GPU

AMD Radeon Instinct MI350X

Edit@1 months ago

Intergrated Memory(VRAM)
Capacity

288 GB

(HBM3e 8192-bit)

Bandwidth

8192 GB/s

1170 Token/s

Vector Compute
FP64
72.09 T
FP32
144.18 T
FP16
288.35 T
BF16
X
INT32
X
INT8
X

AMD Radeon Instinct MI350X General-Purpose Float-Point performance (Vector Performance / Scalar Performance)

FP64: 72.09 TFLOPS

FP32: 144.18 TFLOPS

FP16: 288.35 TFLOPS

Matirx Compute
FP64
72.09 T
144.18 T
FP32
144.18 T
288.36 T
FP16
2306.88 T
4613.76 T
FP8
4613.76 T
9227.52 T
TF32
X
BF16
2306.88 T
4613.76 T
INT16
X
INT8
4613.76 T
9227.52 T
INT4
4613.76 T
9227.52 T

AMD Radeon Instinct MI350X AI performance (Tensor Performance / Matrix Performance)

FP64: 72.09 TFLOPS, with sparsity: 144.18 TFLOPS

FP32: 144.18 TFLOPS, with sparsity: 288.36 TFLOPS

FP16: 2306.88 TFLOPS, with sparsity: 4613.76 TFLOPS

FP8: 4613.76 TFLOPS, with sparsity: 9227.52 TFLOPS

BF16: 2306.88 TFLOPS, with sparsity: 4613.76 TFLOPS

INT8: 4613.76 TOPS, with sparsity: 9227.52 TOPS

INT4: 4613.76 TOPS, with sparsity: 9227.52 TOPS

Hardware Specs
AMD Radeon Instinct MI350X is a 3nm chip, has 185000 million transistors, launched by AMD at 2025. It has 288 GB built-in(On-Board/On-Chip) memory with bandwidth up to 8192 GB/s. It has 16384 general-purpose ALUs(CUDA cores/Shader cores) and 1024 matrix cores(Tensor cores) .
Process Node
3 nm
Launch Year
2025
Vector(CUDA) Cores
16384
Matrix(Tensor) Cores
1024
Core Frequency
1000 ~ 2200 MHz
Cache
256MB