Mark Pillsbury (no_spam@gmail.com) on 12/15/11 wrote:
---------------------------
>David Kanter wrote:
>
>>>4. Bury the animosity toward Nvidia. It hurts
>>> the whole industry when there is no performance
>>> compatibility between Nvidia and ATI graphics.
>>
>> Hrmm, what do you mean? AMD's architecture is
>> more efficient in terms of perf/W and perf/area,
>> so I don't see any reason for them to give up
>> that advantage.
>
>My thinking is that programmers don't have time to write two versions of graphics
>software, one for AMD and one for Nvidia. I wish AMD would swallow their pride
>and support CUDA or Nvidia would provide better support for OpenCL. Even if the
>API is the same, the graphics microarchitectures have diverged too much. The end
>result is that most software doesn't use a fraction of the capability of either company's chip.

Then you should be glad to hear that NVidia wants to open
source their LLVM based CUDA compiler and that the Mesa
developers experiment with LLVM for r600 GPUs.

Both LLVM won't be compatible, but it is a good opportunity.

>>>5. Price the Opteron 6000 series at the same
>>> price per socket as the Opteron 4000 series
>>> to encourage more adoption of high socket
>>> count systems and high core count software
>>> optimized for AMD's microarchitecture.
>>
>>Wouldn't it just be easier to have a bunch of field engineers optimize the applications?
>
>Field engineers can't redesign applications to use more cores. I think both Intel
>and AMD are being penny wise and pound foolish by charging an arm and a leg for
>the versions of their chips needed for three or more socket systems. They should
>love it when someone wants to put a lot of processor sockets in a system. It helps
>get software ready for the next generation chips with more cores.
>
>About the longer term need for AMD to have a 3G/4G wireless modem: I think they
>just need to find 20 to 30 good engineers and give them two years. Wireless technology
>is not rocket science. LTE and GSM support might be enough. Maybe they could license
>some of technology from some small design companies.

Or they could just not care and let the buyers put a chip
for wireless/3G/ 4G or whatever cool next wireless
technology on the board. You can take SoC too far, too.

(Or just license the IP, no need to do it again.)