Nvidia has introduced its new Grace Hopper superchip, and its huge reminiscence is sufficient to make your entire RTX 40-series weep. Additional than that, it is packing eight of them into what Jensen Huang calls “one large GPU” with the DGX GH200—a completely mega superserver to speed up AI.
The superchip comes with 576GB of GPU reminiscence, together with 96GB of HBM3 reminiscence and 480GB of LPDDR5. Contemplating even the 24GB of GDDR6X on the RTX 4090, it makes for stark comparability to what we’re used to over in GeForce land.
Huang says the superchip has “virtually your entire laptop” on it, together with 72 ARM CPUs, and Nvidia’s grabbing eight of them to stuff contained in the DGX GH200—an excellent couple of racks price of AI servers.
The DGX GH200 is large, each actually and figuratively. We’re 150 miles of fibre optic cables, and 2000 followers that may recycle 70,000 cubic toes per minute of air. “It will possibly in all probability recycle the air on this complete room in a few minutes”, says Huang.
And at 40,000 kilos in weight, the DGX GH200 is “4 elephants, one GPU.” In different phrases, you are not going to suit one in your research. However I suppose that is a part of the great thing about cloud primarily based acceleration—you do not have to.
Not that we may afford one anyway, contemplating the worth this monster is more likely to be. Nvidia did not affirm the worth tag we are able to anticipate, however I might guess on it being eye-watering.
Between the eight Grace Hopper chips within the DGX GH200 sits 900GB/s price of chip to chip hyperlinks from NVLink switches. As soon as mixed, that places it altogether within the area of an exaflop in computing energy.
And naturally, between some sporadic Mandarin standup comedy, Jensen panders to the avid gamers within the viewers with “I’m wondering if this might play Crysis? Solely avid gamers will get that joke.”
In fact Nvidia is not transferring away from gaming, regardless of our musings round a current incomes name, however it’s clear the main focus for the long run is on AI. Contemplating Nvidia simply introduced ACE for Video games, an AI toolset that lets builders create NPCs that may discuss again to you in real-time. Nvidia’s already confirmed that the system will run on Nvidia DGX Cloud however may additionally run on a daily ol’ GPU.