The GPGPU revolution is well underway, and the next stop seems to be the lucrative server arena. From an information point of view, a vast majority of servers are the ideal applications for GPGPU acceleration, as they are often required to perform repetitive processing. There is no doubt GPGPUs can accelerate specific server processes several times over CPUs.

However, so far, the only drawback has been the form factor of server systems. Servers are clearly designed for CPU usage, with no real space to fit in graphics cards. It is possible to replace the form to fit in PCI-e GPUs. This is not an option many server data centers are likely to opt for, however. Instead, the two year timeline given by AMD might simply allude to the availability of Fusion APUs.

More next page.


The GPGPU revolution is well underway, and the next stop seems to be the
lucrative server arena. From an information point of view, a vast
majority of servers are the ideal applications for GPGPU acceleration,
as they are often required to perform repetitive processing. There is no
doubt GPGPUs can accelerate specific server processes several times
over CPUs.

However, so far, the only drawback has been the form factor of server
systems. Servers are clearly designed for CPU usage, with no real space
to fit in graphics cards. It is possible to replace the form to fit in
PCI-e GPUs. This is not an option many server data centers are likely to
opt for, however. Instead, the two year timeline given by AMD might
simply allude to the availability of Fusion APUs.

It is the obvious solution – if fitting in GPUs were a problem, simply make it disappear into the CPU die. We will have Llano APUs as early as 2011 with 400 Radeon HD 5000 class shaders which could considerably accelerate certain processes.

The server GPU acceleration market holds much potential for AMD. The only way Nvidia can accelerate servers currently is by selling workstation graphics cards – whose shape and size, as mentioned earlier, is an inconvenience for server systems. While Intel may have the stronger CPU in its upcoming APU (i.e. CPU + GPU), Sandy Bridge’s GPU is unlikely to be much of a GPU accelerator, with Larrabee indefinitely delayed and unlikely to feature in an Intel CPU anytime soon.

The second generation Fusion, which we can assume releases in 2012/13, according to the two-year timeline given by AMD, might be when the Fusion’s CPU integrates a next-generation Bulldozer core with a next-generation Radeon GPU. That could make for a powerful APU – and a perfect combination for next-generation server processors.

Reference: Xbitlabs