Nvida launched a monster field yesterday called the HGX-2, and it’s the stuff that geek goals are manufactured from. It’s a cloud server that’s presupposed to be so highly effective it combines excessive efficiency computing with synthetic intelligence necessities in a single exceptionally compelling package deal.
You recognize you need to know the specs, so let’s get to it: It begins with 16x NVIDIA Tesla V100 GPUs. That’s good for 2 petaFLOPS for AI with low precision, 250 teraFLOPS
for medium precision and 125 teraFLOPS for these occasions while you want the best precision. It comes commonplace with a half a terabyte of reminiscence and 12 Nvidia NVSwitches, which allow GPU to GPU communications at 300 GB per second. They’ve doubled the capability from the HGX-1 launched final 12 months.
Paresh Kharya, group product advertising and marketing supervisor for Nvidia’s Tesla knowledge heart merchandise says this communication pace permits them to deal with the GPUs primarily as a one large, single GPU. “And what that permits [developers] to do is not only entry that large compute energy, but in addition entry that half a terabyte of GPU reminiscence reminiscence as a single reminiscence block of their applications,” he defined.
Sadly you gained’t be capable to purchase one in all these packing containers. In reality, Nvidia is distributing them strictly to resellers, who will doubtless package deal these infants up and promote them to hyperscale datacenters and cloud suppliers. The great thing about this method for cloud resellers is that once they purchase it, they’ve your entire vary of precision in a single field, Kharya mentioned
“The good thing about the unified platform is as corporations and cloud suppliers are constructing out their infrastructure, they’ll standardize on a single unified structure that helps your entire vary of excessive efficiency workloads. So whether or not it’s AI, or whether or not it’s excessive efficiency simulations your entire vary of workloads is now potential in only a single platform,”Kharya defined.
He factors out that is notably necessary in massive scale datacenters. “In hyperscale corporations or cloud suppliers, the primary profit that they’re offering is the economies of scale. If they’ll standardize on the fewest potential architectures, they’ll actually maximize the operational effectivity. And what HGX permits them to do is to standardize on that single unified platform,” he added.
As for builders, they’ll write applications that benefit from the underlying applied sciences and program within the actual degree of precision they require from a single field.
The HGX-2 powered servers shall be out there later this 12 months from associate resellers together with Lenovo, QCT, Supermicro and Wiwynn.