Nvidia Firmly Denies Plans To Build a CPU

Barence writes "A senior vice president of Nvidia has denied rumours that the company is planning an entry into the x86 CPU market. Speaking to PC Pro, Chris Malachowsky, another co-founder and senior vice president, was unequivocal. 'That's not our business,' he insisted. 'It's not our business to build a CPU. We're a visual computing company, and I think the reason we've survived the other 35 companies who were making graphics at the start is that we've stayed focused.' He also pointed out that such a move would expose the company to fierce competition. 'Are we likely to build a CPU and take out Intel?' he asked. 'I don't think so, given their thirty-year head start and billions and billions of dollars invested in it. I think staying focused is our best strategy.' He was also dismissive of the threat from Intel's Larrabee architecture, following Nvidia's chief architect calling it a 'GPU from 2006' at the weekend."

Inaccurate headline

[TheRaven64]

nVidia are building a CPU, a Cortex A9 derivative with a GPU on-die and a load of other nice features. The summary states that they're not building an x86 CPU, but this is not what the headline says.

Difficult

[DrYak]

Microcode-upgrade are possible for CPU that have a huge big complex reprogrammable pipeline like the current top of the line CPUs, or CPU where the pipeline is handled in software (like the Transmeta chips).

GPU, on the other hand, have a very short and simplistic pipeline which is hard-fixed. They draw their tremendous performance, from the fact that this pipeline drives ultra-wide SIMD units which process a fuck-load of identical threads in parallel.

But there nothing much you could reprogramm currently. Most of the die is just huge cache, huge registry files, and a crazy amount of parallel floating point ADD/MUL blocks for the SIMD. The pipeline is completely lost amid the rest.
(Whereas on CPU, even if the cache dwarfs the other structure, there are quite complex logic blocks dedicated to instruction fetching and decoding).

Re:Anyone Surprised?

[morgan_greywolf]

I don't get the legal action part. Is the x86 architecture patented by Intel? Even if it is, wouldn't the patent have expired by now? After all, its more than 30 years old. Do AMD, VIA etc. pay licensing fees to Intel for building processors using the x86 architecture? If so, why cant NVidia?

Yes. Various pieces of parts of the x86 architecture that have been developed within the last 20 years (noteably, stuff related to the IA32 architecture of the 386, 486 and Pentium and later lines) are all still under patent.

Patents filed before June 8, 1995 get the greater of 17 past the patent grant date or 20 years total, whichever is greater.

Re:wouldn't this be a good thing?

[ThisNukes4u]

Actually nvidia doesn't own any fabs, they contract out all their chips to TSMC, same as ati. Although now ati/amd are going to be making their fusion chips at TSMC, so they will definitely have the expertise to make x86 chips in the near future(TSMC will).

Re:Focused

[microbrew_nj]

I can think of a few good reasons for Nvidia to roll their own chipsets. SLI is one. The market for integrated motherboards (with their chipset) is another.

Re:Anyone Surprised?

[Daengbo]

Check the URL before clicking.

Re:Difficult

[billcopc]

Let me guess: you've never read anything about microprocessor engineering, have you ?

What you describe is what every non-engineer dreams of. You want a chip that any idiot can reprogram, without knowing the "less simple" ways of FPGAs. That's kind of like saying you want a car that gets 200 miles to the gallon, can park in a shoebox and carry 20 kids in the back seat - oh, and it drives itself automagically so your kids can take themselves to soccer practice without bugging you.

The reason why no one ever builds such monstrosities is because there is simply no point to it, when you can have purpose-built chips designed and fabbed for a fraction of the cost. People don't stop breathing just because their device needs 2 distinct chips instead of one jesus-truck.

Re:Difficult

[dreamchaser]

What you are describing is a pipe dream. Even *if* they managed to do something like that, performance would be utter crap, die size would be huge, and the odds are it just plain would suck.

Re:Only reason

[Lisandro]

How is a "GPU" different from a "CPU"?

The GPU is a specialized (vector) processor, while the CPU is a general purpose one. What the GPU does, it does great. But its reach ends pretty much there.

The nVidia is programmed with a specific higher-order assembly language, We rely solely on the hardware vendor for tools. I think that this is UNIQUE in the (mass-market) processor world. And this is why Intel, with an x86 compatible GPU is such a threat.

You're confused. Intel is not working on a "x86 GPU". Intel is working on a new GPU design - the kicker being that this is a relatively high performance one, instead of the kind of GPUs they offered so far (feature packed, but lacking in performance). The x86 instruction set has nothing to do with it, and in fact, has nothing to do with GPU programming, which is a completely different beast.

Can anyone else produce an OpenGL shader compiler for the nVidia? Or, better yet, extend it to do NON-shader tasks. How about for the AMD?

If i'm no mistaken, nVidias CG compiler is now open sourced. So yes.