Gigabyte AMD Instinct MI300X 192GB HBM3 OAM GPU Accelerator

The Gigabyte AMD Instinct MI300X 192GB HBM3 OAM GPU Accelerator (GV-AMI300X-192G) is built on AMD's CDNA 3 architecture, integrating a unified GPU compute die and HBM3 memory stack within a single OAM (OCP Accelerator Module) form factor. The MI300X APU features 192GB of HBM3 memory across eight stacks with an aggregate memory bandwidth of 5.3 TB/s, making it the highest memory capacity and bandwidth configuration available in its class at launch. This architecture eliminates the CPU-GPU memory boundary by unifying compute and memory on a single package, enabling massive model weights to reside entirely on-accelerator without offloading.

Designed specifically for large-scale AI inference and training, the GV-AMI300X-192G delivers up to 1,307.4 TFLOPS of FP8 compute performance and 383.9 TFLOPS at BF16, supporting the full ROCm software ecosystem including PyTorch, TensorFlow, and JAX. The OAM mechanical standard ensures compatibility with OCP-compliant UBB (Universal Baseboard) server platforms from major hyperscale and enterprise ODM vendors, enabling dense multi-accelerator node configurations with coherent high-speed interconnect via AMD Infinity Fabric.

For enterprise datacenter operators targeting generative AI, large language model (LLM) serving, and high-performance computing workloads, the GV-AMI300X-192G provides a memory footprint sufficient to run 70-billion-parameter and larger transformer models in full precision without tensor parallelism across multiple devices. The accelerator is suited for deployment in AI factories, national research computing clusters, and enterprise private cloud environments where throughput per rack unit, memory capacity, and software ecosystem openness are primary selection criteria.

Ideal for

• Large language model (LLM) inference serving for models exceeding 70B parameters, fitting full model weights in 192GB on a single accelerator without multi-device tensor parallelism
• Generative AI training runs including foundation model pre-training and fine-tuning at scale within OCP-compliant multi-accelerator server nodes
• High-performance scientific computing and simulation workloads in national labs, research institutions, and engineering environments requiring high memory bandwidth and FP64 throughput
• Enterprise private AI cloud build-outs where open-ecosystem ROCm software compatibility and avoidance of vendor lock-in are architectural requirements
• Multimodal AI and diffusion model inference requiring large activation memory and sustained high-bandwidth memory access patterns
• Data center consolidation projects targeting maximum AI compute density per OAM slot within existing OCP Universal Baseboard infrastructure

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.