Power and Innovation to Drive High-End GPUs in 2009

[mascaras]

Power and Innovation to Drive High-End GPUs in 2009

The year 2008 so far, has been very eventful for the graphics card market. A reinvigorated GPU lineup by ATI, brought in some fierce competition to NVIDIA, which resulted in a tug-of-war with pricing graphics cards in the market, with either company refusing to lose on grounds of pricing. This event, coupled with the announcement of several game titles by game publishers, resulted in bumper-sales of graphics cards, giving the present state of the global economy little or no relevance.

The months to come hold the same amount of importance for both AMD and NVIDIA, where the next round of competition begins with successors to current high-end products being slated. NVIDIA is expected to continue with its monolithic high transistor-count GPU design methodology, while AMD could bring in a little change to the way it uses two efficient GPUs to build powerful products.





The change AMD is planning on is using a Multi-Chip Module (MCM), a design similar to the one on Intel microprocessors, where Intel has been making some of its quad-core processors using two dual-core dice on an MCM. The module facilitates faster interconnection between the GPU cores, than if the cores were to communicate over a PCI-Express bridge chip, or using internal CrossFire ports. This would result in a single GPU with two dice, in essence, a dual-core GPU. The RV870 would be the first to support an MCM design, with unit cores featuring 25% more stream-processors than its predecessor, while upping the rated shader compute power to 1.5 TFLOP/s. In effect, it could feature a 512-bit GDDR5 memory bus, support DirectX 11, and OpenCL (Compute language, a GPGPU API). According to this projection, AMD could be out with a DirectX 11 GPU at least six months ahead of NVIDIA, although it remains to be seen if there are any games that come out that soon to take advantage of the technology. Between Q3 and Q4 2009, AMD could plan on a dual-MCM setup, bringing together two such cores with two unit cores each.

On to the green camp, and NVIDIA seems to have concrete plans to respond to AMD, with the GeForce GTX 270, and GeForce GTX 290 planned for this quarter. The two new SKUs are based on a 55nm silicon fabricated G200, the G200b, also known as GT206. Within this quarter itself, NVIDIA plans to prepare a dual-GPU card based on G200b cores, to lead its front. In Q1 2009, the company is planning its transit to the 40nm silicon process, with GT212. There’s little known about this chip, while GT216 in Q2 shows NVIDIA incorporating GDDR5 memory across a wide memory bus. This will also be the time when NVIDIA plans a dual-GPU card out of these chips. Finally in Q4, comes GT300, the next generation GPU from NVIDIA that brings in support for DirectX 11.

Year 2009 along with the rest of 2008, promise to be times when both graphics giants take computational power to the next level. It remains to be seen, if NVIDIA is able to deliver the punch, in order to rebuild whatever ground it lost to AMD, or if AMD goes for the kill with its innovation. Either way, these are good times for the buyers.

http://www.techpowerup.com/75586/Pow...s_in_2009.html

regards

[Hornet331]

anyone doubts mcm for highend cards?

A RV770 already consume north of 130W (and thats not TDP). Imagine 2 dice close to each other only releasing 120W... how the hell you want to cool 240W on less then 500mm² TEC cooling is stock now?

[Shintai]

anyone doubts mcm for highend cards?

A RV770 already consume north of 130W (and thats not TDP). Imagine 2 dice close to each other only releasing 120W... how the hell you want to cool 240W on less then 500mm² TEC cooling is stock now?

Yep. It makes no sense to make an MCM GPU compared to single dies. Unless you wish to hype something.

Seems techpowerup did the usual junk and copy/pasted stuff.

Only innovation 2009 brings is Larrabee. Else its just more of the same. More shaders, more clocks and shrinks. There simply just aint much innovation to do unless you go seriously radical like Intel.

And I doubt the HD5000 series is DX11. It seems to be shaders, shrink and clocks only. (Dumb shrink so to say).

[Vinas]

while upping the rated shader compute power to 1.5 TFLOP

That is going to be awesome! No good way for a TEC to cool 240w though.

[cegras]

Yep. It makes no sense to make an MCM GPU compared to single dies. Unless you wish to hype something.

Seems techpowerup did the usual junk and copy/pasted stuff.

Only innovation 2009 brings is Larrabee. Else its just more of the same. More shaders, more clocks and shrinks. There simply just aint much innovation to do unless you go seriously radical like Intel.

And I doubt the HD5000 series is DX11. It seems to be shaders, shrink and clocks only. (Dumb shrink so to say).

What's wrong with MCM? Given a well designed interconnect architecture, you're surely saving on massive yield problems then if you produced a 1600 shader monster.

[Hornet331]

What's wrong with MCM? Given a well designed interconnect architecture, you're surely saving on massive yield problems then if you produced a 1600 shader monster.

Plain an simple -> heat. RV770 now emmits ~130W at full load at a die size of 256mm² -> 40nm shrink that which puts the W/mm² even higher -> which means it is harder to cool, now place two of this suckers next to each other you you double that heat output...

MCM is only possible if they drasticly reduce power consumption -> lower then 80W under full load. (remember smithfield was 144W under load) and that thing was hard to cool.

[Klarko]

Sounds kickass, cant wait

[Shintai]

What's wrong with MCM? Given a well designed interconnect architecture, you're surely saving on massive yield problems then if you produced a 1600 shader monster.

So like a 5GB/sec sideport? To transfer what..the frames? Or will you talk about shared memory? Shared memory would also require an actual MPU.

Plus what benefit does MCM gives over the X2 approach? More expensive PCB with more layers to be countered a slightly smaller PCB?

MCM on GPUs just doesnt work like it does on CPUs with a shared FSB on the AGTL+.

[[XC] gomeler]

I hope they indeed do an MCM for their highend GPU. It'll make cooling it a much easier task!

[cegras]

Plain an simple -> heat. RV770 now emmits ~130W at full load at a die size of 256mm² -> 40nm shrink that which puts the W/mm² even higher -> which means it is harder to cool, now place two of this suckers next to each other you you double that heat output...

MCM is only possible if they drasticly reduce power consumption -> lower then 80W under full load. (remember smithfield was 144W under load) and that thing was hard to cool.

Maybe. But ATI said they were going to implement vapour chamber technology, and none of this stupid solid copper block stuff like on the 4870X2. If the shrink goes well or whatever other factors come into play, given a large enough chamber it'll be possible.

That, or they just have two chips connected via a true interconnect to appear as one GPU.

Either way it works.

[Shintai]

R700 was also MCM, minicores, shared memory and whatever. Its amazing what people WANT to believe. We gonna see the same thing all over with RV870/R800 as with the 700 series.

[Macadamia]

ts amazing what people WANT to believe.

Why don't we start with Larabee? It's the closest to vaporware, considering RV870 will tape out in Winter/Q1 latest and will be on time.

It's easier for ATI to make a functional MCM (though I disagree with the rumours) than for Intel to make Larabee not a joke at launch.

[Shintai]

Why don't we start with Larabee? It's the closest to vaporware, considering RV870 will tape out in Winter/Q1 latest and will be on time.

It's easier for ATI to make a functional MCM (though I disagree with the rumours) than for Intel to make Larabee not a joke at launch.

You do know Larrabee did tape out and is in ES now?

The point of MCM and GPUs just doesnt mix. It did for CPUs due to low memory bandwidth, full MPU and so on.

[Miwo]

If its one thing that AMD should have learned from Intel these past years, is that MCM does work and it doesn't matter if you have a real single-die solution. Despite whatever criticisms of MCM solutions being inelegant, sloppy, whatever....... it is the performance numbers that matter in the end. I know this is beating a dead horse, but for lack of a better example: Nobody cared that the K10 was a "true" quadcore.... and we all know the rest.

If MCM effectively doubles GPU power without involving Crossfire support in drivers (ie an MCM config will act as a Single GPU), then it'd definitely be worthwile. It would be a valid solution to having to require Crossfire for more performance along with the driver mess it involves. They can manage heat by using lesser cores. Since their whole architecture is designed to be broken down into SIMD units and modular, they just need to figure out the optimum SIMD configuration that allows for the smallest chip to the highest SP/TMU config.

For example, using 2x cores with 640SP / 32TMU each (ie 4830 specs) on 45nm would keep power consumption down, yet still completely blow away the 4870

[RPGWiZaRD]

Nice, now I know what my next upgrade will be => GT212 or RV870 (I prefer single GPU card solutions tho). GT206 arrives too late IMO concidering it most likely won't make any significant change to current state in GPU market other than NVIDIA maybe starting to sell cards with some profit. The GT206 should be released as soon as NVIDIA can and if there's delays it most likely means there's some problems as I don't see in my logic anything else that makes sense delaying this launch otherwise and wouldn't be suprised if NVIDIA originally tried to aim for a release already a month ago. I have a feeling I'd sleep better like a AMD worker than NVIDIA currently.

[Glow9]

G212 GX2 interesting.... and whats Dual MCM birdged?
They ever going to make the GX2 or X2 series capable of using all 4 outputs for individual screens?

[ryboto]

G212 GX2 interesting.... and whats Dual MCM birdged?
They ever going to make the GX2 or X2 series capable of using all 4 outputs for individual screens?

MCM bridged seems to imply 4xRV870 cores, two MCM modules.

[LordEC911]

Why don't we start with Larabee? It's the closest to vaporware, considering RV870 will tape out in Winter/Q1 latest and will be on time.

RV870 already taped out...

[Bojamijams]

Watercooling will be able to cool these MCM's. However, it will require their own separate loop. I don't think air cooling would be able to cool that.

[Glow9]

MCM bridged seems to imply 4xRV870 cores, two MCM modules.

as in 4 cores on one card or 2 cards crossfired?


Is the GT206 a replacement for the 9800s or onboard/mobile thing? Am I the only one who misses how card names used to be?

[G0ldBr1ck]

Almost 1.5 years ago the Catalyst driver Dev team was working on software for this type of architecture acording to an interview with one of the Dev's. I have no reason to believe that this is not whats coming, if its not than will be with rv970 at the least.

[Klarko]

I hope they indeed do an MCM for their highend GPU. It'll make cooling it a much easier task!

Could you clarify how that is?

[bedlamite]

It's not easy to cool these X2 card under LN2
And there is not many WC blocks and alternate AC systems for them.
But afaik this MCM thing is just a rumour. Release dates too...

[TheKarmakazi]

Could you clarify how that is?

Slap 1 copper pot on it rather than 2 pots which you also have to create/buy a new mountingplate for. He speaks only relative to LN2 cooling though.

[RaZz!]

sounds like a dumb question, but will a mcm card solve problems like micro stuttering that sli/cf graphics cards have these days?