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
Nvidia B300

Nvidia B300

AI accelerator with dual-die design and 288 GB memory

The Nvidia B300, also known as Blackwell Ultra, is a graphics processing unit designed for artificial intelligence workloads. Manufactured by TSMC on the 4NP node, it features 208 billion transistors and 288 GB of HBM3E memory to support large language models.

Share on WhatsApp

  • Contains 208 billion transistors on TSMC 4NP node
  • Features 288 GB of HBM3E memory capacity
  • Delivers 15 petaFLOPS of dense NVFP4 compute
  • Uses dual-reticle design with 10 TB/s interconnect
  • Doubles SFU throughput for attention operations

What it is

The Nvidia B300 is a high-performance graphics processing unit built for artificial intelligence tasks. It belongs to the Blackwell family and is also referred to as Blackwell Ultra. The chip uses a dual-reticle design, connecting two silicon dies to function as a single accelerator. It contains 208 billion transistors and provides 15 petaFLOPS of compute power using the NVFP4 precision format. This hardware targets data centres that run large-scale AI models and inference services.

How it came about

The B300 builds upon the original Blackwell architecture. Nvidia introduced it to address the growing demands of AI factories and real-time services. The design focuses on scaling performance and efficiency beyond the previous generation. It retains the CUDA programming model while introducing new silicon innovations. The chip was developed to support trillion-parameter models without offloading data to slower storage.

Inside the chip

The processor features 160 Streaming Multiprocessors organised into eight clusters. Each multiprocessor holds 128 CUDA cores and four fifth-generation Tensor Cores. These cores support matrix operations in formats such as FP8 and NVFP4. A custom interface called NV-HBI links the two dies at 10 TB per second. The chip includes 288 GB of HBM3E memory with 8 TB per second bandwidth. Special Function Units handle attention-layer calculations twice as fast as the prior model.

How it was made

TSMC manufactures the B300 using its 4NP process node. The dual-die layout allows for a high transistor count of 208 billion. This construction method enables greater silicon utilisation per square millimetre. The memory subsystem uses high-bandwidth modules to ensure rapid data access. The packaging supports coherent memory accesses across the shared L2 cache.

Where it went

This GPU powers AI training and reasoning applications in data centres. It supports models with over 300 billion parameters entirely on-chip. The hardware accelerates transformer attention layers for faster response times. It is suitable for large-batch pre-training and low-batch inference. System integrators use it in rack-scale deployments like the GB300 NVL72.

Versions and family

The B300 is part of the Blackwell Ultra series. It succeeds the standard Blackwell GPU, offering 1.5 times more NVFP4 throughput. Compared to the Hopper H100, it delivers a 7.5-fold increase in specific compute metrics. The available count of streaming multiprocessors and memory capacity may vary by specific stock keeping unit.

Why it matters

The chip aims to lower costs per token generated in AI services. Its efficiency improvements help reduce total processing cycles per query. By keeping large models on-chip, it minimises latency in interactive applications. The architecture supports up to 576 GPUs in a non-blocking fabric via NVLink 5.

Questions people ask about Nvidia B300

What is the Nvidia B300?

The B300 is a graphics processing unit from Nvidia, also known as Blackwell Ultra. It is designed for artificial intelligence workloads in data centres, featuring a dual-die design and high-capacity memory.

How much memory does the B300 have?

The chip includes 288 GB of HBM3E memory. This provides 8 TB per second of bandwidth, allowing large AI models to reside entirely on the chip without needing external storage.

Who manufactures the B300?

TSMC manufactures the B300 using its 4NP process node. The design is by Nvidia, which integrates the silicon into systems for AI training and inference tasks.

What makes B300 different from Blackwell?

The B300 offers 1.5 times more NVFP4 compute throughput than the standard Blackwell GPU. It also features doubled Special Function Unit speed for faster attention-layer processing in AI models.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Nvidia B300 yet, so the facts on this page come from the sources listed above: developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-power, tomshardware.com/pc-components/gpus/nvidias-future-blackwell-ultra-gpu. 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.

    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.