nVidia's x86 project

By power666

^{Tue, August 17th, 2010 at 9:01PM CDT}
Business Week has the recent scoop of nVidia's much rumored CPU design. It is no secret that several former employees of Transmeta are on nVidia's payroll and they were indeed working on a CPU design. The goal: get an x86 compatible CPU to market before both AMD and Intel move entirely to system-on-a-chip architecture. By including graphics on the same chip as a CPU, nVidia will find itself in a niche market where their discrete graphics products become a premium luxury.

The catch for nVidia entering the mainstream processor market has always been a legal matter. The x86 instruction set is controlled by AMD and Intel through various patents. Via is the only company to have an up-to-date license for those patents to create a new x86 CPU design. Via's license is non-transferable so nVidia cannot buy them out and gain the ability to make their own x86 design. nVidia's last legal hope fell two weeks ago when the FTC settled an anti-trust suit. The only good thing for nVidia to come out of that agreement was that Intel would continue to provide an on-die PCI-e controller for a discrete graphics card for several more years.

So how nVidia was giong to enter the x86 market is one of great speculation. Rumors of their x86 project date back years so they did have the time to build an x86 chip from scratch. This direct approach wouldn't last long before Intel would get a legal injunction, block its release and then attempt to sure nVidia into oblivion. With the hiring of several transmeta engineers, it was believed that nVidia would develop their own code morphing schema to emulate the x86 instruction set using a different architecture. It was thought that nVidia would bolt on the necessary code morphing hardware to one of their GPU designs. Performance wouldn't be great but it would be functional to get a foot hold in the market. Another source claims that nVidia was bolting on code morphing hardware to an ARM based design. nVidia currently has licenses for several ARM core designs and presumably a license for the architecture in general. Integrating the additional hardware to accelerate code morphing would elimiate the power advantage ARM currently holds over the x86 market. Furthermore, the ARM designs nVidia has licensed are not performance competitive for desktop usage (they are fast for mobile though). To move into the desktop market, nVidia would also have to wait on a 64 bit native ARM design to emerge. With both AMD and Intel to release hybrid CPU+GPU chips next year, nVidia is under extreme pressure to compete but those forces may crush the company.

9 Comments

IBM exposes the current limits of eDRAM

By power666

^{Wed, July 28th, 2010 at 11:27PM CDT}
The secret for the Xbox 360's graphics performance is the usage of embedded DRAM, eDRAM, inside the same package as the graphics chip. The main limitation of the eDRAM in the XBox 360 is its size: 10 MB. The eDRAM was originally manufactured using NEC's 90 nm bulk process. So how much eDRAM can be placed onto a modern chip? Earlier this year IBM announced the POWER7 chip which featured 32 MB of eDRAM used as a L3 cache alongside 8 CPU cores and tons of IO logic. While the amount is impressive, this isn't the limit due to all the other components on that die. It didn't take IBM that much longer to design a nearly pure eDRAM chip but this time for the z series of mainframes (PDF). The mainframe processor module contains four 5.2 Ghz quad core chips and two L4 cache chips. The L4 cache chips do contain a bit of coherency logic to link between other modules but is mostly composed of 96 MB of eDRAM. The eDRAM chip is 478 mm^2 in size and built upon IBM's 45 nm SOI process. Removing the coherency logic from IBM's chip would allow for capacity to increase to approximately 160 MB.

So what good would a massive eDRAM chip be? Like in the Xbox 360, it could be used for a fast cache for a graphics processor. The main limitation eDRAM isn't used in modern GPU's is that the capacity to fit an HDR frame buffer. A 1920 x 1080 HDR buffering using 128 bit precision would take 32 MB. Generally two buffers are used in the rendering process but a 160 MB buffer can hold five comfortably. Chances are massive amounts of eDRAM will find its way into future chips.

4 Comments

Intel Silicon Photonics Demo

By power666

^{Tue, July 27th, 2010 at 4:55PM CDT}
As clock speeds climb and the transmission voltages dropping for copper wiring, the maximum wire length decreases. One way around this problem is to utilize optical wiring that does not suffer such steady degradation in signal quality. This is why optical transmission is ideal for long cable runs, often measured in kilometers, but the technology does have a distinct downside. Optical transceivers are an expensive component to add and the need for an external transceiver prevents optical technology from being used in ultra short distances.

An Intel demonstration bypasses several limitations imposed by current optical technology. How does Intel manage this? It is all done on single piece of silicon. Intel has mentioned this technology before but this is the first public demo of all the necessary components merged together. This allows for the optical components to be built inside of a processor instead of requiring external components. As an added benefit, the wiring distance is slightly shorter as the wiring for external logic is unnecessary. This will also be a boon for bringing optical transmission between chips on computer motherboards. High speed devices like 100 gigabit ethernet switches will likely need this technology to become economically viable for small to medium installations.



Intel isn't alone in this resarch. IBM has invested into optical technology as well. The crown jewel of IBM's efforts seem to be an economical way to build multilayer motherboards using optical wiring.

1 Comment

Wedding

By NoOutlet

^{Wed, July 21st, 2010 at 9:43PM CDT}
I know it's been a long time since I last posted. All sorts of business has occurred in the interim.
The big news is that I'm getting married in 3 days and 8 hours. It's going to be a wedding on the beach and I'm expecting that it will be amazing. Other news is that my fiancee is a member on the forum. Her username is kalemonger and she has a sub-site dedicated to healthy cooking recipes. For this sub-site, she wanted to post images. For awhile, I was uploading her images but now I've set it so that people with sub-sites can upload images automatically. I think this is a good policy that I wish I had thought of before.
Another bit of news is that I'm re-designing the code and style and working of the site on my own computer.
Another bit of news is that said computer is currently inoperable and so there is no work being done on the redesign for the foreseeable future.
Also, I'll be on honeymoon for 3 weeks after the wedding, so don't expect any official moderation to go on during that time.

15 Comments