AMD's Latest Server Compute GPU Packs In 32GB of Memory

Deathspawner writes: Following-up on the release of 12GB and 16GB FirePro compute cards last fall, AMD has just announced a brand-new top-end: the 32GB FirePro S9170. Targeted at DGEMM computation, the S9170 sets a new record for GPU memory on a single card, and does so without a dual-GPU design. Architecturally, the S9170 is similar to the S9150, but is clocked a bit faster, and is set to cost about the same as well, at between $3,000~$4,000. While AMD's recent desktop Radeon launch might have left a bit to be desired, the company has proven with its S9170 that it's still able to push boundaries.

Comments

Re:What are these used for?

[chuckymonkey]

Not necessarily, in scientific computing cards like this are important. The biggest problem with GPU computing in general is the time it takes to copy from main memory to GPU memory and back. It really makes GPU difficult to work with and generally the gains in parallelization don't really pay off considering the amount of time it takes to make those memory copies. Being able to load more into memory and have it stay there is a big deal.