AMD Instinct MI300X
AI accelerator GPU with 192GB HBM3 memory
The AMD Instinct MI300X is a data centre graphics processor launched in mid-2023. It targets artificial intelligence workloads by offering 192GB of high-bandwidth memory. This design allows large AI models to run on fewer chips than competing solutions.
- Contains 153 billion transistors
- Features 192GB of HBM3 memory
- Delivers 5.2TB/s memory bandwidth
- Uses four CDNA3 GPU tiles
- Launched in mid-2023
What it is
The AMD Instinct MI300X is a specialised graphics processing unit designed for artificial intelligence tasks in data centres. AMD launched this chip in mid-2023 as part of its broader Instinct MI300 family. The device focuses on providing massive onboard memory to handle large language models and other AI workloads. It contains 153 billion transistors and uses four GPU tiles based on the CDNA3 architecture. Its primary distinction is the inclusion of 192 gigabytes of HBM3 memory, which exceeds the capacity of many contemporary competitors.
How it came about
AMD introduced the MI300 series to challenge NVIDIA’s dominance in the AI hardware market. The company aimed to offer an alternative at a time when demand for NVIDIA GPUs led to long supply delays. By launching in mid-2023, AMD sought to capture share from customers who could not wait for other solutions. The MI300X was positioned directly against the NVIDIA H100, targeting high-end AI inference and training markets rather than older generations of accelerators.
Inside the chip
The chip uses a modular design built on a large silicon interposer. This base connects up to eight stacks of high-bandwidth memory and four sets of compute tiles. For the MI300X, all four central tiles are GPU units using CDNA3 IP. This configuration provides 5.2 terabytes per second of memory bandwidth. The large memory pool allows users to run bigger AI models entirely within the GPU memory. This reduces the need to split data across multiple cards or use slower external links, lowering overall system complexity and power use.
How it was made
The MI300X comprises 13 distinct chiplets. These include the base interposer, four network-on-chip units with Infinity Fabric and cache, and eight GPU chiplets. The total transistor count reaches 153 billion. The chip does not include HBM memory stacks or structural silicon in this count. It is packaged in an OAM (Open Accelerator Module) form factor. This standard allows it to fit into existing server infrastructure designed for high-performance accelerators.
Where it went
This processor is intended for data centre environments running artificial intelligence workloads. It supports large-scale AI inference, where models generate responses or predictions. The high memory capacity enables the execution of larger, more accurate models without spanning multiple GPUs. AMD has committed to software compatibility with platforms like PyTorch and Hugging Face. The chip also supports CXL Type-3 devices, allowing for future memory expansion via connected modules.
Versions and family
The MI300 family includes other configurations. The MI300A combines 24 Zen 4 CPU cores with CDNA3 GPU cores and 128GB of HBM3. This variant powers the El Capitan supercomputer. A theoretical CPU-only version could feature up to 96 Zen 4 cores. The modular nature of the platform allows AMD to mix CPU and GPU tiles in different ratios. For example, a future part might use three GPU tiles and one CPU tile for tightly coupled processing.
Why it matters
The launch represented AMD’s most significant attempt to compete in the AI accelerator market. By offering more memory than the NVIDIA H100, AMD provided a technical advantage for specific large-model workloads. The availability of the chip in 2023 gave customers an option during a period of constrained supply for rival products. It established a foundation for AMD’s high-performance compute strategy beyond traditional central processing units.
Problems and reception
Observers noted the focus on memory capacity rather than raw performance metrics at launch. Some commentators questioned the lack of published benchmarks, floating point operations per second, or thermal design power figures. Early discussions highlighted the potential for running large models in memory but expressed caution regarding inference speeds on smaller models compared to established alternatives.
Questions people ask about AMD Instinct MI300X
How much memory does the MI300X have?
The AMD Instinct MI300X includes 192 gigabytes of HBM3 memory. This large capacity allows it to store bigger AI models directly on the chip, reducing the need to spread data across multiple processors.
What is the MI300X used for?
It is designed for artificial intelligence workloads in data centres. Specifically, it targets AI training and inference tasks that benefit from high memory bandwidth and large onboard storage for complex models.
How does it compare to the NVIDIA H100?
The MI300X offers more onboard memory than the NVIDIA H100, which tops out at 96GB in some configurations. AMD positions it as a direct competitor for high-end AI inference, aiming to reduce the number of GPUs needed for large models.
Does it have CPU cores?
No, the MI300X is a GPU-only part. It uses four CDNA3 GPU tiles. Other variants in the MI300 family, such as the MI300A, combine GPU tiles with Zen 4 CPU cores, but the X model is dedicated to graphics and AI acceleration.
Good reading and sources
There is no English Wikipedia article on AMD Instinct MI300X yet, so the facts on this page come from the sources listed above: servethehome.com/amd-instinct-mi300-is-the-chance-to-chip-into-nvidia-, tweaktown.com/news/100187/amd-details-instinct-mi300x-mcm-gpu-192gb-of. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.