[Beowulf] More AMD rumors

Vincent Diepeveen diep at xs4all.nl
Mon Nov 19 12:27:22 PST 2012



On Nov 19, 2012, at 9:07 PM, Douglas Eadline wrote:

>
>> On 11/19/2012 09:28 AM, Douglas Eadline wrote:
>>> Though not formally announced, Semiaccurate seems to have learned
>>> that AMD is leaving the X86 server business.
>>>
>>> http://semiaccurate.com/2012/11/19/amd-kills-off-big-cores-kaveri- 
>>> steamroller-and-excavator/
>>>
>>>
>> of course the article was amended at 10:15am to say that AMD  
>> contacted
>> them to say that Kaveri and the big cores are still on track.
>>
>
> I hope that is true. There was also the "looking for a buyer"
> rumor that surfaced last week. In my unofficial discussions with
> AMD, they said they are looking forward to APU on server (which
> flies in the face of the semiaccurate claim) and that the new
> roadmap will get them to back to competing.
>
> The key will be sales of the three major game consoles, all of
> which AMD won video deals. If they can work their roadmap
> while the console revenues keep them going things might
> be looking good in 2-3 years. For now, Intel owns the
> street. Although Intel seems to have their howitzer pointed
> at NVidia at the moment -- MIC/Phi and all that.
>

Intel is a high margin manufacturer - HPC has lots of commissions who  
ignore some of the
realities, with a huge budget for just a few cpu's, and who prefer  
benchmarks; it's easy to target
a single benchmark always.

Find me 1 benchmark where AMD cpu's managed to turbo and where intels
machines didn't profit a lot from turboboost, to give 1 example.

> But I have to give Intel soem major geek credit.

Intel gives a lot of stuff for free, you wonder why the geeks love  
intel?

HPC is easy to manipulate in this manner. If there is 1 knowledgeable  
dude in a commission,
that's a lot for HPC. Finding a way to convince him somehow to  find  
your brand of cpu's better is
total peanuts if you invest into that. We know intel has been doing  
that.

Note i wouldn't say they bribed anyone - but let's put it this way -  
some of the reports i read i would arguably
call written by someone IQ90 who didn't figure out yet how to google  
yet.

Based upon such reports decisions have been taken and will get taken  
and this is where intel is really good.
I don't claim they bribed anyone.

If we look in other areas of IT we see how near the top bribing is a  
big problem for buying in hardware.
In most cases both the bribing company as well as the victim soon  
shuts up about it and nearly nothing gets leaked
as it damages both of them. Yet this is a major problem of HPC.

AMD never gave away anything AFAIK.

That's one of the reasons that kills AMD currently - they never gave  
away presents to benchmarking geeks.

Also i try to remember to you how the GCC compiler last few years  
improved a lot for intel and not for AMD.

AMD didn't deliver.

As for HPC i want to remind you how in gpgpu they had to fix rather  
few tickets and the few i filed, i can't say
they were very helpful. I got contradictary answers from the helpdesk  
versus what engineers posted and they didn't
solve my request - making me belief that the helpdesk just simply  
lied and that an engineer posting was correct.

A list of promises from AMD what their gpgpu should be able to do.  
We're 5 years later now and it still doesn't work well.
Besides that it takes always months and months before a new driver is  
there.

There is no place for AMD in HPC simply - their helpdesk is not working.

> Bringing the
> original Star Trek bridge to SC12 was the highlight of the show.
> And yes I sat in Kirk's chair and have the pictures to prove it.
>
>
> --
> Doug
>
>
> --
> Doug
>
>
>
>>
>>
>>
>>
>> --
>> Mailscanner: Clean
>>
>
>
> --
> Doug
>
> -- 
> Mailscanner: Clean
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
> Computing
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list