Biren BR100
AI GPU for data centre compute workloads
The Biren BR100 is a graphics processing unit designed for artificial intelligence and high-performance computing in data centres. Built on a 7nm process with 77 billion transistors, it uses a chiplet design to deliver high memory bandwidth and compute power for deep learning tasks.
- Contains 77 billion transistors on a 7nm process
- Uses two GPU compute tiles with 256MB L2 cache
- Features 64GB of HBM2E memory
- Supports PCIe Gen5 and CXL interfaces
- Includes dedicated tensor cores and video engines
What it is
The Biren BR100 is a graphics processing unit built for heavy computational tasks in data centres. It targets artificial intelligence training and inference, offering an alternative to established market leaders. The chip features a large die size and high transistor count to support complex mathematical operations required by modern machine learning models. It supports standard interfaces like PCIe Gen5 and CXL for integration into server systems.
How it came about
Biren developed the BR100 to compete in the high-end AI accelerator market. The design was presented at Hot Chips 34, highlighting its architectural choices. The company focused on cost efficiency and performance relative to existing solutions like the NVIDIA A100. The development involved creating a modular chiplet-based structure to allow for different product variants without complete redesigns.
Inside the chip
The processor uses a single instruction multiple thread architecture. It divides compute units into flexible chunks of execution units. A dedicated tensor core handles matrix mathematics, while a tensor data accelerator manages data movement. The chip introduces a TF32+ data type, which sits between TF32 and FP32 precision. Each streaming processing center includes 8MB of L2 cache, totaling 256MB across the two compute tiles. Video encoders and decoders are integrated to assist with video-based AI workloads.
How it was made
Manufactured on a 7nm process node, the chip contains 77 billion transistors. The total die area is 1074mm squared, with each compute tile measuring 477mm squared. It employs a chiplet design with two GPU compute tiles and two HBM2E memory packages. The die-to-die interconnect offers 896GB per second of bandwidth. This modular approach allows Biren to create the BR104 variant for PCIe slots without a new tape-out.
Where it went
This GPU is intended for data centre servers running AI and general-purpose compute workloads. It fits into OAM platforms and supports the OCP OAM Universal Baseboard. Systems using this chip, such as the Biren Hearten server, likely house eight GPUs with external connectivity. It is designed to handle deep learning tasks and video processing efficiently within enterprise hardware environments.
Versions and family
The BR100 is designed for OAM modules, while the BR104 variant targets PCIe connections. Both share the same underlying modular design, allowing Biren to reuse the core architecture. The BR104 does not require a new design layout, leveraging the chiplet structure of the main BR100. This strategy enables the company to address different server form factors with minimal engineering overhead.
Why it matters
The chip represents a significant effort by a Chinese firm to enter the global AI hardware market. With 77 billion transistors, it rivals the scale of contemporary competitors. Its performance claims suggest it exceeds the NVIDIA A100 in deep learning scenarios. The inclusion of advanced interconnects and high-bandwidth memory aims to provide a competitive option for data centre operators seeking alternatives to dominant suppliers.
Questions people ask about Biren BR100
What is the Biren BR100 used for?
It is a GPU designed for data centre compute and artificial intelligence workloads. It handles deep learning training and inference, as well as general-purpose computing tasks in server environments.
How much memory does the BR100 have?
The chip includes 64GB of HBM2E memory. This is distributed across two memory packages connected to the GPU compute tiles via high-speed interconnects.
What is the difference between BR100 and BR104?
The BR100 is designed for OAM modules, while the BR104 is a variant for PCIe slots. They share the same modular chiplet design, allowing the BR104 to be produced without a new full design layout.
What process node is used for the BR100?
The chip is manufactured using a 7nm process technology. It contains 77 billion transistors and has a total die area of 1074mm squared.
Good reading and sources
There is no English Wikipedia article on Biren BR100 yet, so the facts on this page come from the sources listed above: tomshardware.com/pc-components/gpus/blacklisted-chinese-ai-gpu-champ-b, servethehome.com/biren-br100-gpu-for-datacenter-compute-and-ai-workloa. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
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.