It seemed like for a while the graphics processing unit (GPU) was on the verge of swarming into everyday enterprise infrastructures. The more companies like Nvidia, Intel and AMD pushed the platform's capabilities, the more it appeared that we would soon see GPUs pushing alphanumerical data for standard business applications.
That drive took a hit this week with Intel's decision not to ship its Larrabee processor this year, opting instead to fold much of its technology into forthcoming multicore lines for server and laptop products. But even then, it probably won't make it into the next generation of core chips, namely the 22 nm Sandy Bridge architecture due out next year.
That leaves the market wide open for Nvidia and AMD, for the moment. And indeed, companies like IBM are still on the hunt for a dynamic GPU-friendly architecture. Witness the latest iteration of the iDataPlex line, the dx360 M3 server that sports a pair of PCIe x 16 slots designed for Nvidia's Tesla devices capable of delivering up to a Tflop of processing capability.
There is also a new generation of GPU clustering technology about to hit the channel. NextIO is demoing a new vCORE appliance that sits on top of the rack and can pool up to eight double-wide GPUs. It also has the ability to dynamically reconfigure server/GPU connections on the fly. That means any server in the cluster can access a GPU when necessary, provided it has a suitable PCIe connection. No word yet on when an actual product will be available.
Meanwhile, Nvidia has teamed up with Mellanox for advanced GPU clustering under the InfiniBand protocol. The agreement matches Nvidia's GPUDirect system with Mellanox's ConnectX-2 40 Gbps adapter as a means to offload memory and other functions from CPUs in the cluster. The companies claim they can boost application delivery some 40 percent, even in high-load environments, and cut overall latency by 30 percent.
Now, it can be argued that none of these solutions is geared toward standard enterprise environments. And it's true that the HPC crowd is most likely to take immediate advantage of advanced GPU clustering.
But HPC solutions have a way of trickling down to the mainstream, due to both lower costs and ever-increasing data loads. So even if the top chip maker has foregone discrete GPU processing for the moment, there is still plenty of advanced technology available for those who need it.