[Beowulf] Inside the Titan Supercomputer: 299K AMD x86 Cores and 18.6K NVIDIA GPUs

[Vincent Diepeveen]

On Oct 31, 2012, at 4:33 PM, C. Bergström wrote:

> On 10/31/12 08:17 PM, Vincent Diepeveen wrote:
>> Let's see. PDF about K20 says:
>>
>> K20 gpgpu: "3x the double precision performance compared to the
>> previous generation... M2090.."
>>
>> M2090 : 665 gflop.
>>
>> See: http://www.nvidia.com/object/tesla-servers.html
>>
>> A theoretic peak of: 18.6k * 0.665 tflop * 3 = 37.107 Pflop
> I can't validate or attest to the accuracy, but the *actual* numbers
> I've seen in other articles is 1.16 (not 0.665 * 3) per K20 and around
> 20Pflop for Titan.  We have to wait until Top500 is published to get
> actual numbers

The Nvidia PDF on their homepage says it's 3x faster than the M2090.

So to calculate the peak gflops that's 3 x 0.665 = nearly 2 Tflop for  
each GPU.

How much that results into for the sporthall top500 benchmark is less  
relevant i'd say,
we may safely assume that they are not using this benchmark for  
production codes...



> -----------
> Anyone with more 1st hand knowledge of the matter please feel free to
> correct me
> _______________________________________________
> 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