AMD Executive Elite System Builder AMD Partner Logo

AMD Instinct MI350P GPU - System Overview

mi350p-4.pngAdditional Information

As part of the Instinct MI350 Series, the AMD MI350P GPU features a PCIe 5.0 x16 interface and is a full-height, full-length, 2-slot design with passive cooling. It draws up to 600W typical board power (TBP), configurable down to 450W, via a 12V-2x6 connector. Its multi-chip design includes 4x accelerated compute dies and a single I/O die, delivering 128 compute units, 8,192 stream processors, and 512 matrix cores at a peak engine clock of 2,200 MHz. The GPU can also be partitioned into up to 4x instances of 36GB each.

The MI350P comes with 144GB of HBM3E memory spread across 4x HBM3E stacks. It uses a 4,096-bit interface and delivers up to 4 TB/s of memory bandwidth. The card also includes 128 MB of AMD Infinity Cache as a last-level cache. Full-chip ECC is supported, along with page retirement and page avoidance.

AMD Instinct MI350P GPU - Specifications

Architecture

- 4th Gen AMD CDNA architecture

Memory

- 144GB HBM3e

Memory Bandwidth

- 4 TB/s

Memory Interface

- 4096 bits

AMD Infinity Cache (Last Level)

- 128MB

Form Factor

- FHFL 2-slot PCIe CEM Card

Maximum Power TBP

- 600W (configurable to 450W)

I/O Interconnect

- 1x PCIe Gen5 x16 (128 GB/s)

Partitions

- Physical Partitions: Up to 4 @ 36GB
- Memory Partitions: 1

Compute Units

- 128

Matrix Cores

- 512

Stream Processors

- 8192

Peak Engine Clock

- 2.2 GHz

Performance

- FP16 VECTOR (TFLOPS): 72
- FP16 VECTOR (TFLOPS) with Sparsity: N/A

- FP16 MATRIX (TFLOPS): 1150
- FP16 MATRIX (TFLOPS) with Sparsity: 2300

- BF16 (TFLOPS): 1150
- BF16 (TFLOPS) with Sparsity: 2300

- INT8 MATRIX (TOPS): 2300
- INT8 MATRIX (TOPS) with Sparsity: 4600

- MXFP8 (TFLOPS): 2300
- MXFP8 (TFLOPS) with Sparsity: N/A

- OCP-FP8 (TFLOPS): 2300
- OCP-FP8 (TFLOPS) with Sparsity: 4600

- MXFP6 (TFLOPS): 4600
- MXFP6 (TFLOPS) with Sparsity: N/A

- MXFP4 (TFLOPS): 4600
- MXFP4 (TFLOPS) with Sparsity: N/A

- FP64 VECTOR (TFLOPS): 36

- FP32 VECTOR (TFLOPS): 72

- FP64 MATRIX (TFLOPS): 36

- FP32 MATRIX (TFLOPS): 72

Decoders

- 2 groups for HEVC/H.265,AVC/H.264, VP9, or AV1

JPEG/MJPEG Codec

- 20 cores, 10 cores per group

Lithography

- TSMC 3nm/6nm FinFET

I/O Dies (IODS)

- 1

RAS Features

- Full-chip ECC memory, page retirement, page avoidance

AMD Instinct MI350P GPU - Documentation

Consult
the Experts!
Talk to one of our trained professionals to create a system to your exact specifications.
Contact Us