Updated
Updated · Wccftech · Aug 4
SK hynix, SanDisk Unveil 512GB HBF Standard for AI Accelerators as Bandwidth Reaches 3TB/s
Updated
Updated · Wccftech · Aug 4

SK hynix, SanDisk Unveil 512GB HBF Standard for AI Accelerators as Bandwidth Reaches 3TB/s

3 articles · Updated · Wccftech · Aug 4

Summary

  • SK hynix and SanDisk introduced High Bandwidth Flash as a new memory tier for AI accelerators, positioning stacked NAND between HBM and enterprise SSDs to ease inference bottlenecks.
  • Up to 512GB per package and 0.4TB/s to 3TB/s bandwidth are meant to let large models keep hot data in HBM while storing far more parameters in nearby flash.
  • A 500-billion-parameter model would need roughly 500GB, making full placement in HBM uneconomic and giving HBF a role as capacity expansion rather than an HBM replacement.
  • UCIe support is intended to let HBF connect across CPUs and GPUs and help build a broader compatible ecosystem through the Open Compute Project.
  • The launch also sharpens competition in AI storage-memory architectures, with Samsung already pushing NAND-based CMX solutions for faster GPU offloading in AI servers.

Insights

As chipmakers bypass traditional standard bodies, will their high-speed gamble revolutionize AI infrastructure or dangerously fragment the market?
With AI pushing hardware to its limits, could this new middle-tier memory be the secret to cheaper, faster inference models?