NVIDIA’s Volta technology powers its specially made V100 GPU, which has found its way into data center servers announced by Dell EMC, Hewlett Packard Enterprise, IBM, and Supermicro. All of these data center server systems take advantage of NVIDIA’s powerful V100 GPUs, with none of them sporting only one. This makes them ideal for not just normal number crunching, but also neural networking, machine learning, and high-power serving functions like administering multiple virtual users through thin clients, or serving VR and other high-throughput content. These GPUs are fully compute-oriented, which means that they’re optimized for this sort of purpose.
Dell has unveiled the PowerEdge R740 and an XD variant, which can use up to three V100 GPUs, as well as a PowerEdge C4130 model that can not only use four V100 GPUs, but can even link them up with NVIDIA NVLink technology. HP went a bit more high-end with their offering, and put an AI-friendly eight units in the HPE Apollo 6500 system, though it uses slower PCI-E linkage between components, as well as the less powerful HPE ProLiant DL380, which has three V100 GPUs in it. IBM, meanwhile, will be updating the entire next generation of the IBM Power Systems server hardware series with NVIDIA’s powerful new GPUs in varying numbers. Finally, Supermicro is rounding things out by introducing a workstation, an advanced compute machine for data analytics and number crunching, and three different AI rigs of various power levels.
The Volta-equipped V100 GPU offers up to 120 teraflops of raw power for certain tasks such as machine learning. To put that in perspective, NVIDIA’s newest consumer-grade Titan X gaming GPU offers just over 12 teraflops of GPU compute for gaming. Much of the power in these GPUs comes from optimization, rather than raw computing power in the form of clock speeds, cores, and other hardware details. The Volta technology on board the V100 makes it well-suited to raw compute applications. Meanwhile, it excels on entirely new levels when put into AI situations, thanks to AI-centric features like tensor cores, custom-made bus lines, and custom architecture that puts things that would be used frequently in AI, like memory chips, closer together.