Enervation: How to make something new unintelligible
|
|
|
Nvidia denies it's moving to x86
IN THE FACE OF overwhelming evidence on the world wide web, Nvidia is denying that it is working on an x86 clone to rival anything that Intel comes up with.
Last week the rumour mill was working overtime, manufacturing hell on earth for the Green Goblin by claiming it had hired Transmeta technicians to take on the mighty Chipzilla on its own turf.
The rumour seemed to be gaining so much traction that Nvidia CEO Jen-Hsun Huang went on record to deny it.
He told shareholders last week that he was only interested in visual and parallel computing, GPUs in servers and those sorts of things, perhaps with a dash of mobile just to spike the taste buds.
The rumours were started, and continued, by Doug Freedman of Broadpoint AmTech, who thinks that Nvidia will have to go to designing x86 chips within a year to preserve both GPU and chipset revenue.
Freeman's claims do not seem to be going away and some elements seem to suggest that he has an inside source at the Green Goblin. For example, he said that Nvidia might use Global Foundries as a manufacturing partner and last week Nvidia complained that TSMC was not allocating it enough capacity. That doesn't mean it's moving, but it is one shoe dropped.
However, Huang seemed to forget that complaint when he rebuffed the rumour. He told analysts after the shareholders meeting that "Global Foundries is AMD's fab. Our strategy is TSMC."
There would be a number of reasons that Nvidia would not want to officially announce a low powered x86 chip and it would be acceptable if Huang tried to string out the press as long as possible on such a plan.
Does this mean that Nvidia is still going ahead with the plan? It is clear that Huang has to do something now that the two companies are at war over chipsets.
A large chunk of Nvidia's business comes from its chipset business and it would not seem to be a good idea for the company to walk away from it. An x86 CPU clone might be a sharp move.
But we would have thought it would take the Green Goblin more than a year to get it into production, unless this has been planned for a lot longer than the rumour mill claims.
What is possible is that Nvidia is starting to think longer term. A much forgotten announcement by Bill Dally, Nvidia's chief scientist, indicated something that has been on Nvidia's mind is called an "Nvidia Exascale Machine", which the graphics company is scheduled to release in 2017.
The Exascale GPU will pack 2400 throughput cores (7200 FPUs) and 16 CPUs on a single chip with a TDP of 300W, delivering up to 40TFLOPS of single-precision floating-point processing power or 13TDLOPs of double-precision floating-point processing capabilities. Each node of the chip will feature 128GB of memory, 2TBps bandwidth and 512GB of Phase-change or Flash memory for checkpoint and scratch space.
However that sort of hardware is a long way into the future and does not solve Nvidia's immediate problems. It might be that development of x86 technology would fill a short term hole.
Either way it will take something a more than a Huangdenial to make these rumours go away. µ.
Share this:
Share
Transmeta did not have an x86 license and didn't need one. Its IP is would now be from an IP firm, though Nvidia may have bought a license before Transmeta ingloriously dissolved. Nvidia did buy a license for the Transmeta Longrun low power tech if memory serves.
Transmeta nailed Intel for $250 million with a patent infringement suit and the management then sold out as essentially a round of financing for a startup.
AMD also made a few Transmeta processors with a spin on the CMS interpreter code for Microsoft and its pay to play home computer project. Transmeta cores ported to a modern process with enough cache could likely be competitive as a low power x86. However, multiple cores would not be able to cache common optimized code since that code would differ due to different execution history. Each core would have to cache that code separately in its very own "level 2" and the chip overall could have a common "level 3" of unprocessed x86 binary.
Transmeta did make a few missteps like losing a year by changing fabs. But it had no room for errors. The same tactics that got Intel in trouble in Japan for its actions against AMD strangled Transmeta.
Can't wait for THREE companies tacking things onto the x86 instruction set. There's only, what, 5000 instructions now?
yeah right. My conspiracy theory (a.k.a made up bullshit) is this. Nvidia is looking to fight Laffabee on its own terms. Think about it. Intel drives the development tools quite well, so if they can get x86 emulation a la transmeta-styles, hook it into their other bits of IP, chipsets and GPU, then they can have a larrabee compatable solution, however they have their own GPU do the heavy lifting, thus outperforming anything Intel comes up with, yet still keeping their current GPU designs. Keeps a foot in both camps and prepares them for more atom competition.
Probably sounder more coherent in my head, but you get the gist. Hey it is a *theory* so someone tell me how it won't work.......
bugger - you other clever dicks (and dickesses?) have beaten me to it...
I don't think Transmeta had an x86 licence either (please someone in the know correct me if I'm wrong).
AFAIK their code morphing translated the x86 instructions into the chips own native VLIW instruction set, thus letting them get around patents or something like that?
I think Nvidia is trying to find a way to get an integrated solution for it's Fermi platform.
They obviously have the GPU, they also have the chipset knowledge. The only thing missing is a CPU. Any CPU, even a crippled, barely functional one would actually do. Just to run the OS and the X86 apps who would act as a shell to manage the GPGPU array in the background and simple I/O to/from storage. The real horsepower will come from the GPGPU anyway, no need for a fast or complex CPU.
This way, NVDIA could sell a 100% NVDIA solution and do practically whatever it wants, without begging anything to anyone. The X86 JIT visualization might be a way to circumvent the X86 licensing problem and make for a much more flexible and cheaper solution on the long run. That would also fit with the hiring of Tranmeta engineers.
Ramon
A licence is not needed to emulate. The ARMs are faster and faster every day. NV could even integrate many ARM cores and GPGPUs with x86 emulation/translation into a single super fast machine. i386 is sufficient compatibility. Even MMX is unnecessary. The GPU does the heavy stuff (hopely in a seamless way), the ARM cores run the system. QEMU anyone?
Yes, that was what I was wondering. Nvidia need that licence or they're totally stuck. I wonder if they could buy another company (e.g. VIA) who do have an x86 licence?
This is all pointless. Nvidia doesn't have an x86 license which means they can't manufacture or sell x86 chips, so where is the point in this?
Moreover, the ARM is a small core. A modern x86 occupies 200 million transistors. You could pack a thousand ARMs into that.
Interesting idea.
The best bet for basic x86 compatibility for NVIDIA would be to write a software x86 JIT for their ARM-based Tegra line - like the old Alpha FX!32 software. They might even be able to hardware accelerate some or all of the process. This could then let them run x86 software on ARM, for backward compatibility reasons.
Performance wouldn't be a target of such an emulator. It would allow old business applications to run inside Windows running on the emulator at acceptable performance. It's not for Photoshop but for MyCompany'sStupidTPSApplication.
Hence the hiring of Transmeta engineers, who know a lot about translating x86 instructions via a combination of hardware and software. This could form the basis of a custom x86 decoder block for the Tegra ARM cores (ARM supports multiple ISA decoders, e.g., ARM, Thumb2, Jazelle, etc).
Most People Think Nvidia will Have to Go Toe2Toe with next chipset, LEO & Make cpu that fits thurban. Of Course NOT, CPU for Tablet or Other Programable DeveiceIS More Likely Start.
Maybe ITS NOT that Hard & Entire Show Will Open Up. Except in Ultee'THUNK Nvidia IS Intel. What Will INTEL Do, Castrate Evil GREEN? Will ATI Laugh OFF Nvidia? Await Leadership of DRASHEK to Find OUT.
TS vondrashek
I agree with you, but offer one counterpoint. When Nvidia got into the chipset business everyone said the same thing, and the Nforce/2/3 were solidly performing chipsets.
VIA, on the other hand, took YEARS before you could even begin to safely poke a stick at their chipsets, and even then, their success came solely because AMD wouldn't seriously get into the Chipset game on the K6/Athlon line. VIA, for some reason, refuses to do anything useful or spectacular despite having every opportunity literally laid at is feet naked, begging to be taken advantage of.
So many articles about the nVidia x86 rumours make it sound like any old company can just set up a skunkworks and crank out a competitive x86 processor in a jiffy. BS. It has taken the likes of AMD and intel decades to get their x86 processors to be where they are today, and they have patented every step of the way. Even VIA have been at it for a while and they are no where near competitive. There is no way even a company like nVidia could produce a part anywhere near the low end of what intel and amd produce without treading all over a boatload of their patents. Cue many lawsuits and patent infringement notices, making sure even if such a processor was designed that it would never see the light of day.
These investment analysts, and the journalists that spread these rumours, simply show how little they know about the industry with every word they utter.
Nice story without the usual Nick Farrell dribble/sarcasm overkill.
If he writes more like this there is hope.