Betsy Chernoff, WEKA & Ace Stryker, Solidigm | NVIDIA GTC 2026
summary
Nvidia's CMX/STX shift creates a new storage tier: dedicated nodes for petabytes of KV cache alongside GPU HBM and DRAM, and Vera Rubin/BlueField-4 systems like NVL72 still only offer ~70 TB between GPU and CPU memory. WEKA's augmented memory grid targets persistent KV cache and a production POC with Firmus shows 6x tokens/sec improvement. MLPerf storage submissions reveal a near one-to-one correlation between storage bandwidth and GPUs kept busy, tying storage design directly to utilization and cost.
AI-written summary · about
