Blackwell, Decoded
A plain-spoken tour of NVIDIA’s Blackwell architecture — the chips, engines, and rack-scale systems built for the age of AI reasoning. Every spec from the technical brief, explained for people who know software but are new to AI hardware.
In 2024–25, AI stopped being only about training bigger models and started being about making them think. That shift — toward models that reason step-by-step before answering — demands far more compute, delivered in real time. Blackwell is NVIDIA’s answer.
This course turns a dense 31-page technical brief into eight short, visual modules. Work through them in order for a complete picture, or jump to whatever you need. Your progress is saved on this device as you go.
#Start the course
Why Blackwell?
Why AI suddenly needs a new kind of computer — and what “reasoning” changes.
□ 5 min · The shift to reasoning AIMODULE 02The Three Scaling Laws
Pre-training, post-training, and test-time scaling — the three ways compute buys intelligence.
□ 6 min · How compute buys intelligenceMODULE 03The Blackwell GPU
Inside the chip: two dies fused into one, 208B transistors, and the new low-precision math.
□ 8 min · A new class of chipMODULE 04The Engines
The five specialized engines: Transformer, Decompression, RAS, Secure AI, and attention.
□ 9 min · Five specialized acceleratorsMODULE 05NVLink & Scaling Up
How 5th-gen NVLink and NVLink Switch make 72 GPUs behave as a single giant one.
□ 6 min · Turning many GPUs into oneMODULE 06GB300 NVL72 Rack
The flagship GB300 NVL72 rack: full specs and the 50×/35×/30×/25× gains.
□ 9 min · The flagship reasoning systemMODULE 07GB200 & HGX Servers
The GB200 NVL72 and the air-cooled HGX B300 / B200 servers for x86 data centers.
□ 8 min · The wider Blackwell familyMODULE 08Inference Deep-Dive
How trillion-parameter models are split across GPUs to balance speed and throughput.
□ 7 min · Parallelism for huge modelsWhy reasoning AI needs new hardware; what makes the Blackwell GPU a “new class” of chip; how its engines and NVLink turn 72 GPUs into one; and how the GB300 NVL72 delivers 50× the AI-factory output and 25× lower cost per token versus the previous generation.
Educational material generated by Claude from NVIDIA’s published Blackwell Architecture Technical Brief (V2.1). Figures are NVIDIA projections “subject to change” and are reproduced here for learning. All trademarks belong to NVIDIA Corporation. Not affiliated with or endorsed by NVIDIA.