Blackwell
Nvidia GPU microarchitecture for AI and gaming
Blackwell is a graphics processing unit microarchitecture developed by Nvidia. Announced in March 2024, it succeeds Hopper and Ada Lovelace. It targets generative AI workloads and consumer gaming with new tensor cores and dual-die packaging.
- Uses TSMC 4NP and 4N custom process nodes
- GB100 die contains 104 billion transistors
- Dual-die B100 connects via 10 TB/s NV-HBI link
- Introduces fifth-generation Tensor Cores
- Supports MXFP4 and MXFP6 low-precision formats
What it is
Blackwell is a graphics processing unit microarchitecture created by Nvidia. It serves both datacentre compute needs and consumer gaming markets. The design features dedicated dies for each purpose. Nvidia announced the architecture in March 2024 at its GTC keynote. It replaces the Hopper design for servers and Ada Lovelace for desktops. The name honours mathematician David Blackwell. This generation focuses heavily on generative artificial intelligence tasks.
How it came about
Nvidia revealed the Hopper architecture in March 2022. High demand for Hopper products caused long delivery delays in 2023. The company sold half a million H100 accelerators in one quarter alone. This success pushed Nvidia market value past two trillion dollars. Roadmaps initially listed the next chip as Hopper-Next. In October 2023, Nvidia confirmed the Blackwell name and B100 accelerators. The official launch occurred in March 2024. Reports later noted a design flaw that reduced yields. TSMC and Nvidia collaborated to fix this issue before mass production.
Inside the chip
The architecture uses fifth-generation Tensor Cores for AI calculations. It supports new low-precision formats like MXFP4 and MXFP6. These formats improve efficiency for generative model inference. Ray tracing cores now include a Triangle Cluster Intersection Engine. This helps render fine details such as hair. An AI Management Processor built on RISC-V handles scheduling. This offloads work from the central processor. Consumer chips support DirectX 12 Ultimate and Vulkan 1.4. Compute Capability versions 10.x and 12.x are introduced.
How it was made
TSMC manufactures Blackwell chips using custom process nodes. Datacentre parts use the 4NP node, an enhancement of 4N. Consumer parts use the standard 4N node. The GB100 die holds 104 billion transistors. This is a thirty percent increase over the previous Hopper die. The B100 accelerator combines two GB100 dies in one package. They connect via a ten terabyte per second link. This interconnect uses the NVLink 7 protocol. The package sits on a silicon interposer using CoWoS-L technology.
Where it went
Blackwell powers the GeForce RTX 50 series for gamers. It also drives the RTX PRO Blackwell series for workstations. In datacentres, it forms the basis of the HGX B200 board. The NVL72 rack-scale system uses seventy-two GPUs. Nvidia promotes these for generative AI training and inference. Major tech firms including Google and Meta have endorsed the platform. The architecture supports foundation models from various AI developers. It connects to ARM-based Grace CPUs in server configurations.
Versions and family
The GB100 die serves datacentre applications. The GB202 die is the largest consumer variant. It measures 750 square millimetres and contains 24,576 CUDA cores. This is significantly larger than the previous AD102 die. The GB203 consumer die has less than half the cores of GB202. This gap is wider than in previous generations. The B200 accelerator uses two GB100 dies. The GB100 single die is also available. Successor architecture Rubin is planned for future release.
Why it matters
Blackwell aims to define the generative AI era. Nvidia invested around ten billion dollars in research. This included development of the NV-HBI die interconnect. The architecture allows full cache coherency between dual dies. It acts like a single large piece of silicon. Production for 2025 was reported as sold out by late 2024. The design pushes against semiconductor reticle limits. It demonstrates a shift towards yearly datacentre product releases.
Problems and reception
A design flaw was reported in October 2024. It affected manufacturing yields but was functional. Nvidia and TSMC resolved the issue collaboratively. CEO Jensen Huang described the fix as successful. Industry analysts noted strong pre-orders for the silicon. Some engineers criticized the high cost of the proprietary interconnect. They suggested alternative networking standards could be cheaper. Despite this, major cloud providers committed to the platform.
Questions people ask about Blackwell
What is the Blackwell microarchitecture?
Blackwell is a GPU microarchitecture developed by Nvidia. It succeeds Hopper and Ada Lovelace. It is designed for both datacentre AI workloads and consumer gaming applications.
Who manufactures Blackwell chips?
TSMC manufactures Blackwell chips. They use custom process nodes called 4NP for datacentre parts and 4N for consumer products. These are enhancements of standard TSMC technologies.
How many transistors does Blackwell have?
The GB100 die contains 104 billion transistors. The B100 accelerator uses two of these dies in one package. This results in a total of 208 billion transistors for the dual-die unit.
What is new in Blackwell Tensor Cores?
Blackwell introduces fifth-generation Tensor Cores. They support new microscaling formats like MXFP4 and MXFP6. These allow more efficient low-precision computations for generative AI models.
Good reading and sources
- Nvidia sold half a million H100 AI GPUs in Q3 thanks to Meta, Facebook — lead times stretch up to 52 weeks: ReportTom's Hardware, Article
- All About Nvidia Chips, AI Hype and What Lies AheadBloomberg, Article
- Nvidia CEO Jensen Huang announces new AI chips: 'We need bigger GPUs'CNBC, Article
- GTC Wrap-Up: ‘We Created a Processor for the Generative AI Era,’ NVIDIA CEO SaysNvidia, Article
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 2026-11-15.