AMD Bulldozer as server and HPC

[Carlis]

It appears that the bulldozer chip is not competitive in the desktop segment in most scenarios, but from the review here on AT it seemed to do fairly well with very threaded applications. This raises the question; can it be competitive in the server and HPC markets? Of course, HPC is a very small market but there amd is doing very well. Given that there is no shortage of threads in many server applications, amd might do better there than in the desktop segment. Though it appears that the power usage is larger than for intel systems, and that's bad for servers.

For some reason amd continues to make profit, yet the share price has certainly plummeted. I was waiting for the bulldozer review to decide on buying shares, but given the weakness of BD in the desktop segment, I am not going to buy at 4.92 usd (price right now).

 

[alyarb]

http://www.ornl.gov/info/press_releases/get_press_release.cfm?ReleaseNumber=mr20111011-00

 

[tatertot]

Assuming BD was great in the server space, one would think that AMD might have had some basic server benchmarks to release along with/before the Zambezi disaster, in order to cushion the blow.

Has anyone seen any spec_rate benches for Valencia or Interlagos yet? Anything from TPC?

 

[jpiniero]

http://www.ornl.gov/info/press_releases/get_press_release.cfm?ReleaseNumber=mr20111011-00

I imagine AMD is very proud to sell Cray Opertons at a huge discount for the sole purpose of feeding nVidia GPUs.

 

[alyarb]

Their chem simulators do not run on Tesla. The first upgrade phase is moving the platform to Interlagos and a 1000-card peripheral array of Fermi accelerators "to begin modifying their scientific applications." So maybe a part of their simulators will run on Tesla in two years if they are lucky.

The second upgrade in 2H2012 will be the 7,000-18,000 Kepler array. Since they don't yet know how many GPUs they are buying, it is quite obvious that they have no way of guessing how useful the GPUs will be, or how lopsided their load balancing will work. All of the large custom GPGPU jobs will have growing pains because you have to individually partition and delegate a workload to the CPU or GPU. It is not likely that they will use Interlagos as a thread scheduler for the GPUs. Fermi incorporates pretty fine-grained scheduling already and when you are approaching ten million concurrent threads I'd rather let the cuda framework handle it.