This year at the GPU Technology Conference (GTC18) our roaming camera crew was lucky enough to catch Dolly Wu, Inspur’s VP & GM of their Datacenter/Cloud division. Inspur has been the fastest growing server vendor over the past couple of years, according to Gartner, and we expect the company to become an increasingly important player in the global market.
Wu discusses several topics in the video below, including the introduction of their first Power 9 based server. The new box is all about acceleration and has all of the bells and whistles, starting with NVLink 2.0, which, at 300 GB /s (150 GB/s each way) is 2.8x faster than PCIe 4.0.
This is a big differentiator for the Power-based systems. The IBM fueled boxes can use NVLink to pass data from CPUs to GPUs and GPUs to GPUs at native NVLink rates. However, x86 systems equipped with NVLink still have to communicate through the system PCIe interface, which is limited to 64 GB/s with the newest 4.0 version.
The new Inspur Power 9 system also sports PCIe Gen 4.0, 8 memory channels, and Open CAPI 2.0, which connects to storage or network and boasts a 2x advantage over PCIe 4.0.
Dolly believes that Inspur’s key differentiation is that they work directly with customers (currently in the US, these are hyperscale types) to discover their hardware pain points and then design systems to directly address the problems. They’re a total solution provider, meaning they cover the gamut from x86 2-4-8 way systems, to Power 9 systems, to one of the most GPU dense servers in the industry (8 GPUs in 2U).
They recently announced their Olympus 4-socket system which they designed for Microsoft along with another server they’ve developed with Baidu. This server connects up to 16 GPUs to a single 4-way server. Baidu is using this box for workloads ranging from brain simulations to autonomous driving simulation.
Keeping all of this hardware cool is a problem. With this in mind, Inspur is implementing liquid cooling for their most dense systems – like the 2-way, 8 GPU box mentioned above.