Up to 16 thousand AMD Instinct MI300X accelerators can be combined in the Oracle cloud

The Oracle Cloud Infrastructure (OCI) cloud platform leverages AMD Instinct MI300X accelerators and the AMD ROCm open software stack as part of the latest BM.GPU.MI300X.8 instances. They are designed for demanding AI workloads, including processing large language models (LLMs) with hundreds of billions of parameters.

BM.GPU.MI300X.8 Bare Metal instances combine eight Instinct MI300X accelerators, each of which carries 192 GB of HBM3 memory with a bandwidth of 5.3 TB/s. It is claimed that the Hugging Face OPT model with 66 billion parameters can be run on a single Instinct MI300X.

The instances include two Intel Xeon Sapphire Rapids processors with 56 cores. The amount of DDR5 RAM is 2 TB. Eight NVMe SSDs with a capacity of 3.84 TB each are used to store data. The instances include a frontend network with a throughput of 100 Gbps, and the cluster uses 400G (×8) connections.

Image Source: AMD

The subscription price for BM.GPU.MI300X.8 is $6/hour for each accelerator. The OCI Supercluster platform based on AMD Instinct MI300X can combine up to 16,384 accelerators, which allows customers to flexibly scale computing resources to suit their tasks, the company says. It is noted that the emergence of new instances will provide OCI clients with more opportunities in terms of organizing high-performance and efficient computing when implementing complex AI projects.

In addition, the company also offers clusters with NVIDIA accelerators: 32768×A100, 16384×H100, 65536×H200 and 3840×L40S. And in the first half of 2025, a zettaflops cloud AI supercomputer consisting of 131,072 NVIDIA B200 (Blackwell) accelerators will become available to customers.