the other method you can only take one sample at a time, unless you're using FETCH4. On the HD 2xxx products, PCF should be equivalent in performance to bilinear, I believe.
Yeah, when I tested texture formats the R600's PCF performance was indeed similar to bilinear, and thus very competitive. Although I'll admit I wasn't 100% sure about these results, because it was in OGL and some formats were not exposed yet, and I didn't implement a way to check if the correct results were given...
In addition to that, G80 has a slight problem with single-channel stuff, which apparently also applies to PCF (although I only tested FP32 depth texture PCF) so that also makes R600 look even better comparatively.
Presumably, the reason why AndyTX's demo runs faster on G80 is that the pass rendering *to* the shadowmap must be running much faster than on the R600. If you compare G80's Z-write rate to R600's, that shouldn't be very surprising either.
Oh, and since I didn't chime in this thread yet apparently (oops!), incredibly great answers Eric, thanks a lot for taking the time for all this!
In addition to that, G80 has a slight problem with single-channel stuff, which apparently also applies to PCF (although I only tested FP32 depth texture PCF) so that also makes R600 look even better comparatively.
Presumably, the reason why AndyTX's demo runs faster on G80 is that the pass rendering *to* the shadowmap must be running much faster than on the R600. If you compare G80's Z-write rate to R600's, that shouldn't be very surprising either.
Oh, and since I didn't chime in this thread yet apparently (oops!), incredibly great answers Eric, thanks a lot for taking the time for all this!
If someone can russian: http://www.gamedev.ru/community/gamedev_lecture/articles/r_e_n_de_r
Page 3: something about shadows.
STALKER uses shadow maps; DF24 and fetch4 for ATi (RV515 (?), RV530, RV560/570, R580), it's not necessary to enable it (fetch4) per a modification.
Maybe there is a conflict and STALKER can't recognize the R600 correct (Stalker thinks: "What? There is an ATi graphics card inside and it can PCF and fetch4? *error*").
(The author of this lecture about STALKER's renderer says, that X-ray engine is doing everything with half precision, float32 mod forces full precision (fp32).)
Dear ATi guys: Where is graphics card assisted video transcoding?
Will ATi/AMD release finally a new version of Xcode with GPU support?
What's about DXVA 2.0/EVR? There is an interesiting thead in Doom9's forum: http://forum.doom9.org/showthread.php?t=125835
OpenGL guy
Veteran
S.T.A.L.K.E.R. properly uses PCF on R600.
Unknown Soldier
Veteran
Hopefully AMD learn and fix the R600 most if not all issues with it's next release. Would be nice it the card uses a lot less power too.
US
US
Fix what?
In our Tests: Nothing. It's actually one of the few tests where the R600 really shines. Maybe it depends strongly on the scene being tested?
That's pretty interesting as most reviews put this GPU far behind. But did you test with cat 7.6 maybe? And I heard there are some problems with AF through CCC?
As a GPU junky and enthousiast I'm getting a 1 GB GGDR4 board in next week, so I can put it up against my 8800GTX. I'm planning to CF'ire 2 of these beauty's later this summer.
- Status
