I just wanted to confirm - but if you don't care: http://www.computerbase.de/artikel/...3850_rv670/8/#abschnitt_clive_barkers_jericho
AMD: R7xx Speculation
- Thread starter Unknown Soldier
- Start date
- Status
In STALKER and Geforce 8s case, that is true AFAIK.
UE3 being partially deferred, want to draw parallels to that?
Using DX10+ shader resolve does seem not worth it when a seperate shaderpath has to be done, and in DX10 code execution itself it does seem like hacks aren't given much freedom. No forced TSAA, no custom filters, no funky SSAO shaders...
@Lukfi
But that still doesn't explain how the XTX almost caught up with the 3870 with AA when without it, it loses by 10 frames.
Jawed
Legend
HD2900XT is 13% faster than X1950XTX. Based on clocks it should be 14% faster.
HD3870 is slower than HD2900XT on that page, whereas here it is the same speed:
http://www.computerbase.de/artikel/...e_9800_gx2/9/#abschnitt_clive_barkers_jericho
26.1fps for HD3870 (but HD2900XT not listed).
This page:
http://www.computerbase.de/artikel/...hd_3850_x2/8/#abschnitt_clive_barkers_jericho
shows no recent improvement in RV670 performance.
Meanwhile, HD3850 512MB is clocked 2.7% faster than X1950XTX, but HD3850's memory is 82.8% of X1950XTX's memory. 22.5fps for HD3850 is 97.4% of X1950XTX's 23.1.
Jawed
Lukfi: No, I don't think so. As seen in one of Jawed's links, the difference seems to be smaller than 1%.
Jawed:
I was under the impression, you wanted to be shown an instance, where RV670 was faster than R580 with AA applied - hence my question, if you'd limit that to specific conditions and/or models. You didn't bother to comment on that, so there you have it. Personally, I don't care about this very much, as almost anything can be shown by cherrypicking reviews and or benchmarks.
edit:
Besides - what's your point? That this is an exception from the rule? I'll give you that. That this benchmark is not sound? Maybe - I don't know, but CB has changed test platforms between my link and your two.
Jawed:
I was under the impression, you wanted to be shown an instance, where RV670 was faster than R580 with AA applied - hence my question, if you'd limit that to specific conditions and/or models. You didn't bother to comment on that, so there you have it. Personally, I don't care about this very much, as almost anything can be shown by cherrypicking reviews and or benchmarks.
edit:
Besides - what's your point? That this is an exception from the rule? I'll give you that. That this benchmark is not sound? Maybe - I don't know, but CB has changed test platforms between my link and your two.
Last edited by a moderator:
=>Quasar: Of course I was referring to the chart in your link. I know that 256 MB is usually not a problem for Radeon cards (as opposed to GeForces), but apparently Jericho doesn't like 256MB cards, whether they have good memory management or not.
=>Tchock: You're right, it's strange. Seems the particular game is tailored for R580 and not for R600/RV670...
=>Tchock: You're right, it's strange. Seems the particular game is tailored for R580 and not for R600/RV670...
WTF even 8X AF is more expensive than 8X AA? This is unacceptable, 8X or 16X AF are the most important settings for increasing overall visual quality (I have a FW900 CRT so I can get away without AA or with only 2X), if this doesn't improve in the 4000 series then no buy from me.
Hmmm do any reviews corroborate that? It would be ironic if all the hype around the AA performance hit was due to AF being turned on at the same time.
Would have been nice if B3D had done an analysis of the performance hit for enabling various levels of AA and AF on R600 or RV670 similiar to what was done for G80. You know it would be the only analysis of that type...further improving B3D's reputation and raking in the page hits
Would have been nice if B3D had done an analysis of the performance hit for enabling various levels of AA and AF on R600 or RV670 similiar to what was done for G80. You know it would be the only analysis of that type...further improving B3D's reputation and raking in the page hits
annihilator
Newcomer
AF is a form of texture filtering, which was horrible in R600 and RV670; which is why AF sucks on HD2xxx and 3xxx.
Hopefully, the rumors we're hearing are true and texture units have been indeed doubled in the RV770; but if all the rumors are true that would mean a twice increase in computational power (1GFLOPS) which looks like texture units could still be a bottleneck. :S
Hopefully, the rumors we're hearing are true and texture units have been indeed doubled in the RV770; but if all the rumors are true that would mean a twice increase in computational power (1GFLOPS) which looks like texture units could still be a bottleneck. :S
Jawed
Legend
I presume you meant RV670 slower than R580.
The lack of AA-only tests and the use of immature drivers certainly causes a lot of grief if one wants to try to understand an architecture. It'd be nice if _xxx_ could point us in the direction of the compelling evidence he has for "a slower AA implementation".
Take the FEAR results in the test you linked. They show HD2900XT as only 11% faster at 1280 16xAF/4xAA. But 13% faster at 1600. CPU limited? Driver-threading problem? Is R6xx more CPU-limited at lower resolutions than R5xx? Vista versus XP problem? etc. We know these variables are in the mix.
There were clear cases in the early days (erm, months) of R600 where performance was way below X1950XTX. Sadly the latter has dropped off most benchmark charts so it's hard to track. It would be nice to have some results that we can discuss in technical terms.
A 4xMSAA resolve shader:
Code:
struct samples
{
float4 colorA : color0;
float4 colorB : color1;
float4 colorC : color2;
float4 colorD : color3;
};
float4 main(samples IN) : COLOR
{
return (IN.colorA + IN.colorB + IN.colorC + IN.colorD) * 0.25;
}
Code:
ps_3_0
def c0, 0.25, 0, 0, 0
dcl_color v0
dcl_color1 v1
dcl_color2 v2
dcl_color3 v3
mov r0, v0
add r0, r0, v1
add r0, r0, v2
add r0, r0, v3
mul oC0, r0, c0.xThat shader isn't actually of the right form because it's using a vertex attribute to provide the samples to the shader - whereas in reality four registers (r0...r3, say) would be populated with the samples (since R600 can dump data directly into the register file - I presume this is the "special path between RBEs and shaders" that's alluded to). When the correct shader is executed R600's throughput would be 16 pixels per clock.
I'm an HLSL noob so apologies for the code
If R580 is faster than 16 pixels per clock in resolving 4xMSAA then R600 loses. To be honest I don't know R580's resolve rate for 4xMSAA data. Potentially it's complicated by the fact that a render target is a mixture of single-sample tiles and 4-sample tiles. It's also complicated by the format of the samples, int8 or fp16.
Jawed
so both the 2900xt + hd3870 are slower per clock than a 1900xt, thats sad so much for architectural improvements
Jawed
Legend
http://www.computerbase.de/artikel/...d_3870_rv670/5/#abschnitt_aa_und_afskalierungWould have been nice if B3D had done an analysis of the performance hit for enabling various levels of AA and AF on R600 or RV670 similiar to what was done for G80. You know it would be the only analysis of that type...further improving B3D's reputation and raking in the page hits
Jawed
chavvdarrr
Veteran
someone wishing to comment on english ?
Last edited by a moderator:
Whoa, 2900 and 3800 series take massive hits with 8X HQ or better AF, almost 50% with 16X HQ! The biggest hit for 8800GT with 16X HQ AF is 20%. They better have fixed this for 4800.
Jawed
Legend
Aha, interesting.
In the past AMD people have been adamant that texturing hardware isn't used. Or, at the very least, that texture fetches aren't performed in order to get samples into the shader hardware.
As far as I can tell the RBE sample de-compression system must be used to extract samples for MSAA resolve - so it's a question of whether a route from the RBEs to L2 texture cache is available...
Well, unless a patent document turns up, we'lll prolly never know the route that samples take.
Jawed
What is the difference between HQAF and regular AF? And which do most reviewers use?
Regular AF doesn't appear to be much performance hit, and if that's what reviewers use it's not affecting benchmarks.
- Status
