View Single Post
  #6  
Old August 14th 08, 10:19 PM posted to alt.comp.hardware.amd.x86-64,alt.comp.periphs.videocards.nvidia,alt.comp.periphs.videocards.ati,comp.sys.intel,comp.sys.ibm.pc.hardware.video
[email protected]
external usenet poster
 
Posts: 4
Default TheInquirer: 'Why AMD should merge with Nvidia'

On Aug 14, 1:19 pm, Robert Myers wrote:
On Aug 13, 6:26 pm, Yousuf Khan wrote:
Beyond that, AMD's Fusion CPU/GPU looks like it will be a better bet for
floating point domination than Larabee/Nehelem combination.


"Floating point domination" could mean any number of things. If, by
that, you mean domination of HPC (including applications like
animation), we'll just have to see. Cost, power consumption, and the
usability and stability of the API will be more important than raw
flops.


Specifically, AMD and Intel are both attempting to extend their SIMD
instruction sets for the next generation of floating point. Intel
calls theirs AVX (Advanced Vector eXtensions), while AMD calls theirs
SSE5. The two instructions aren't compatible with each other but they
do basically the same thing, so it is a forking of the SSE standards.
Both will be introducing 3-operand SSE instructions, leaving behind
the current 2 operand variety.

However it's how they implement the SIMD engine in the background that
makes the difference. Intel will be using multiple little Pentium I
cores, which is the basis of its Larabee project. And AMD will be
implementing its ATI latest graphics cores. AVX will be a frontend for
Larabee, while SSE5 will be a frontend for the ATI GPU. Both
instruction sets are supersets of the existing x86 instruction set
therefore they will be easy to program for in certain ways. Intel's
Larabee will depend on certain amount of super-sized multi-threading
of the software to get the most out of its Larabee cores. AMD won't
need that much multi-threading of the software since, GPU's are
already highly parallelized by definition. Will it be easier for
compilers to create the level of multithreading that Intel requires,
or will they be more comfortable just throwing the data at the GPU and
letting the GPU sort it out for them? We'll have to see how that
plays.

Marketing will play a big role, and both IBM and Intel, which is the
real competition, are much better situated to get their products
placed. Both have a much better track record than AMD, which did make
some dents with Opteron and Hypertransport. I think that blip is
past, but, as I said, we'll just have to see.


Intel has finally caught up to all of the technology that AMD did
introduce about 5 years ago. But this CPU/GPU hybrid with a friendly
CPU frontend is a new direction that Intel can't take yet. Intel is
attempting to emulate a GPU with Larabee, but how good it's going to
be is questionable. Intel is excitedly talking about starting Larabee
out with 32 cores and then expanding that out to 64 cores. By
comparison Nvidia's latest GPU (GTX 280) already has 240 cores, while
AMD's latest single-chip GPU (HD 4870) has 480 cores.

Legit Reviews - NVIDIA GeForce GTX 280 Graphics Cards by EVGA and PNY
- NVIDIA Brings The Muscle - GeForce GTX 280
http://www.legitreviews.com/article/726/1/

Radeon HD 4000 series specifications surfaces | NordicHardware
http://www.nordichardware.com/news,7356.html

Also AMD has been concentrating on optimizing the double-precision FP
performance, a clear sign they are looking for a bigger market for
their GPUs than just graphics. Nvidia suffers a huge performance hit
when comparing double-precision to single-precision, something like a
4x decrease. AMD is only suffering a bit more than 2x decrease so far.
Even with these double-precision performance hits they are still
faster than Cell at double-precision. The following are comparisons
showing older versions of Cell and AMD GPU; Cell comes out at 14.6
GFlops and 200 GFlops for ATI Firestream 9160 (basically Radeon HD
3870) for double-precision.

Berkeley Lab Researchers Analyze Performance, Potential of Cell
Processor
http://www.lbl.gov/CS/Archive/news053006a.html

AMD FireStream - Wikipedia, the free encyclopedia
http://en.wikipedia.org/wiki/AMD_FireStream

Yousuf Khan