AMD unveils Instinct MI350P with 144GB memory, its first 600W PCIe card with 12V-2×6 power connector

AMD Instinct MI350P launches with 4.6 PFLOPS MXFP4 compute

AMD has launched the Instinct MI350P, a new PCIe accelerator based on the CDNA 4 architecture. The card is designed for air-cooled servers and targets AI inference deployments that do not require a full OAM platform.

The MI350P is the PCIe version of AMD’s MI350 series. It has 128 Compute Units, 8,192 Stream Processors and 512 Matrix Cores. AMD lists the peak engine clock at 2.2 GHz. Basically half of the MI350X/355X specs.

The card has 144GB of HBM3E memory across a 4096-bit interface. AMD lists peak memory bandwidth at 4 TB/s. The GPU also has 128MB of last-level cache and supports full-chip ECC memory.

AMD claims up to 4.6 PFLOPS of MXFP4 or MXFP6 matrix performance. The card also reaches 2.3 PFLOPS in MXFP8 and OCP-FP8, 1.15 PFLOPS in FP16 and BF16 matrix workloads, and 36 TFLOPS in FP64. Structured sparsity doubles several listed 8-bit and 16-bit performance figures.

Source: AMD

The MI350P uses a full-height, full-length dual-slot PCIe 5.0 x16 design. AMD lists board length at 10.5 inches, or 267 mm. The card uses passive cooling and a 12V-2×6 power connector. Its default board power is 600W, with a configurable 450W mode. Basicaly this is AMD’s first 12V-2×6 card that can reach 600W.

Specifications
VideoCardz.com Instinct MI350P🆕 Instinct MI350XInstinct MI355X
Picture
Form factor PCIe add-in card OAM moduleOAM module
Architecture CDNA 4 CDNA 4CDNA 4
Process TSMC 3nm / 6nm FinFET TSMC 3nm / 6nm FinFETTSMC 3nm / 6nm FinFET
Stream processors 8,192 16,38416,384
Compute Units 128 256256
Matrix cores 512 1,0241,024
Peak engine clock 2200 MHz 2200 MHz2400 MHz
Transistors 73 billion 185 billion185 billion
HBM memory 144GB HBM3E 288GB HBM3E288GB HBM3E
Memory bus 4096-bit 8192-bit8192-bit
Last-level cache 128MB 256MB256MB
ECC Full-chip ECC Full-chip ECCFull-chip ECC
TBP 600W max, 450W configurable 1000W1400W
External power 12V-2×6 54V UBB54V UBB
Bus PCIe 5.0 x16 PCIe 5.0 x16PCIe 5.0 x16

Source: AMD