NVIDIA B200
芯片
GPU
发货中
Blackwell: two reticle-limit dies fused into one GPU — the 2025 workhorse.
您正在查看一条芯片记录：该处理器是什么、其关键规格、谁设计它、谁制造它、谁供应其零部件，以及它被用在哪里。
Blackwell hit the reticle limit and kept going: the B200 joins two 104-billion-transistor dies with a 10 TB/s die-to-die link so they behave as one 208-billion-transistor GPU. It introduced FP4 precision for inference, moved to CoWoS-L packaging, and in its rack-scale GB200 form made liquid cooling mandatory at the frontier. Every major cloud deployed it through 2025.
规格与关键参数
每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。
Overview
Designer
NVIDIA
Architecture
Blackwell
Announced
March 2024 (GTC); volume through 2025
Status
Shipping
Manufacturing
Foundry / process
TSMC 4NPVendor spec
Transistors
208 billion (2 × 104B dies)Vendor spec
Die-to-die link
NV-HBI — 10 TB/sVendor spec
Packaging
CoWoS-L, 8 HBM3E stacksVendor spec
内存
Memory
192 GB HBM3E (8 × 24 GB stacks)Vendor spec
Memory bandwidth
8 TB/sVendor spec
算力
BF16 tensor (dense)
2.25 PFLOPSVendor spec
NVIDIA datasheets often headline throughput with structured sparsity; figures here are dense unless noted.
FP8 tensor (dense)
4.5 PFLOPSVendor spec
FP4 tensor (dense)
9 PFLOPSVendor spec
FP4 with second-generation Transformer Engine is new in Blackwell
电力
TDP
up to ~1,000 (HGX, configurable); 1,200 in GB200 configurations WVendor spec
Physical
Form factors
SXM (HGX B200), GB200 superchip
Interconnect
NVLink 5 — 1.8 TB/sVendor spec
价格
Price
~$30,000–40,000 per GPU indicated by NVIDIA leadership in 2024Reported
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。 +9 更多分组见下方。
设计方
由……制造
工艺节点
封装
封装服务商
内存技术
内存供应商
内部包含
支持方
前代产品
来源
NVIDIA Blackwell page ↗ · DGX B200 datasheet ↗
记录审阅截至 2026-01-15 — 查看 方法论.