The AMD MI300X utilizes AMD Infinity Fabric Link technology which improves I/O connectivity delivering a bi-directional memory bandwidth of 128GB/s between each of the installed GPUs. Designed on CDNA 3.0 architecture, this card is a powerful server GPU accelerator with 192 GB of HBM3 memory. It has a 8192-bit memory interface with a peak memory bandwidth of 5.3 TB/s. This card houses 19456x Stream Processors, 304x Compute Units, and 1216x Matrix Cores. Peak half precision (FP16) performance comes in at 1.3 PFLOPs, peak single precision (FP32) performance comes in at 163.4 TFLOPs, and peak double precision performance comes in at 81.7 TFLOPs.
Designed with a passive cooling solution, it features an internally shared 256MB AMD Infinity Cache between compute units. On an industry-standard-based universal baseboard (UBB 2.0) platform with 8-node configuration, it offers a bi-directional Infinity Fabric bandwidth of 896GB/s. individually, each card holds 8x Infinity Fabric Links which have peak speeds of 128 GB/s. The theoretical speed would be 1 TB/s. The thermal design point (TDP) of this card is 750W with a recommended system PSU with a rating of 1150W.