Next Gen Intel iGfx?

[AnarchX]

Ironlake, Nehalems IGP (said by Expreview), seems to appear on some Intel driver pages:
http://downloadmirror.intel.com/12091/ENG/GFX_Vista32_15.6.0.1322_PV_Intel_relnotes_gfx.htm

Look like it still based on current GMA-architecture and not Larrabee.

 

[DavidC]

Now i am thinking, with some further optimization, better clock to lock performance, and double shader amount with the die shrink to 45nm, Higher Clock speed, or double the clock speed.

The IGP in the Nehalem has 12 shaders, and 2 more fixed function units. Probably higher clock speed, but nothing more is known yet. It's disappointing the shader increase is only 20%, but oh well.

Probably the increased fixed function units will do something beyond the 20% increase in shaders. In the Gen 5(G45), the Vertex Shader performance increased by 2x which is why the improvements are greater than the simple changes we know.

Ok some latest test shows G45 is about 100 - 200% faster then G35. But it is still about 10 - 30% slower then ATI 780G.

I now hope Intel could have some more performance tuning from Drivers. And compatibility and stability.

The performance went up by a lot with newer drivers. The HKEPC site test got 1180 in 3DMark06 with the 15.9.8 drivers. I've seen tests with beta drivers and older 15.9.3 drivers which had same chipset scoring ~750. That gives me hope it'll at least catch up to the 780G with the final versions.

 

[AnarchX]

Clarkdale with a 900MHz and TDP >20W IGP?

A entry D3D11 add-in-card should be in the same power range, but lot faster...

 

[Davros]

well at least theyre going to make a dx11 gpu
which is good considering thy only need dx9 for aero

 

[Kaotik]

well at least theyre going to make a dx11 gpu
which is good considering thy only need dx9 for aero

DX9 for Aero, but DX10 for WDDM1.1 and IIRC 10.1 had some improvements on Aero too (not effect-wise of course)

 

[Entropy]

Actually the first Intel cpu to use the i prefix was the iAPX 432 in 1981. The i960 followed 1988. Jeez, you guys are a bunch of puppies.

If you count the 80960 as an i960, you shouldn't forget the i860, the Larrabee of its day.
It is worth noting that in spite of Intels strength in the marketplace and technologically, all of the above processors flopped, the iAPX 432 and the i860 quite spectacularly. I don't think Larrabee will ever really amount to anything in the marketplace. It simply doesn't bring enough benefit to justify itself. It's a preemptive strike against a dead-end branch in computational evolution.

 

[mczak]

Clarkdale with a 900MHz and TDP >20W IGP?

Oh hmm, that's interesting. It would make sense that the i5-661 has higher TDP (due to faster graphics, though it seems a bit too much more) but I fail to see why it would miss some features of the i5-660 (vpro, vt-d, txt) - well except maybe if that's the home-user version where neither of these are useful (but the faster graphics might be handy).

 

[mczak]

Ironlake, Nehalems IGP (said by Expreview), seems to appear on some Intel driver pages:
http://downloadmirror.intel.com/12091/ENG/GFX_Vista32_15.6.0.1322_PV_Intel_relnotes_gfx.htm

Look like it still based on current GMA-architecture and not Larrabee.

It's way too early for Larrabee-based IGP.
btw the open-source linux drivers already support this too so you can see some differences to other i965-based chips (it's named "IGDNG" for now), though I don't know if the driver actually really works on it .

 

[mczak]

The IGP in the Nehalem has 12 shaders, and 2 more fixed function units. Probably higher clock speed, but nothing more is known yet. It's disappointing the shader increase is only 20%, but oh well.

I always thought it doesn't have really much alu capacity, but actually that's wrong. Those shader units are 4 wide (*) so that's 40 MAC per clock (can't do MAD). Of course (at least for G45 and older), in contrast to AMD and Nvidia those also have to perform clipping, attribute interpolation (granted nvidia does interpolation in the shader too but can use special function unit, G45 merely has instructions which help with that but can't do it in parallel).

(*) from the released documentation. The instruction set is actually 8-wide (you can in fact compact under some circumstances to 16-wide but instructions are just run twice after decoding), but it is stated that physical size is 4-wide, there's also a 4-wide mode but if you use that spec says there's no useful work being done in second clock.

 

[Gubbi]

If you count the 80960 as an i960, you shouldn't forget the i860, the Larrabee of its day.
It is worth noting that in spite of Intels strength in the marketplace and technologically, all of the above processors flopped, the iAPX 432 and the i860 quite spectacularly. I don't think Larrabee will ever really amount to anything in the marketplace. It simply doesn't bring enough benefit to justify itself. It's a preemptive strike against a dead-end branch in computational evolution.

They didn't succeed on the desktop, but the i860 and i960 were succesful in embedded solutions, the i860 in Postscript printers, the i960 in automotive solutions.

One of the reasons they didn't succeed on the desktop (micro architectural quirks aside) was because of x86 momentum. Larrabee stands to benefit from that. Consequently, IMO, you cannot compare LB to those former Intel architectures.

Cheers

 

[Simon F]

They didn't succeed on the desktop, but the i860 and i960 were succesful in embedded solutions, the i860 in Postscript printers,

IIRC the i860 was also in some SGI workstations.

 

[dkanter]

They didn't succeed on the desktop, but the i860 and i960 were succesful in embedded solutions, the i860 in Postscript printers, the i960 in automotive solutions.

One of the reasons they didn't succeed on the desktop (micro architectural quirks aside) was because of x86 momentum. Larrabee stands to benefit from that. Consequently, IMO, you cannot compare LB to those former Intel architectures.

Cheers

LRB isn't exactly 100% x86 compatible, it doesn't run SSE AFAIK.

DK

 

[Florin]

They didn't succeed on the desktop, but the i860 and i960 were succesful in embedded solutions, the i860 in Postscript printers, the i960 in automotive solutions.

The i960 was also quite common as a RAID controller, lasted way longer than it was competitive performance-wise.

 

[Gubbi]

LRB isn't exactly 100% x86 compatible, it doesn't run SSE AFAIK.

DK

Not yet.

I'm still viewing LRB as a will-it-float-or-will-it-sink test from Intel. If it is succesful (even moderately so) as a massively cored graphics chip, I'd expect the next LRB (or the next one after that) to be a convergence device, with real good per core x86 performance and a superset of all the architectural extras.

How hard would it be to make those super wide vector execution units do SSE by jumping through some extra decode hoops ?

Cheers

 

[mczak]

LRB isn't exactly 100% x86 compatible, it doesn't run SSE AFAIK.

By that definition even a Pentium II isn't x86-compatible.

How hard would it be to make those super wide vector execution units do SSE by jumping through some extra decode hoops ?

Should be doable I think but I don't know it would make much sense (though I didn't look at all instructions to see if all of them could be more or less trivially implemented with some decoding voodoo). It would also only run at 1/4 the performance of the lrb vector instructions (128bit vs. 512bit).

 

[aaronspink]

The i960 was also quite common as a RAID controller, lasted way longer than it was competitive performance-wise.

Define performance-wise...

There's performance and then there is PERFORMANCE. The later means that the thing actually works. The horror stories in the industry over buggy firmware for raid controllers are pretty numerous.

 

[Thorburn]

The i960 was also quite common as a RAID controller, lasted way longer than it was competitive performance-wise.

Also it was the main CPU in Sega's Model 2 arcade platform (Daytona, Sega Rally, etc).