Nvidia warms up the datacenter

Nvidia's Fermi GPGPU chip gets squeezed into servers.

Server vendors Supermicro and Appro have managed to work out how to fit Nvidia's Tesla GPGPUs into servers.

Given that the Fermi architecture was engineered to excel at GPGPU tasks, the Green Goblin must have been worried that the chip's thirst for power might lead to vendors shunning the high performance computation (HPC) oriented version of the chip, dubbed Tesla. Instead both Supermicro and Appro have managed to fit Tesla GPUs into a 1U 19-inch rackmount chassis.

That's quite an achievement given that 1U is the same height as a 5.25-inch drive bay, meaning cooling is severely restricted. While Supermicro decided to play it safe and stick two Tesla GPUs in its 1U number crunching hotbox, Appro went whole hog and stuck four Tesla chips into its 1U offering.

Interestingly, Appro opted for Intel's older Xeon 5500 and 5600 chips in its four Tesla server whereas Supermicro offers AMD's 12-core Opteron 6100 series chips in its dual Tesla 1U server. The trade-off, understandable given the cooling restrictions, means that those who want maximum compute power from their Appro server will have to optimise their code, relying less on the CPU. Supermicro has standardised its 1U offering on Intel's quad-core and hexa-core processors.

Supermicro also announced a 4U unit with four Tesla GPUs that has hot-swappable 1400W power supply units. Few would question the computational abilities of either server but rather might be concerned about their ability to fill a rack with them, given the power requirements and heat output of these units.

These boxes are being pitched to the scientific research community as well as natural resource exploration and financial simulation customers.

Both vendors deserve notice for engineering ways to stick Nvidia's latest GPGPU chip in places where few thought it would go. Now we guess they'll find out whether any customers will actually buy them.