So, who will design AMD graphics in the future?

15 May 2008 | 10:52 BST

By Nebojsa Novakovic

WHILE Nvidia's Jen Hsun Huang sometimes gloats a tad rudely over the competition's headaches, there could be one point justifying his recent analyst talk confidence - the leapfrog game between Nvidia and AMD-ATI is expected to continue as usual between the HD4870 series, then GeForce 9900 series, then their respective 40nm shrinks early next year.

It's business as usual, with the two parties doing a regular leadership switch.

What about deeper into 2009? Yes, you'll all mention Intel Larrabee - after all, that is Nvidia bosses' favourite topic these days it seems. They don't even seem to mention DAAMIT much in that contest. Why?

One reason seems to be that steady stream [Shurely 'flood'? - Ed] of departures from the GPU division of AMD. Are they saying "Daamit, I had enough of this"? Whatever the case, it won't affect the R700 series and probably the 40nm shrink either. But, whatever the roadmap in place, future GPU designs can't be done without the right people.

This could be bad news if not properly addressed. I happen to quite like the current batch of ATI GPUs - while they may not be the highest performers, they are pretty fast still for comparatively low power consumption and, of course, price.

Also, they handle desktop graphics well, and now OpenGL looks better with the new drivers. The branching-capable architecture supports computational GPU work for the past three generations, and the RV770X2 will provide very nifty dual GPU shared memory system, rather than the usual duplication, that will be even more useful in computational work then Crossfire gaming. For example, two GPUs could work in parallel on a single problem and dataset, just like a dual-CPU system does today.

Also, despite the eternal Intel-AMD rivalry, the Crossfire parallel GPU operation continues to be supported on most Intel platforms. The X38 and X48 support it today - as do mainstream P35 and P45 chipsets. With Nehalem's arriv al in five months or so, both the X58 TylersburgDT desktop and TylersburgDP dual-CPU workstation chipsets will support 2 x PCIe x16 v2 Crossfire operation right out of the box. So, dual RV770X2 cards might be seen in many early Nehalem gaming rigs - especially since there's no news of an Nvidia Nehalem chipset yet.

What could AMD do next? Let's assume the worst, and no replacement of all those lost people. If fresh blood doesn't come in, one solution, in my mind, is merging the GPU group into the CPU one, and working out a new strategy where one or more R700 like units are integrated into future high-end CPUs - if AMD stays in the high-end CPU business long enough, that is.

A CPU with four X86 and two RV770 cores on a single die in 32nm process might not be that unviable, if the memory bandwidth issues are taken care of (quad channel DDR4 or something). With minor feature updates along the way, that would make the R700 series work useful for at least two generations of high-end Fusion GPUs. That could also be an interesting answer to Larrabee.

Even without it, as we can see, the Intel platform advantage, coupled with 40 nm shrink, can help the AMD GPU group survive the next nine to twelve months just fine. But beyond that, a major recruitment drive of design teams and associated positions is vital.

Put simply, too much brain drain lowers the collective IQ, and we don't want Nvidia to aim for near 100 per cent discrete GPU market share in late 2009, with only Larrabee standing in its way.

Now, was that another reason for Nvidia not exactly liking it?

© 2007 Incisive Media Investments Ltd. 2007