Nvidia's Vera Rubin CPU, GPU roadmap charts course for hot-hot-hot 600 kW racks

Nvidia's rack-scale compute architecture is about to get really hot. On stage at GTC on Tuesday, CEO Jensen Huang teased the tech titan's next-generation of high-end datacenter GPUs and CPUs codenamed Vera, Rubin, and Rubin Ultra. This is going to get a little complicated, so just bear in mind: Vera is a CPU architecture, …

[UnknownUnknown]

My Dad’s House

That would be ideal to heat my elderly ‘hot like the Sahara’ Dad’s house. He would save tons on his gas billl, and there is plenty water. Broadband is a bit crummy as a little rural but 5G available and the electricity is smart meter so we can bill you for that..I’m sure we can get BT to do a business fibre or SoGEA circuit.

He doesn’t drive anymore se we could get probably 8 standard cabs in the garage.

[Roger Greenwood]

Power

Cooling is one thing, but imagine the size of the backup battery/UPS/generators required for these things. Put a few together and pretty soon you need your own power station, no wonder SMRs are being worked on, and taken seriously.

[DS999]

Re: Power

Are they using battery backup for AI datacenters? This isn't like traditional servers where you're a bank that can't afford to lose a single transaction, LLM training can save checkpoints and restart if power is lost, and interference could redirect queries to another datacenter.

You'd need battery backup for the cooling since the heat doesn't immediately disappear just because of power loss, but I don't know that you'd need it for the AI server hardware itself. Anyone here actually working with this stuff who can comment?

I think the SMRs are to avoid dealing with the grid because they might have to wait years for the power they need to become available if new high lines have to be built or existing ones have to be expanded. If they can generate the power locally they can put the DC almost anywhere without regard for local power pricing or availability.

[Anonymous Coward]

Surely, you're joking.

You've got to love a good Feynman book reference........

[cyberdemon]

Re: Surely, you're joking.

And no doubt, "Feynman" will be a 1MW rack..

Given that these cards plug in to 48V busbars, that's 20kA needed on the busbar for 1MW, or 12.5kA for 600kW...

Surely you're joking, indeed, Mr. Feynman!

Each rack of "GPUs" will need half a dozen racks of power supplies at that rate! And cables the girth of firehoses.. Just mental.

[IvyKing]

Re: Surely, you're joking.

They may have to go to 380VDC busbars, which would reduce current to 1.6kA. Still pretty mental.

[HuBo]

To SMT or not to SMT

Interesting that the 88-core Vera CPU will have 2x SMT. It's rare in ARM (it's not in Neoverse N and V but apparently made its way at 2x in E1).

I can't remember seeing SMT on Ampere's ARM CPU roadmaps. They are very datacenter- and efficiency-oriented but I guess favor very large core counts instead (eg. 512), as would I.

Meanwhile, yes, IBM POWER9/10/11 (Spinal Tap CPU coming this year) have done SMT4 and SMT8 for a while, and Sun's historical 4-way, 8-cored, CMT8, Victoria Falls chip (256 threads per box) was once a thing, a boat anchor, an iceberg, that sunk the company.

Not sure what Nvidia's up to with this (what kind of workloads will benefit? Desktop?).

[Anonymous Coward]

A 600kW heater is usally the size of a 20ft container

..and would typically heat something like 10L a second of cold water to 90C, for example in a big building like a hospital or hotel.

600kw doesn't really belong in one rack. It doesn't really belong in a regular datacentre either.

[Decay]

600 kw/h is 2 million btu!!!!!!!!!!

If the average datacenter is about 100,00 square feet and lets assume 12 foot ceiling height you could heat the volume from 20 degrees C to 60 in 45 minutes with one rack and nothing else in the datacenter creating heat, no heat loss or gain through the envelope. This assumes a completely empty space, no other racks or equipment in the building.