Machine learning workloads have an insatiable appetite for DRAM capacity, and the cost of scaling that capacity is a growing concern. Flash memory is significantly cheaper per gigabyte than DRAM, which raises an obvious question: can it be used to satisfy that demand? A discussion at Hot Chips 2026 takes up this idea under the banner of High Bandwidth Flash, or HBF.
The premise is straightforward. Instead of relying solely on DRAM, systems could tap into flash memory's cost advantage. The challenge is that flash is traditionally slower, so making it viable for ML requires delivering it at high bandwidth. HBF appears to be a proposal for exactly that, though the source does not provide technical specifications.
As with any new memory technology, the feasibility depends on whether the bandwidth gap can be closed enough for real workloads. The article frames HBF as a possible way out of the DRAM capacity crunch, but it stops short of declaring a solution. The source is a single preview, so there is no competing viewpoint to compare against.