What are Extreme Pipelines?
- Thread starter rms
- Start date
Hellbinder said:WHAT??
You better go check what i posted an Nvnews almost a year ago.
I posted a lot of Things, a lot of different times. Not all of them were intended on being totally accurate. What i have never failed to say even when messing around is its a *16* pixel Architecture.
Oh HB, don't be like that. I was referring to something you wrote similar to "sometimes 8 pipelines are like 16" or something which really to me bolstered the rumor of 8 extreme pipes. No worries anyhow, I still read your posts without hesitation.
WaltC said:trinibwoy said:
I believe I answered it. Heh...
Come on now, this is just informal chatting, not "3d-Graphics Hour With Your Host Captain Kirk, Brought to You in Thrilling Technocolor By Our Sponsor, TWIMTBP," is it?...
I should go sleep. Can't afford to stroll into work late again
Any chance of hard NDA lifted info before 1am EST ?DerekBaker
Newcomer
I guess this is an extreme pipeline:
Derek
Derek
martrox said:Chalnoth said:Inquirer is always correct. They make sure to cover every possible angle of an issue, so something is always right
Yup...it's called the "shotgun" approch to journalism......
I think journalism is a bit of a stretch. Most journalists don't seem to print every rumour that gets passed their way.
Actually, this is quite similar to the NV40 pixel shader. Instead of the term "co-issue" you just have "1 3-comp vect op + 1 scalar op", instead of "dual-issue" you just have two alus.DerekBaker said:I guess this is an extreme pipeline:
Some differences of course remain - NV40 has more flexible co-issue (can issue 4+0, 3+1, 2+2), the r420 can only do 4+0, 3+1, the r420 has the advantage of a dedicated texture address alu (nv40 needs to sacrifice 1 fp alu for that). There are of course small differences (like the NV40 fp16 free normalize and the mini-alus of the NV40 - no idea if the r420 has something similar), but overall the raw shading power seems to be quite similar.
Oh, and the r420 pixel shader pipe looks quite similar to the r3x0 pipe, except that there's now a second alu, so there's nothing really "extreme" about it.
In comparison to this rough analysis of NV40 versus R3xx and MDolenc's fillrate tester:
(counting per listed assembly op in the shaders)
IPC/pipe is still about 1 for the add/mad shaders, so there seems to be no improvement in handling these ops (0.97 versus the 0.98 I got for the R3xx).
For the PerPixelLighting test, X800 IPC seems to be about 1.42 versus 1.24 for the 9800.
The NV40 was 1.32 full precision, and 1.96 partial precision, I presume due to extracting normalization for its special functionality. Based on that assumption: for partial precision performance outside of the assumed normalization functionality, calculated by counting normalization as 1 extracted instruction instead of however many assembly instruction, indicates an IPC of about 1.27.
Overall, this seems to indicate a minor improvement in general case IPC for R3xx->R420, though the cause isn't fully clarified. My guesses remain something like faster execution of ops like rsq (how many clocks on R3xx?) or dp3 ability added to 2nd ALU (this latter seems to fit the 2 or 3 dp3 opportunities that seem likely given dependency and the impact relative to R3xx IPC).
(counting per listed assembly op in the shaders)
IPC/pipe is still about 1 for the add/mad shaders, so there seems to be no improvement in handling these ops (0.97 versus the 0.98 I got for the R3xx).
For the PerPixelLighting test, X800 IPC seems to be about 1.42 versus 1.24 for the 9800.
The NV40 was 1.32 full precision, and 1.96 partial precision, I presume due to extracting normalization for its special functionality. Based on that assumption: for partial precision performance outside of the assumed normalization functionality, calculated by counting normalization as 1 extracted instruction instead of however many assembly instruction, indicates an IPC of about 1.27.
Overall, this seems to indicate a minor improvement in general case IPC for R3xx->R420, though the cause isn't fully clarified. My guesses remain something like faster execution of ops like rsq (how many clocks on R3xx?) or dp3 ability added to 2nd ALU (this latter seems to fit the 2 or 3 dp3 opportunities that seem likely given dependency and the impact relative to R3xx IPC).
Xmas said:
Read the review!
I just feel all used, cheap, and dirty over the whole "extreme" pipelines thing now.
demalion said:
Why would you extract free norm vs any of the other "special functions" like POW, LIT, SINCOS, etc? If you just want to measure vector IPC, you also need to correct for special functions of ATI's scalar ALUs as well. That doesn't seem to be a fair comparison.
Normalization takes 3 instructions. This lower IPC, it would raise it. First of all, it's an additional instruction which be issued in parallel. Secondly, it raises the effectively how many instructions are issued by 3 since you are able to dispatch 3 instructions in the time needed to dispatch 1 instruction. For example, a 5 instruction long shader considering of a NRM (3 instructions) + vector op, plus a final combine, would execute in 2 cycles. On another architecture, it would take 4 cycles.
Possible. Hard to tell though, even if we know a lot more about the structure of the NV40 than we did with NV3x, too much of internal details are unknown - for instance ATI could have larger L1 texture caches than Nvidia, which might negate the potential performance disadvantage of not having a L2 texture cache.Tahir said:
Similar threads
