GS-GPU-12Engineering datasheet

AMD Instinct MI325X

Statusshipping
Verified2026-10-14
Download datasheet (PDF) ↓

The Instinct MI325X is the memory-upgraded refresh of the MI300X: the same CDNA 3 compute (304 compute units) with 256 GB of HBM3E at 6 TB/s and a 1,000 W peak board power, launched 10 October 2024. It stays on the OAM form factor with Infinity Fabric, passively cooled, so it slots into MI300-generation platforms where the board supports the higher power. GPU Smith specs it where memory capacity per GPU is the constraint but a move to the MI350 generation (and its ROCm/datatype changes) is not yet justified.

256 GB HBM3E ECC
Memory · 8192-bit
6 TB/s
Memory bandwidth · peak
2.61 PFLOPS
FP8 matrix · dense; 5.22 PFLOPS with sparsity
1.3 PFLOPS
FP16 / BF16 matrix · dense; 2.61 PFLOPS with sparsity
163.4 TFLOPS
FP64 matrix · vector 81.7 TFLOPS
1,000 W peak
Board power · OAM, 54V UBB
Compute01/05
ArchitectureCDNA 3 · TSMC 5nm / 6nm FinFET
Compute units304
Stream processors19,456
Matrix cores1,216
Peak engine clock2,100 MHz
FP8 matrix2.61 PFLOPS · dense; 5.22 with structured sparsity
FP16 / BF16 matrix1.3 PFLOPS · dense; 2.61 with structured sparsity
TF32 matrix653.7 TFLOPS · 1.3 PFLOPS with sparsity
INT82.6 POPS · 5.22 with structured sparsity
FP32163.4 TFLOPS · matrix and vector
FP6481.7 TFLOPS vector · 163.4 TFLOPS matrix
Transistors153 billion
Memory02/05
Capacity256 GB HBM3E · full-chip ECC
Interface8192-bit · 6 GHz memory clock
Bandwidth6 TB/s · peak
Last-level cache256 MB · Infinity Cache
Interconnect & form factor03/05
Form factorOAM module · passive
Host busPCIe 5.0 x16
Infinity Fabric links8 · 128 GB/s peak per link
Power & thermal04/05
Board power (TBP)1,000 W peak
Power delivery54V UBB
CoolingPassive OAM
Software & platform05/05
Software stackROCm
Launch date10 October 2024 · per AMD product page
RASPage retirement, page avoidance · SR-IOV supported
Field notes
  • Compute is identical to the MI300X on paper; the gain is 256 GB at 6 TB/s (33 percent more capacity, 13 percent more bandwidth by arithmetic from AMD's figures). Decode-bound inference benefits; compute-bound work does not.
  • 256 GB lets a 70B model in FP16 (about 140 GB of weights) leave over 100 GB for KV cache on a single GPU.
  • No FP6 or FP4 hardware; if quantised throughput is the plan, the MI350 series is the generation that adds MXFP4.
  • The 1,000 W figure is a peak board rating; validate sustained clocks in your chassis.
Questions we get on this part

How much memory does the MI325X have?

256 GB of HBM3E at 6 TB/s peak bandwidth with full-chip ECC.

MI325X vs MI300X?

Same CDNA 3 compute. The MI325X has 256 GB HBM3E at 6 TB/s versus 192 GB HBM3 at 5.3 TB/s, and a 1,000 W versus 750 W peak board power.

What is the MI325X power draw?

AMD lists 1,000 W peak board power over 54V UBB.

Is the MI325X replaced by the MI350 series?

The MI350 series (MI355X, MI350P) is the newer CDNA 4 generation with MXFP4 and MXFP6; the MI325X remains on AMD's product list.