Replies to post #98592 on Intel Corp (INTC)

[mmoy]

> but, wouldn't NVidia have at least as much IP in this arena as
> Intel???

Why would you ask that?

Yes, a graphics card does a lot of operations in parallel. Far more than a CPU can do (of that type). Great. Let's say that you take a thousand calculators and put them on your desk. A harmonica beats those calculators at playing a tune.

Computer architecture is hard to explain but if you want a better idea of the problem: learn a programming language that compiles to assembler code and then write a program that's a few pages long. Get the compiler to spit out the assembler code and then look at the assembler code to see if you could write it faster using multiple threads. Okay, that's an awful lot of stuff to learn. Computer architecture is an undergrad course. So is assembler. So is compilers. So is performance work. Intel and Microsoft are offering large amounts of money to universities to do parallel compiler research. If it were really easy to get a bunch of small processors to improve performance, they would not have to do this.

[mas]

Intel has Larrabee which is a lot of Pentium I type cpus all working in unison so Intel has that Niagara type throughput side covered too if it ever takes off in x86.

[chipguy]

wouldn't NVidia have at least as much IP in this arena as Intel

Nvidia has deep knowledge about making graphics processors.

Graphics is the most ideal workload for sea of simple processor
type chips you could imagine - nearly infinitely sub-divisible
homogenous workload with next to no inter-thread communication
or serialization overhead.

General purpose computing workloads that PCs and servers run
are far lumpier, less divisible, and often highly asymmetric with
serial bottlenecks. Threads and processors often share data
intensely and with fine granularity. Nvidia's institutional know
how building GPUs is next to useless for designing uncores
for large scale CMP chips for general purpose computing.