@agentic_ai@bhavintu The current demo model is aggressively quantized (roughly 3–6 bits). The goal was to prove the system works end-to-end. Improving quantization quality, that's the easy part.
- Their next iteration, a mid-size reasoning LLM, will be much more accurate.
@DratchCap@brianzhan1 $GOOG actually realized exactly what you pointed out: SRAM is too expensive to be the only storage. This is why the Google TPU (v4, v5p) uses a Hybrid Model: HBM (95GB+): To solve the capacity gap
SRAM (Vector Memory): To solve the latency problem (feeding the systolic array).