Nvidia has yet to announce any new consumer GPUs since the GTX 1070 Ti last year, but in the meantime, the company is talking up its DGX-2. It's the first single server capable of delivering a whopping two petaflops of computational power. It's equipped with 16 Tesla V100 GPUs based on the company's Volta architecture.
This monstrously powerful system uses a new GPU interconnect fabric called Nvidia NVSwitch that allows the GPUs to simultaneously communicate at a record speed of 2.4 terabytes per second. From Nvidia's vantage point, this makes the DGX-2 in effect a single GPU.
"The world wants a gigantic GPU, not a big one, a gigantic one, not a huge one, a gigantic one," Nvidia CEO Jensen Huang said right before unveiling the DGX-2 at its GPU Technology Conference (GTC) that is currently taking place. Each GPU can communicate with every other GPU at 300GB/s, thanks to the NVSwitch topology.
Nvidia didn't build the DGX-2 to run Crysis, though it surely could without breaking a sweat (with appropriately optimized drivers, of course). The system is designed for deep learning, and to that end, Nvidia says it wields the deep learning processing power of 300 servers occupying 15 racks of datacenter space, while being 60x smaller and 18x more power efficient.
"Clearly the adoption of GPU computing is growing and it’s growing at quite a fast rate," Huang said. "The world needs larger computers because there is so much work to be done in reinventing energy, trying to understand the Earth’s core to predict future disasters, or understanding and simulating weather, or understanding how the HIV virus works."
Each of the Volta GPUs inside the DGX-2 has 32GB of HBM2 memory, for a total of 512GB. (The 32GB Tesla SKU is also new.) The DGX-2 also boasts 30TB of NVMe SSD storage, a pair of Intel Xeon Platinum processors, 1.5TB of system memory, high-end networking amenities, and a dozen of the aforementioned NVSwitches. Each of those switches offers 5x higher bandwidth than the fastest PCIe switches, which Nvidia says will help developers break through previous system limitations and run much larger datasets.
This is not a system you could bring home and plug into your wall socket. Fully equipped with 16 GPUs, the DGX-2 is powered by a 10kW power supply (redundant 5kW, probably), according to AnandTech. It also weighs a hefty 350 pounds and costs $399,000.
Nvidia has begun taking orders for the DGX-2 and has lined up buyers in Cray, Hewlett Packard Enterprise, IBM, Lenovo, Oracle, Supermicro, and Tyan. Skynet is apparently also interested in buying a few, or so we've heard.