New standard announced today: HBF sits between HBM and enterprise SSDs, using stacked NAND flash (8-16 layers) for up to 512GB capacity and 3TB/s bandwidth.
Why this matters for AI infrastructure: - Current bottleneck: HBM is fast but capacity-limited.
SSDs are capacious but too slow for inference.
HBF targets the middle ground.
- 500B parameter models can fit in HBF without the cost of all-HBM storage - UCIe support means direct connection to CPUs and GPUs - SK hynix and SanDisk are building an open ecosystem for compatible products At BHK Cloud, this is exactly the direction we're positioned for.
Affordable cloud storage ($2.49/TB) paired with GPU compute ($0.15/hr) is the practical foundation for AI teams building on these emerging standards.
Buy Now Pay Later - zero upfront, pay after 1 month.
No free trials, no gimmicks.
Originally posted on reddit