AMD Instinct MI450 Helios 2026: Is ROCm Finally CUDA-Ready?

For three years, the honest answer to “can we run this on AMD?” was “for inference, if you’re patient.” The AMD MI450 Helios rack changes that. It’s a 72-GPU, liquid-cooled system with 432GB of HBM4 per GPU, and it enters production deployments this quarter. OpenAI’s first 1GW tranche and Oracle’s 50,000-GPU supercluster get the first racks. With rack-scale AMD capacity finally being built, we can settle the ROCm 7 vs CUDA argument with real workloads instead of vendor slides. Our short answer: ROCm is ready for most production inference. For large-scale training, treat it as ready only once you’ve proven it on your own code.

What’s actually new with AMD MI450 Helios

Helios is the first rack-scale system AMD designs and sells as a single product, and it’s AMD’s answer to Nvidia’s NVL72. Each rack holds 72 Instinct MI455X GPUs (the MI450-series flagship and the only part that ships in Helios) across 18 compute trays. The trays pair the GPUs with 6th Gen EPYC “Venice” CPUs and Pensando “Vulcano” NICs. AMD rates the rack at 2.9 exaflops of dense FP4, 1.4 exaflops of FP8, 31TB of HBM4, and about 1.7 PB/s of aggregate memory bandwidth. Each GPU gets 432GB of HBM4, and the MI455X combines TSMC N2 compute dies with N3 I/O dies. GPUs communicate over UALink-over-Ethernet on Broadcom Tomahawk 6 switches, which delivers 260 TB/s of scale-up bandwidth. Scale-out runs at 2,400 Gb/s per GPU. There’s no NVLink anywhere in the rack.

The customer list is what makes this more than a spec sheet. OpenAI’s 6GW agreement starts with a 1GW MI450 deployment in the second half of 2026. The first tranche of AMD’s warrant for up to 160 million shares vests when that gigawatt goes live, so AMD has a financial reason to deliver. Oracle committed to a 50,000-GPU OCI supercluster starting in calendar Q3 2026, with expansion in 2027. Since then, Meta has signed up for as much as 6GW and Microsoft Azure plans to use Helios for inference. Anthropic announced up to 2GW, backed by an AMD equity investment of up to $5 billion. At Advancing AI 2026 in July, AMD said Helios was in full production, with revenue shipments starting at the end of Q3 and ramping into the first half of 2027. Reported rack pricing is about $5.25 million.

The software shipped alongside the hardware. ROCm 10.0 arrived in late August 2026 with ROCm.ai, which adds a new rocm CLI, “AMD Skills” plugins for Claude Code, Cursor and Codex, and Hyperloom, an agent that automates inference optimization. AMD claims an average 3.3x inference gain and 2.4x training gain over ROCm 7 on the same MI355X hardware. Read the fine print, though. Math libraries for gfx1250 (the MI455X’s architecture target) have been landing since ROCm 7.14, but the ROCm 10 changelog still lists rocSHMEM support for MI455X as single-node only.

Why it matters

• Memory per GPU is the headline. 432GB against Rubin’s 288GB means a trillion-parameter-class MoE model at FP4 fits on fewer GPUs, and long-context KV caches have more room before you have to shard. For inference-heavy teams, that’s worth more than peak FLOPS.
• Real pricing leverage for renters. A credible Nvidia alternative in 2026 GPU clouds gives buyers a second quote at rack scale. Even if you stay on CUDA, AMD capacity will probably lower what you pay.
• An open scale-up fabric. UALink-over-Ethernet on merchant Tomahawk 6 silicon breaks NVLink lock-in at the rack level. That matters to anyone building multi-vendor fleets.
• Frontier labs are now doing ROCm’s QA. About 14GW of commitments from OpenAI, Meta and Anthropic means someone will write, tune and upstream AMD kernels for frontier-scale models, whether or not you do it yourself.
• Inference is ahead of training. AMD’s own numbers (3.3x inference vs 2.4x training) and the single-node rocSHMEM caveat point the same way. Serving is production-grade today; multi-node training on MI455X is still unproven.
• Supply diversification. When Rubin allocations tighten, and they will, running on both vendors keeps you shipping.

How to use it today

MI455X instances aren’t generally rentable by the hour yet. You can still validate your stack on MI355X (gfx950), which is generally available on OCI and other clouds and runs the same ROCm 10 software. Here’s the playbook we’d run.

1. Audit the repo for CUDA-only dependencies. Most PyTorch code ports without changes. The trouble lives in extensions, pinned wheels, and hand-written kernels.

   grep -rnE "cupy|bitsandbytes|apex|flash_attn|nvidia-|\.cu\b|cuda_runtime" \
     --include="*.py" --include="*.txt" --include="*.toml" --include="*.cu" .

2. Check the node with the ROCm CLI. Before you install anything, run rocm examine to see what the machine has and what’s missing.

   curl -fsSL https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.sh | sh
   rocm examine --json

3. Install ROCm and PyTorch from TheRock wheels. ROCm 10 ships as pip-installable SDKs. Swap gfx950 for your target.

   python -m venv rocm-env && source rocm-env/bin/activate
   pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ \
     "rocm[libraries,devel,device-gfx950]"
   pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ \
     "torch[device-gfx950]" "torchvision[device-gfx950]" torchaudio
   rocm-sdk version
   rocm-sdk test

4. Sanity-check PyTorch. On ROCm, PyTorch keeps the torch.cuda namespace, so most device code works unchanged.

   import torch
   print(torch.version.hip)              # set on ROCm builds, None on CUDA
   print(torch.cuda.is_available(), torch.cuda.device_count())
   print(torch.cuda.get_device_name(0))
   x = torch.randn(8192, 8192, device="cuda", dtype=torch.bfloat16)
   print((x @ x).float().abs().mean())

5. Serve your model with vLLM. The official ROCm image covers gfx942 and gfx950. Pin a tag you’ve tested.

   docker run -it --rm --network=host --ipc=host \
     --device /dev/kfd --device /dev/dri --group-add video \
     --security-opt seccomp=unconfined \
     -v ~/.cache/huggingface:/root/.cache/huggingface \
     --entrypoint bash vllm/vllm-openai-rocm:v0.25.0
   
   # inside the container
   vllm serve Qwen/Qwen2.5-72B-Instruct --tensor-parallel-size 8 --port 8000

6. Benchmark tokens per dollar. Ignore AMD’s “up to 30% more tokens per dollar” claim until you’ve measured your own traffic shape.

   vllm bench serve --model Qwen/Qwen2.5-72B-Instruct \
     --dataset-name random --random-input-len 2048 --random-output-len 512 \
     --num-prompts 500 --max-concurrency 64

   Divide output tokens per second by the instance’s hourly price, then run the same test on your current Nvidia instance.

7. Port custom kernels and test collectives. Use HIPIFY for CUDA C++, move new kernels to Triton where you can, and run RCCL all-reduce before you trust any multi-GPU training number.

   hipify-perl my_kernel.cu > my_kernel.hip.cpp
   
   git clone https://github.com/ROCm/rccl-tests && cd rccl-tests
   make HIP_HOME=$ROCM_PATH
   ./build/all_reduce_perf -b 8 -e 8G -f 2 -g 8

   If you use Claude Code, Cursor or Codex, install AMD Skills from the marketplace. AMD argues that coding agents writing against its open compiler stack erase the porting cost that used to be CUDA’s moat. That’s plausible for inference kernels, and it’s the claim to test first.

How it compares: AMD MI450 Helios vs Nvidia Vera Rubin NVL72

Spec	AMD Helios (72x MI455X)	Nvidia Vera Rubin NVL72
HBM4 per GPU	432GB	288GB
HBM4 per rack	31TB	20.7TB
Aggregate memory bandwidth	~1.7 PB/s	~1.4–1.6 PB/s (Nvidia has quoted both)
FP4 compute per rack	2.9 EF dense	2.5 EF NVFP4 training / 3.6 EF NVFP4 inference
Scale-up fabric	UALink-over-Ethernet, 260 TB/s	NVLink 6, ~260 TB/s
Scale-out per GPU	2,400 Gb/s (Vulcano)	1,600 Gb/s
Software	ROCm 10 + ROCm.ai; MI455X multi-node support still maturing	CUDA 13.x; mature everywhere
Shipping	Revenue shipments end of Q3 2026, ramp into H1 2027	Partner shipments autumn 2026
Rack price	~$5.25M (reported)	Not published

On paper, Helios wins on memory capacity and scale-out bandwidth, the two specs that matter most for serving large models. Nvidia still wins on software depth, and nobody has independently validated rack-scale results on either platform. Treat both vendors’ per-rack claims as projections until MLPerf numbers and real cloud pricing arrive.

What’s next

The next two quarters are about execution. Watch whether OpenAI’s first gigawatt goes live on schedule; the vesting warrant makes that milestone public. Watch whether Oracle publishes MI455X shape names, regions and per-hour pricing, because that’s when “Nvidia alternative GPU cloud 2026” becomes something you can budget for. And watch the ROCm changelogs. AMD has moved to a roughly six-week release cadence, and the line we care about is MI455X rocSHMEM and RCCL moving from single-node to multi-node across the UALink fabric.

Further out, AMD has committed to yearly Helios refreshes. The CDNA 6-based MI500 is due in 2027 with copper and optical interconnect options, followed by the MI600 in 2028. Analysts note that AMD’s architectural disclosures are thinner than Nvidia’s, so plan around shipped products, not roadmap slides. Nvidia’s Rubin Ultra will join the fight in 2027.

Our take: teams that mostly serve open-weight models through vLLM or SGLang should qualify AMD now. The software is good enough and the memory advantage is real. Teams training frontier-scale models with custom CUDA kernels should run a pilot on MI355X now and wait for Q4 multi-node data before committing to rack-scale MI455X capacity.

Frequently Asked Questions

Is ROCm finally CUDA-ready for production?

For inference on mainstream PyTorch, vLLM and SGLang models, largely yes. ROCm 10 claims 3.3x inference gains over ROCm 7 on the same hardware, and frontier labs run production traffic on it. For large multi-node training with custom CUDA extensions, it isn’t proven on MI455X yet, so budget engineering time for porting and validation.

When can I rent AMD MI450 Helios capacity?

Revenue shipments start at the end of Q3 2026, and Oracle’s 50,000-GPU OCI supercluster is the first public cloud deployment. Broad on-demand availability will realistically arrive in late 2026 and early 2027. Until then, MI355X instances are the best way to test the same software stack.

What’s the difference between MI450 and MI455X?

MI450 is the product series name used in the OpenAI and Oracle announcements. The MI455X is the series flagship and the part that actually ships inside Helios racks. It carries 432GB of HBM4 and about 40 petaflops of MXFP4 compute.

Do I need to rewrite my CUDA code for ROCm 7 or ROCm 10?

Plain PyTorch code usually runs unchanged, because ROCm builds keep the torch.cuda API. Custom CUDA C++ kernels need HIPIFY or a Triton rewrite. Libraries with CUDA-only binaries, historically including some quantization and attention packages, need ROCm builds or replacements.

Is AMD MI450 Helios cheaper than Nvidia Vera Rubin?

AMD claims up to 30% more tokens per dollar than Vera Rubin NVL72, but that figure comes from internal modeling, and Nvidia doesn’t publish rack pricing. Reported Helios pricing is about $5.25 million per rack. The number that matters is your own tokens per second divided by the hourly price, so benchmark both.

Why does 432GB of HBM4 matter?

It fits bigger models and longer context windows on each GPU, so you need less tensor parallelism and less cross-GPU traffic. For memory-bound inference, especially large MoE models and long-context agent workloads, that usually matters more than peak FLOPS.