Pure sees GPU servers constraining datacenter power budgets

Datacenter operators buying Nvidia GPUs will find their electrical power budgets constrained and buy flash instead of disk storage, says Alex McMullan, Pure Storage International CTO.

Alex McMullan, Pure Storage
Alex McMullan

McMullan offered this opinion at a discussion in London early last week. He thinks that as datacenter operators buy more GPUs, which need 10 kW each, they will find the GPUs eat up more of the overall power budget, leaving less for everything else.

Every rack of networking and storage gear needs power to function. You can reduce storage rack power draw by switching, McMullan says, to efficient flash drives and away from hard disk racks.

McMullan claims that environmental concerns will increasingly drive datacenter operators to manage their electricity consumption more closely and reduce it where they can.

The need for organizations to manage their datacenter power budget while becoming more AI-centric and buying more GPUs means that disk could be squeezed out of datacenters, thus helping to bring about Pure’s contested claim that no more new disk drives will be sold after 2028.

Addressing that claim directly, he said that with 100 TB flash drives coming, and 300TB after them, they would have much more capacity than HDDs, which have a roadmap to 40 or 50 TB in a few years. Based on this, you could imagine one 100 TB SSD would hold more data than three 30 TB disk drives and consequently need less power, less cooling, and less rack space as well.

A flash drive rack with these drives would hold the same amount of data as three and a third disk drive racks, which supports his idea that datacenter operators, when faced with electricity power budget limitations, will prefer high-capacity flash over low-capacity disk drives.

Currently only Pure has a public roadmap to 300 TB flash drives. The commercial SSD manufacturers, apart from Samsung, don’t publish such ambitious capacity roadmaps, McMullan suggested, because the cost of developing such drives in terms of drive-level controller circuitry and firmware to handle the day-to-day read and write operations and allied signal processing, error correction and garbage collection is too high.

He has heard talk of a mainstream flash array vendor developing their own proprietary flash drives to reach greater-than-HDD capacities and so match or attempt to match Pure.

B&F asked him for his thoughts on whether PLC (5bits/cell) flash would be used in flash drive products and he opined: “No. It’s a non-starter.”

That’s because, he said, its endurance is too low, meaning it would require a lot of over-provisioning to produce drives lasting for a five-year warranty period. Also the data access times are longer than with TLC (3bits/cell) and QLC (4bits/cell) flash because of he extra voltage levels involved, as well as the controller circuitry being more complex.

Even QLC flash could become difficult to handle at high capacities, in a controller sense, and there could be a reversion to TLC flash. You could reach QLC drive capacity levels by adding more layers to a TLC die, and the resulting flash chip would not be physically larger than a QLC chip with the same capacity. The resulting extra-layer TLC chip would also have longer endurance than its QLC sibling, faster data access speed, and need less electricity for its operations.

This means that Pure high-capacity flash drives could use TLC NAND and not QLC. But McMullan’s main message was that watts could help kill disk drives as much as raw bit cost, in that flash drive TCO, which takes electricity consumption into account, will be less than HDD TCO. If true, what’s not to like about that?