News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
AI assistant can help me overclock? COMPUTEX NV Showcase Experience

Blackwell

Nvidia's GPU architecture for AI datacentres and gaming, launched in 2024

Photo: 极客湾Geekerwan, CC BY 3.0, via Wikimedia Commons

Blackwell is a graphics processing unit microarchitecture from Nvidia, announced at GTC on 18 March 2024 and shipping from Q4 2024. It follows Hopper in the datacentre and Ada Lovelace on consumer cards, and pairs two large dies into one package to push AI compute past the limits of a single silicon chip.

Share on WhatsApp

  • Announced at GTC on 18 March 2024, shipping from Q4 2024
  • GB100 die holds 104 billion transistors at the reticle limit
  • Two GB100 dies linked at 10 TB/s by NV-HBI, 208 billion transistors total
  • GB202 consumer die: 750 mm2 with 24,576 CUDA cores
  • Fifth-gen Tensor Cores add MXFP4 and MXFP6 low-precision formats

What it is

Blackwell is the GPU microarchitecture Nvidia announced at its GTC keynote on 18 March 2024, with products reaching customers from Q4 2024. It covers two markets at once: datacentre accelerators for training and running generative AI models, and consumer chips for gaming and workstation cards such as the GeForce RTX 50 series. The design is named after David Blackwell, an American mathematician known for work in game theory, probability, information theory and statistics. Its main idea is to link two large silicon dies so they behave as one, letting Nvidia go past the size a single die can reach.

How it came about

Nvidia launched its Hopper datacentre architecture in March 2022, and demand for it ran hot through the 2023 AI boom, with H100 server lead times of 36 to 52 weeks. Roadmaps had called the next step simply "Hopper-Next" until an October 2023 investor presentation named Blackwell and the B100 and B40 parts. The official reveal came at GTC in March 2024, focusing on the B100 and B200 accelerators, an eight-GPU HGX B200 board and the 72-GPU NVL72 rack. Gaming was left out of that keynote. A design flaw was reported in October 2024 and fixed with TSMC.

Inside the chip

Blackwell splits into separate dies for datacentre and consumer work. The datacentre GB100 die holds 104 billion transistors and sits at the reticle limit, the largest area lithography can pattern. Two of them are joined by a 10 TB/s link called NV-HBI, built on NVLink 7, and share cache coherently. Fifth-generation Tensor Cores add MXFP4 and MXFP6 low-precision formats, and a second-generation Transformer Engine handles quantisation. Fourth-generation ray tracing cores add a Triangle Cluster Intersection Engine and Linear Swept Spheres for fine detail such as hair. A RISC-V AI Management Processor schedules work on the GPU itself.

How it was made

Datacentre parts use TSMC's custom 4NP process, an enhancement of the 4N node used for Hopper and Ada Lovelace, while consumer parts use 4N. GB100 carries 104 billion transistors, about 30% more than the 80 billion of Hopper's GH100. Because there is no major node jump, gains come from architecture rather than shrinking. Two GB100 dies, 208 billion transistors in total, sit on a silicon interposer made with TSMC's CoWoS-L 2.5D packaging. On the consumer side, the GB202 die measures 750 mm2.

Blackwell
A drawing made by EveryChip from what this page says. It is a diagram, not a picture of the real chip.

Where it went

The datacentre launch centred on the B100 and B200 accelerators, the HGX B200 board carrying eight GPUs, and NVL72, a rack-scale system with 72 GPUs. Nvidia positioned the platform around its ARM-based Grace CPU as well as the accelerators, and quoted backing from the leaders of Google, Meta, Microsoft, OpenAI and Oracle. By November 2024 Morgan Stanley reported that the whole of 2025 Blackwell output was already spoken for. Consumer cards use GDDR7 memory and PCIe 5.0, while datacentre parts use HBM3E and PCIe 6.0. At CES 2025 Nvidia named Flux, Meta AI, Mistral AI and Stability AI models for the platform.

Versions and family

The family runs from datacentre dies under the GB100 codename to consumer dies under GB20x. GB202 is the largest consumer die, with 24,576 CUDA cores, well ahead of GB203, which has less than half that number. At 750 mm2 it is Nvidia's biggest consumer die since the 754 mm2 TU102 of the Turing generation in 2018. Blackwell replaces Hopper in servers and Ada Lovelace on desktops, and Rubin is named as the next architecture after it.

Why it matters

Blackwell arrived as Nvidia's answer to enormous demand for AI compute, following a year in which the company reportedly sold 500,000 H100 accelerators in a single quarter and passed a $2 trillion market value. Its dual-die approach, and the NV-HBI link that makes it work, showed how far packaging could substitute for a bigger single die. Nvidia said around $10 billion went into developing that interconnect. The architecture also moved datacentre roadmaps to yearly releases aimed at both x86 and ARM systems.

Problems and reception

A design flaw reported in October 2024 was described by Jensen Huang as functional in nature and as having pushed yields down; it was corrected together with TSMC. Jim Keller, an engineer behind AMD's K7, K12 and Zen designs, disputed Nvidia's $10 billion interconnect spend, arguing Ultra Ethernet could reach the same result for $1 billion instead of the proprietary NVLink route. Demand stayed strong regardless, with 2025 production reported sold out by November 2024.

Questions people ask about Blackwell

What is Blackwell?

It is Nvidia's GPU microarchitecture announced at GTC on 18 March 2024, covering both datacentre AI accelerators and consumer gaming and workstation chips. It follows Hopper and Ada Lovelace and is named after mathematician David Blackwell.

Why does it use two dies instead of one?

The GB100 die already sits at the reticle limit, the largest area lithography machines can pattern. Pairing two of them with a 10 TB/s NV-HBI link lets the package act like one large chip with shared cache, reaching 208 billion transistors.

Who manufactures it?

TSMC builds Blackwell chips. Datacentre products use the custom 4NP process, an enhancement of the 4N node used for Hopper and Ada Lovelace, while consumer products use 4N. The two dies sit on a CoWoS-L 2.5D silicon interposer.

What is new for AI work?

Fifth-generation Tensor Cores support sub-8-bit MXFP4 and MXFP6 formats, and a second-generation Transformer Engine handles quantisation. Nvidia claims 20 petaflops of FP4 compute for the dual-GPU GB200 superchip, excluding gains from sparsity.

What replaced it?

Rubin is named as the successor to Blackwell. Blackwell itself replaced Hopper in the datacentre and Ada Lovelace for consumer products.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Blackwell (microarchitecture)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

Photo credits on this page

The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

The weekly digest

Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.