Alex Woodie, of HPCwire, wrote an excellent article on the rise of SRAM.
SRAM-based chips are emerging as a key architecture for AI inference, addressing the “GPU memory wall” bottleneck that limits how much data can be cached during AI sessions. Unlike HBM used in GPUs, SRAM sits on-chip directly, delivering 100–150 TB/s bandwidth versus HBM’s 1–2 TB/s per stack. Nvidia acquired Groq for $20 billion to gain its SRAM-based LPU. D-Matrix’s Corsair chiplet (now in full production) delivers 2,400 TFLOPS FP8 at 150 TB/s from a PCIe card. Cerebras went public at a $56 billion valuation. Gimlet Labs is building a multi-silicon inference cloud optimizing workloads across both GPU and SRAM architectures.
Go here to get the full and very worthwhile story.
