AMD data centre GPUs
AMD Instinct GPUs and ROCm: MI300 to Helios
How AMD’s Instinct accelerators have evolved from MI300X to the MI400 series and the Helios rack, how their chiplet architecture differs from other designs, and why ROCm is the software layer that decides whether they fit your workload.
Export controls: AI chips, servers, networking and memory can need an export licence depending on destination, end user and end use under US, UK and EU rules. We screen every project for this. See what applies.
Why AMD matters
- A credible second source. Instinct is the most direct alternative to NVIDIA’s data-centre GPUs, with very large HBM capacity per GPU and a rack-scale system arriving in 2026.
- Chiplet design. Instead of one or two large dies, Instinct packages several compute chiplets on advanced silicon with stacked HBM, which helps AMD pack more memory around the compute.
- Open ecosystem bets. AMD backs open scale-up and scale-out standards (UALink, Ultra Ethernet) and an open software stack rather than a proprietary interconnect.
Generation by generation
- MI300X (2023). CDNA 3 with 192 GB of HBM3 at about 5.3 TB/s. Its memory capacity made it attractive for serving large models on fewer GPUs.
- MI325X (2024). The same compute with 256 GB of faster HBM3E at about 6 TB/s, aimed at memory-bound inference.
- MI350 and MI355X (from June 2025). CDNA 4 with 288 GB of HBM3E, up to about 1.4 kW per GPU for the liquid-cooled MI355X, and new FP4 and FP6 formats. MI350X and MI355X use the same silicon, with the MI355X qualified for higher power.
- MI400 series and Helios (2026). CDNA 5, with compute chiplets on a 2 nm class process and HBM4. The MI455X carries 432 GB of HBM4 and about 320 billion transistors, according to reports. The MI430X adds up to 288 TFLOPS of hardware FP64 for HPC, and the MI440X is an 8-GPU on-premises server.
- Helios rack. 72 MI455X GPUs and 18 EPYC Venice CPUs (Zen 6, up to 256 cores) in one liquid-cooled rack with about 31 TB of HBM4 and Pensando networking. AMD quotes about 2.9 exaflops of FP4. Reports name server makers such as HPE, Lenovo and Supermicro as partners, and cloud deployments as planned, so check what is live today.
Side by side
| MI300XCDNA 3 | MI325XCDNA 3 | MI355XCDNA 4 | MI455X / HeliosCDNA 5 | |
|---|---|---|---|---|
| Status | Shipping since 2023 | Shipping since 2024 | Shipping since June 2025 | Reported in production and shipping in 2026 |
| Memory | 192 GB HBM3 | 256 GB HBM3E | 288 GB HBM3E | 432 GB HBM4 |
| Memory bandwidth | About 5.3 TB/s | About 6 TB/s | About 8 TB/s | About 20 to 23 TB/s (sources differ) |
| Board power | About 750 W | About 1,000 W | Up to about 1.4 kW | Not confirmed in the sources we read |
| Process | 5 nm and 6 nm chiplets | Same compute as MI300X | 3 nm class | 2 nm class compute chiplets, about 320B transistors (reported) |
| New precision | FP8 | FP8 | FP4 and FP6 | MXFP4 and MXFP8: up to 40 and 20 PFLOPS (AMD figures) |
| Scale-up fabric | Infinity Fabric | Infinity Fabric | UALink tunnelled over Ethernet (128-GPU racks shown) | UALink over Ethernet, about 260 TB/s per rack |
| Rack system | 8-GPU platform | 8-GPU platform | 8-GPU board, liquid and air options | Helios: 72 GPUs, 18 Venice CPUs, 31 TB HBM4 |
Bandwidth and rack figures differ between AMD material and press reports, and most performance figures are AMD claims. Treat them as direction, not a guarantee.
ROCm: the platform layer for AMD
- What it is. ROCm is AMD’s open software platform for GPUs: compilers, runtime, drivers and libraries, with PyTorch and Hugging Face support.
- ROCm 7 (June 2025). Added distributed inference through vLLM, SGLang and llm-d, and AMD claimed a large training gain over ROCm 6.
- ROCm.ai (announced July 2026). AMD announced a new generation of tooling, available from August 2026, with a command-line interface and developer tools. AMD claims further inference and training gains over ROCm 7, which is a vendor claim to test.
- What it means for you. ROCm is younger than CUDA, so validate your exact models, frameworks and versions before you commit. That check is part of our assessment. See the platform comparison.
- Keeping it secure. Drivers, runtimes and containers are part of your attack surface, so patching belongs in the maintain stage of our agent security and defence practice.
What it means for your data centre
- Power and cooling. MI355X-class parts and Helios racks are liquid cooled, so facility readiness sets your timeline. See the rack power and cooling guide.
- Networking. Helios uses Ethernet-based scale-up (UALink over Ethernet) and Ultra Ethernet scale-out, which suits teams standardising on Ethernet. See AI networking.
- Memory supply. HBM3E to HBM4 ties availability to memory supply. See memory suppliers.
- Alongside NVIDIA. Many estates run both. Compare with NVIDIA’s generations on your workload before you pick.
Roadmap
- MI500 and the next Helios (planned for 2027). AMD has outlined a further generation after MI400. Treat it as a plan, not a shipping product.
Where this fits in the stack
Common questions
How does MI355X compare with NVIDIA B200?
Both are current-generation accelerators with large HBM3E memory, and the MI355X is liquid cooled. AMD leads on memory per GPU in some configurations, while NVIDIA leads on software maturity and rack-scale NVLink. The right answer depends on your model and software, so benchmark your own workload.
What is Helios?
AMD’s rack-scale system: 72 MI455X GPUs and 18 EPYC Venice CPUs with Pensando networking and UALink over Ethernet scale-up, designed to compete with NVIDIA’s NVL72 racks.
Is ROCm ready for production?
For many mainstream frameworks and models, yes, and ROCm 7 improved distributed inference. Validate your specific stack first, including driver and framework versions.
Can I buy Instinct for any country?
No. Export rules depend on the chip, destination, end user and end use. We screen every project before we specify hardware.
Sources and dating
Figures are as of 18 November 2026, from AMD material as reported by NAND Research, XPU.pub and WhatPSU, plus AMD’s earlier MI300 and MI325 specifications. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against current datasheets when we specify a build. AMD, Instinct, EPYC, ROCm, Helios and related names are trademarks of Advanced Micro Devices, Inc. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: AMD AI Infrastructure · NVIDIA Data Centre GPUs · Intel AI Accelerators · Other Chip Providers · Data Centre Infrastructure
Planning an AMD build
Get the numbers for your workload
