February 14th, 2025

SanDisk's New High Bandwidth Flash Memory Enables 4TB of VRAM on GPUs

SanDisk has introduced High Bandwidth Flash (HBF) memory technology, providing up to 4TB of VRAM for GPUs, targeting AI applications with high throughput, lower power consumption, and future open standard plans.

Read original articleLink Icon
SanDisk's New High Bandwidth Flash Memory Enables 4TB of VRAM on GPUs

SanDisk has unveiled its new High Bandwidth Flash (HBF) memory technology, which aims to provide up to 4TB of VRAM for GPUs, matching the bandwidth of high bandwidth memory (HBM) while offering significantly higher capacity. This technology is designed for AI inference applications that require high bandwidth and capacity with low power consumption. HBF utilizes multiple high-capacity 3D NAND arrays accessed in parallel, allowing for enhanced performance. The architecture involves stacking multiple HBF core dies, interconnected using through-silicon vias (TSVs), on a logic die that can access these arrays simultaneously. SanDisk's HBF is expected to be a cost-effective alternative to HBM, delivering 8 to 16 times the capacity at a similar price point. While it may not match DRAM in latency, it is targeted at read-intensive applications where high throughput is more critical than ultra-low latency. SanDisk has plans for future generations of HBF and aims to establish it as an open standard within the industry. However, specific performance metrics and endurance details have not yet been disclosed.

- SanDisk's HBF technology enables up to 4TB of VRAM for GPUs.

- HBF matches HBM bandwidth while offering higher capacity and lower power consumption.

- The technology is designed for AI inference applications requiring high throughput.

- SanDisk aims to establish HBF as an open standard in the industry.

- Performance metrics and endurance details for HBF are currently undisclosed.

Link Icon 4 comments
By @layla5alive - about 2 months
The lack of write endurance on these would make these chips throwaway after some amount of use. As much as the performance characteristics would be excellent, I really hope we don't go down the path of building GPU ASICs which will eventually totally cease to function when their NAND arrays fail. I'd hope they would then fall back to functioning without NAND, but there's almost no chance modern corporations would care about supporting their ability to function after that happens... if we do this, parts of computing history will be very difficult to preserve long term. And I know Apple is already doing this. sigh.
By @throwaway81523 - about 2 months
Maybe Optane got discontinued too early.
By @dummydummy1234 - about 2 months
So, what is the bandwidth? What is the bandwidth once the cache is reached.