Single HBM stack reads Wikipedia 20x per second, HBM5 to double bandwidth by 2028
lemire · x · 2026-09-29
- AI processors use High Bandwidth Memory; the current HBM4 family exceeds 2 TB/s per stack (up to 64 GB, often 32 GB per stack), with accelerators stacking multiple modules for even higher total bandwidth.
- Perspective: all uncompressed Wikipedia text is 100 GB, so a single HBM stack can read it over 20 times per second.
- HBM5 is expected by 2028 at roughly twice the speed (>4 TB/s per stack), reading Wikipedia 40+ times per second. Since bandwidth is often the bottleneck, 2028 AI chips should be much faster on this basis alone.
More from Infra
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16
- Colossus 2's 1.2M B200s could generate ~$400B/year in tokens at DeepSeek Flash rates — bookwormengr · 2026-11-16
- BIS issues rule blocking remote access to US AI chips via third-country data centers — teortaxesTex · 2026-11-16
- Qdrant 1.20.1: 50% fewer cold-storage reads, ~2x faster quantization, agent-ready setup — qdrant_engine · 2026-11-16