Journal 11, November 5

Nvidia Making x86 CPU With Ex-Transmeta Brains?

The x86 ISA (instruction set architecture) originally developed by Intel has had a life of continual legal battles wins and losses on both sides of the Intel fence. Intel has long had competitors that also produce x86 designs under license namely AMD and Cyrix owned by VIA.

Intel has recently made a bold move that upsets the balances a bit. Intels latest designs integrate a memory controller onto the CPU die. The impact of this is that other companies would be forced to use an separate memory controller or license Intel's on chip memory controller assuming Intel would even be willing to license it at all. Nvidia made the next move by directly accessing intel's memory controller in their latest chipset designs. Intel of course retaliated and Nvidia has subsequently ceased chipset development as far as is know to the public.

An exception to that rule would be Fujitsu that currently still produces Sparc based designs for high performance computing needs and is touted as the fastest CPU .

Nvidia has never been in the CPU business but has lots of high proformace design experience. Much like the article I think that Nvidia aims to add support for executing x86 binaries on their hardware.

There has been speculation for some time that they would do this and Nvidia's CEO has even threatened it a time or two! I find it rather intriguing that Nvidia has hired many former Transmeta employees working for them possibly to work on x86 compatibility for their GPUs. In my opinion Transmeta's biggest development was a x86 compatible processor that did not use the x86 instruction set in hardware. This allowed them to translate x86 or any instruction set within reason into their own instruction format with good proformace due to their design.

In my opinion if Nvdia were to use a similar translation technology as Transmeta used and implemented it into their GPUs it might spur progress with parallel processing since in theory programs running on the main CPU could be migrated directly onto the GPU if it were determined that it were a multi-threaded program capable of benefiting from the GPU's massive parallel architecture.

Similar ideas are also in the works at Intel on the Larrabee project and also at AMD on the Bulldozer project. Its good to see that Nvidia is not going to lie down on this one and let Intel and AMD get too far ahead.