Virtual Cores/Shaders on ATi Radeon GPUs?

[ichime]

This the first time I'm hearing about this and oddly enough, it's being done on a mobile GPU(s) rather than on a desktop chip. Long story short, a guy made a VBIOS for the Radeon 6970M in the M18x that threads the physical shaders of the 6970M (960 shaders, based off the Desktop 6870) into Virtual shaders; similar to Intel's hyperthreading technology. Here is a link to the thread (it's in Chinese)

http://dell.benyouhui.it168.com/thread-1747026-1-1.html

Does anyone know if this is legit or not?

[STEvil]

texture fillrate doubled... hmm.

[zalbard]

Well, that does look promising.
Not sure if fake, though... Sounds a bit too complicated for a software mod.

[andos]

6990M has 1120 shaders though o.O

[zanzabar]

6990M has 1120 shaders though o.O

the 6990m is based on the 69xx desktop were the 6950/70m is based on the 68xx desktop.

[Burninator]

very interesting - would be epic if this is reproducible on a desktop card

[JaD]

It's a 6970, it shows up as 6990 because the driver ID was changed, the guy says it can be confirmed by the shader count (960). But does GPU-Z actually measure the shader count or does it use a database?
About the fillrate... increasing the shader count shouldn't alter it, unless GPU-Z derives the number of TMUs dividing the total number of shaders by the amount contained in a SIMD unit, 80 in VLIW5 architectures, as there's 1TMU per SIMD engine; even if that's the case, the displayed fillrate at 800MHz should be 76.8, not 72.6: I call fake.

[STEvil]

GPU-Z is probably using a database to detect the number of shaders and convert to fillrate since you can read later in the thread and see 1x, 2x, 3x, and 4x shader numbers giving up to 132gb

[W1zzard]

does anyone have a laptop with that gpu to add proper suppport for it to gpuz?

[WangChung]

I thought GPU-Z actually counts the shaders, as when you do the 6950=>6970 desktop card mod it changes the number of shaders shown in the display after the mod is completed. Correct my understanding if that's wrong.

[zalbard]

It doesn't count shaders, no...

Edit: wait, if it does, then how come it didn't for Cayman before a software update?

[W1zzard]

I thought GPU-Z actually counts the shaders, as when you do the 6950=>6970 desktop card mod it changes the number of shaders shown in the display after the mod is completed. Correct my understanding if that's wrong.

it does for desktop gpus, and it tries for mobile but its probably not reading them right

[zir_blazer]

Google Translator results: This, this and this too.

What I understand is that it add a sort of SMT support that allows it to see twice the Shaders it physically has as with Hyper Threading. How on the earth they do this is not something that I understand, because editing a BIOS the way they're doing it is not an easy feat, if not nearly impossible. Instead of changing some values (Like what you do with Radeon BIOS Editor to mess with PowerPlay Power States values), he basically should have need to alter how the GPU works with some BIOS modifications (That according to that is a "hybrid BIOS" of a 6970M and 6990M), something that I don't see viable because it means you should have tons of white papers about how the GPU works (That generally are all propietary and few people should have access to that info) or a great deal of reverse engineering. If AMD aren't capable of introducing such a feature themselves, you can't expect someone else to do so. But maybe this guy as access to such data...

Small personal note
Friends yesterday, and a jar of (northern, old altar friends know...) QQ to talk about AW, I suddenly thought of the inspiration for so many years of research on AMD, with the University of Washington where I have to do a compilation of AMD work friends, then help you much. However, because this project is almost unprecedented, there is no reference and reference, and all had to start from scratch.

[M.Beier]

Very interesting reading, too bad google translate isng super.

[W1zzard]

Can someone try this test build on a Blackcomb GPU: http://www.techpowerup.com/wizzard/GPU-Z.exe

It uses the shader detection method from Cayman

[Lanek]

I don't know why, im a bit dubitative ..... If it really work, good. but otherwise, i pain to see how the "virtual shaders" are used, or can communicate then with the memory controller etc etc. and how this don't make the card struggle then with Triangle processors etc..

[nn_step]

I'm voting hoax, since even from a theoretical view because the chips just don't have that much processing power.

It may however spoof benchmarks for better performance.

[[XC] hipno650]

hmmm im not buying it just yet. it's a cool idea in theory but I fail to see how some average joe came up with some home brew Vbios that made this HUGE change in rendering and the like without any hardware changes and without a major company figuring this out before. you would think AMD or Nvidia would have tried this sometime ago if it were that easy...

[STEvil]

Well, look how long it took for the "Ricky Tweak" (multi-threaded HAL on single core system) to come about, or multi-thread/cpu systems at all.

[3NZ0]

Looks interesting, I'd give it a go but I'm unsure of what I have to really do.

[W1zzard]

Looks interesting, I'd give it a go but I'm unsure of what I have to really do.

if you have a cayman mobile (=blackcomb) gpu, you could start helping by posting a gpuz screenshot using the gpuz build i posted a few posts further up