At this point, it is clear that NVIDIA had the right idea when it theorized that strong enough GPU compute accelerators would end up supplying the bulk of a supercomputer's performance. AMD is now offering a new product of this type.
Long story short, AMD has formally launched the FirePro S10000 compute accelerator card, with two Tahiti graphics processing units and 6 GB of GDDR5 ECC memory.
One single adapter has a peak single-precision floating-point performance of 5.91 TFLOPS, and a double-precision performance of 1.48 TFLOPS.
It is being advertised as a rival to NVIDIA's Tesla cards and Intel's Xeon Phi x86 accelerators, the latter of which are finally shipping to customers.
Virtualization technologies, like Microsoft RemoteFX, are fully supported, allowing multiple users to run common office applications from a single GPU.
Direct GPU pass-through support for VMware ESXi/ESX and Citrix XenServer is supported as well, so that advanced workstation graphics applications can be utilized (one GPU per user model).
“The AMD FirePro S10000, equipped with our Graphics Core Next Architecture, enables server graphics to play a dual role in providing both compute and graphics horsepower simultaneously,” said David Cummings, senior director and general manager of professional graphics at AMD.
AMD gave its FirePro S10000 four mini-DisplayPort connectors, plus dual-link DVI. Useful for multi-monitor workstations.
This is actually the main difference between this product and NVIDIA's Tesla K20, or other GPU accelerators: video ports actually exist.
Since normal supercomputer accelerators only perform parallel computing tasks, they don't get video outputs. AMD's new Graphics Core Next adapter, on the other hand, “enables server graphics to play a dual role in providing both compute and graphics horsepower simultaneously.”
"This is executed without compromising performance for users while helping reduce the total cost of ownership for IT managers,” David Cummings said.