Table of Contents
Blackwell (GB202 / GB203 / GB206) is NVIDIA’s 2025 architecture spanning consumer (RTX 5060-5090) and workstation (RTX 6000 Pro) cards. The marketing emphasises FP4; the practical impact is broader.
TL;DR
Blackwell brings 5th gen tensor cores with native FP4 / FP8 / FP16, GDDR7 memory (~2× bandwidth of GDDR6X), improved MIG-style partitioning on the Pro variants, and longer driver support cadence. For AI: ~2× throughput vs Ada at FP8, plus FP4 for the cutting edge.
Architectural changes
- 5th gen tensor cores: native FP4 (NVFP4 / MX-FP4), FP8 (E4M3 / E5M2), BF16
- GDDR7 memory: 1,792 GB/s on 5090 vs 1,008 GB/s on 4090
- L2 cache larger: helps with memory-bound LLM inference
- PCIe Gen 5: doubles host bandwidth (mostly relevant for training)
- Improved power management: better sustained boost clocks under load
AI relevance
- FP8 hardware path: ~1.5-2× throughput vs FP16 with <1% quality drop
- FP4 hardware path: ~2× throughput vs FP8 (limited model support so far)
- Memory bandwidth: ~78% uplift over Ada — directly improves LLM decode
- 32 GB VRAM on flagship consumer card (5090) — first card to fit 14B FP16 single-card
Verdict
Blackwell is a meaningful generational leap for AI inference. FP8 is the practical win; FP4 is the future. For new deployments, Blackwell-class hardware is the right pick.
Bottom line
For AI inference in 2026, default to Blackwell. See 5090 spec breakdown.
