[petsc-dev] http://www.hpcwire.com/hpcwire/2012-11-12/intel_brings_manycore_x86_to_mar ket_with_knights_corner.html

[Karl Rupp]

>> In terms of raw numbers, $2,649 for 320 GB/sec and 8 GB of memory is quite a
>> lot compared to the $500 of a Radeon HD 7970 GHz Edition at 288 GB/sec and 3
>> GB memory. My hope is that Xeon Phi can do better than GPUs in kernels
>> requiring frequent global synchronizations, e.g. ILU-substitutions.
>
> for comaprision the tesla 2090 we have costed over 3k each - and provides 177 GB/sec [with 6GB ram]
>
> http://www.nvidia.com/object/tesla-servers.html

Oh, yes, that reminds me of one of the great moments of my GPU computing 
experience so far. We had a GTX 470 (~$400) in our test machine, but 
somehow hoping to get access to a Tesla because of the much higher 
double precision peak performance reported. We begged NVIDIA and they 
finally donated a Tesla 2050 for testing purposes. So, the day the board 
arrived, we were all excited about the expected extra punch for our 
iterative solvers to observe. What we ended up with was ~10% below the 
performance of the GTX 470 even in double precision, just because all 
these operations are so heavily memory-bandwidth limited. Also, for most 
simulations we carried out even the GTX 470 was sufficient, and when we 
really wanted to go to the limits of the work-station, the 3 GB on the 
Tesla wasn't enough either.

Lesson learned.

Best regards,
Karli