The Many Facets of Intel and Nvidia’s Hatred

For all of the unpleasant legal and PR exchanges between Nvidia and Intel, they probably aren’t ready to stop doing business together, at least not yet. Both have explained that CPUs and GPUs will undergo a convergence; both feel that they will be the winner and the other company a loser. Obviously, Intel is backing the CPU, and Nvidia is backing the GPU.

While the underlying tension doesn’t help, the questions over Nehalem chipsets and the Ion platform are likely related to a more immediate concern of Intel’s.

Intel probably isn’t trying to push Nvidia out of these markets entirely, but they do have some rules they expect Nvidia to play by. Ion, by allowing users to perform tasks with an Atom CPU that used to require a more expensive Core 2, costs Intel the difference between selling Atoms and Core 2s. We covered Intel’s concerns with Ion in more detail here.

Nehalem poses a similar issue, but at the higher end of the market. Intel has three sockets lined up for Nehalem, each socket appears only on motherboards with certain features, and each socket only fits a certain performance range of processor. Nvidia could drop a chipset for a low-end Nehalem with more motherboard functionality than Intel wants at that level.

The point of contention is likely which sockets will be available on motherboards with SLI at a full x16 lanes of bandwidth per card. Intel figures if you’re buying more than one GPU, you can afford a top-end Core i7 CPU as well. Nvidia would obviously rather see multiple-graphics cards in a wider range of hardware. We discuss the Nvidia Nehalem chipset question further here.

Last April, Nvidia and Intel exchanged some unkind words. At the heart of that exchange, the true conflict between the two was revealed. Specifically, they have highly divergent opinions about the convergence of GPU and CPU. Convergence of these two is already taking place, and will be a dominant source of change over the next decade.

The theory behind processor convergence is that CPUs are very programmable and flexible, but only beginning to become more parallel (i.e. have more cores), while GPUs are very parallel, but only beginning to develop more flexibility.

Current convergence efforts revolve around using GPUs to do some of the work of CPUs, called GPGPU (General Purpose computing on the GPU), such as Nvidia’s CUDA architecture. Intel doesn’t want to miss this boat, but are coming at it from the other direction.

They are working on Larrabee, essentially a GPU made up of several modified CPU cores. It is likely to have a lot of the flexibility required to drive GPGPU applications, but Intel is aiming it squarely at the consumer market. That will be a tough nut to crack, though. Unlike the High Performance Computing market running GPGPU stuff, which Nvidia enjoys with the CUDA based Tesla line, the consumer market wants the best for their buck in terms of graphics. Consumer GPGPU applications like Badaboom are at most a secondary consideration.

There are indications that Larrabee will have trouble matching Nvidia and AMD GPUs in terms of raw graphics, at least in an attractive package. Rumors and observations include that Larrabee will be a massive, 300W chip on an expensive, 12-layer, PCB.

They have a few: after all, it’s a multifaceted conflict. For one thing, Nvidia plans to go ahead with shipping Nehalem motherboard chipsets. If Intel and Nvidia don’t come to an agreement, it will take years for this to play out in the courts. Nvidia can’t be expected to sit on their hands and lose a generation or two’s worth of market share and development costs. If they lose, they might pay some heavy fines, but it’s a risk they can’t afford not to take.

At first, it looked like Nvidia was going to take a hard line with Ion 2. It was originally believed it would support Via’s Nano CPU, instead of Intel’s Atom. There were also rumors that Nvidia would get into the x86 processor game. Since then, Nvidia has announced, perhaps in some conciliatory decision reached with Intel behind the scenes, that Ion 2 will support the Nano and not only Intel’s Atom, but their Celerons, Pentiums, and Core 2s.

That might resolve the short term segmentation concerns, but Nvidia’s CEO also said at an investor conference that it was more a matter of when than if Nvidia will get into the main processor market. It is important to note that this was in the context of Systems On a Chip, meaning low power chips used in MIDs and such. There was no talk of making something faster than Nehalem or equally dramatic.

Then again, the mobile side is where the growth is, and Intel is pretty defensive about its x86 patents. AMD and Global Foundries are finding out just how defensive Intel is (explained here) the hard way. If the convergence battle indeed sees Nvidia get into the x86 market, they will probably hear from Intel’s lawyers, no matter how carefully they tread.