Blackwell
Blackwell is a GPU microarchitecture introduced by NVIDIA, targeting high-performance computing, artificial intelligence training, and inference applications. The architecture improves computational throughput and energy efficiency relative to its predecessor, supports large-scale parallel computing tasks, and incorporates optimizations specific to AI workloads such as large language models. The Blackwell architecture is deployed across multiple NVIDIA GPU product lines serving data center and professional computing domains.
Core Facts
NVFP4 is a 4-bit floating-point (FP4) quantization format introduced by NVIDIA, designed specifically for the hardware acceleration capabilities of Blackwell architecture GPUs
NVIDIA's H100/H200 series GPUs hold more than 80% of the AI training and inference market.
The RTX 5090 has 32 GB of VRAM.
Timeline (last 90 days)
Running this engine requires a Blackwell GPU (the kernels use the NVFP4 hardware path), about 70GB of space on fast NVMe, and 32GB of RAM.
On newer-generation GPUs that support FP8 compute, such as Hopper and Blackwell, FP8 can achieve nearly double the throughput improvement while using about half the memory of BF16.
The weights of NVIDIA's FP4 version of Qwen3.8 cannot be read by upstream vLLM; a specific community-built container must be used, and only Blackwell-architecture GPUs are supported.
The NVFP4 4-bit format can only be executed natively by Blackwell; non-Blackwell platforms (such as Hopper) can only run it as W4A16.
Nemotron 3.5 Super VL replaces a pure attention decoder with Mamba state-space blocks, uses a self-developed RADIO vision encoder, and natively supports NVFP4 four-bit quantization on the Blackwell architecture.
PCIe 5.0 x16 unidirectional bandwidth is about 64 GB/s, while a single Blackwell's HBM bandwidth is on the order of TB/s, so the cost-effectiveness of UVA offload depends on the access frequency of the offloaded data.
TokenSpeed's MLA kernel optimized for the Blackwell architecture has been adopted by vLLM.
Blackwell is the codename of NVIDIA's architecture following Hopper, corresponding to chips such as B100/B200.
NVIDIA is using open-source models to build a moat for its own hardware ecosystem, and the limited support for the NVFP4 format on non-NVIDIA hardware objectively enhances the appeal of the Blackwell platform.
Only Blackwell's fifth-generation Tensor Cores provide FP4 matrix multiplication instructions.
40 more timeline events
All Facts (20)
NVFP4 is a 4-bit floating-point (FP4) quantization format introduced by NVIDIA, designed specifically for the hardware acceleration capabilities of Blackwell architecture GPUs
90%VerifiedNVIDIA's H100/H200 series GPUs hold more than 80% of the AI training and inference market.
90%VerifiedThe RTX 5090 has 32 GB of VRAM.
80%VerifiedCore innovations of the Blackwell architecture include the second-generation Transformer Engine (supporting FP4 precision), the NVLink Switch system (supporting interconnection of 576 GPUs), 20 petaFLOPS of FP4 compute, and 192 GB HBM3e memory (with bandwidth reaching 8 TB/s).
80%VerifiedBlackwell's second-generation Transformer Engine adds FP4 precision support, which in theory can improve Transformer model training throughput by 2–4x compared to the H100.
75%VerifiedNVIDIA designed a dedicated Tensor Core instruction mma.kind::mxf4 for NVFP4 in the Blackwell architecture (SM 100/103, such as the B100 and B200 series), and adopted an FP8 scaling factor mechanism with one scale per group of 16 elements.
75%VerifiedNVIDIA Blackwell 架构于2024年发布，是继 Hopper（H100）之后的新一代数据中心 GPU 架构
75%VerifiedThe NVIDIA Blackwell architecture (GB200/B200 series) introduces native FP4 tensor cores at the hardware level
75%VerifiedBlackwell (B100/B200 series) was released in 2024, manufactured using TSMC's 4NP process, with a single GPU integrating 208 billion transistors
75%VerifiedThe NVIDIA RTX 5090 is based on the Blackwell architecture, is equipped with 32 GB of GDDR7 memory, and has a theoretical FP16 computing power exceeding 100 TFLOPS.
70%VerifiedThe RTX 5070 belongs to NVIDIA's Blackwell architecture consumer graphics card product line and features approximately 12 GB of VRAM.
65%VerifiedThe RTX 5090 is based on the Blackwell architecture, equipped with 32 GB GDDR7 memory with a bandwidth of 1.79 TB/s, a significant leap compared to the RTX 4090's 1 TB/s.
65%VerifiedThe RTX 5090 is based on NVIDIA's Blackwell architecture.
65%VerifiedThe RTX 5090 is based on NVIDIA's Blackwell architecture.
65%VerifiedThe GB300 is NVIDIA's flagship data center chip in the Blackwell Ultra series, integrating HBM3e memory with a single-card capacity of 288 GB.
65%VerifiedThe Blackwell architecture is named after American mathematician and statistician David Blackwell.
65%UnverifiedGB200中的'G'代表Grace CPU，'B'代表Blackwell GPU
95%Unverified2024年GTC上发布了Blackwell架构GPU和NIM（NVIDIA Inference Microservices）等产品
90%Unverified每颗Blackwell GPU拥有高达192GB HBM3e显存
90%Unverified同一组编译选项在 Ampere、Hopper、Blackwell 等不同 GPU 架构上的表现可能截然不同
90%