整櫃式型態,整合自家CPU、GPU、NVLink及InfiniBand等高速網通技術。其中,
GB200又分為NVL36及NVL72,每櫃分別搭載達36及72顆GPU,而NVL72將為
NVIDIA主推組態,但因其設計較為複雜,估計2024年底將以NVL36先導入前期試煉,
以求快速進入市場(Time to Market)。” (節錄自 TrendForce)
(備註:NVL 72為NVIDIA完整產品,NVL 36架構使客戶保有部分客製化空間)
Comparison of NVIDIA AI Chip Specifications
Memory Clock
Memory Bandwidth
VRAM
Interconnects
GPU
GPU Transistor Count
TDP
Manufacturing Process
Architecture
GB200
H100
B200
B100
8Gbps HBM3E
2x8TB/sec
384GB
8Gbps HBM3E
8Gbps HBM3E
5.23Gbps HBM3
3.35TB/sec
2x NVLink 5 (1800GB/sec)
2x Blackwell GPU
416B
2700W
TSMC 4NP
Grace + Blackwell
8TB/sec
192GB
NVLink 5 (1800GB/sec)
Blackwell GPU
8TB/sec
192GB
NVLink 5 (1800GB/sec)
Blackwell GPU
208B
208B
1000W
700W
TSMC 4NP
TSMC 4NP
Blackwell
Blackwell
80GB
NVLink 4 (900GB/sec)
GH100
80B
700W
TSMC 4N
Hopper
Note: All related data and information are based on official NVIDIA announcements
Source: TrendForce, Mar., 2024
輝達AI晶片 B系列與H系列比較表(圖表來源:TrendForce)