Will Your AI run on AMD?

More teams are taking a hard look at AMD chips this year, and the reason is simple: there aren’t enough NVIDIA GPUs to go around. The newest ones are sold out into late-2026, with a backlog in the millions of units, and most of what’s available has already been committed.

You can get AMD chips today, and they’re every bit as capable. A single MI350X carries 288GB of memory, enough to hold a large model on one card, posts strong inference numbers on benchmarks, and serves tokens at a lower cost than most teams are used to paying.

So why isn’t everyone already building on them? Mostly because of rewrite fear, months spent porting code off CUDA, NVIDIA’s software layer –  before anything runs.

That fear is outdated, though. For most teams, AI on AMD chips just runs, and for most use cases – the switch is just a toggle instead of a complete rewrite.

Most “CUDA code” isn’t really CUDA code

The fear assumes your code is full of CUDA. It mostly isn’t. You write your models in PyTorch [the framework nearly all AI is built on], and PyTorch is what talks to the GPU for you. CUDA sits inside PyTorch’s layer, not in the code you wrote. Your models aren’t really tied to NVIDIA, they run wherever PyTorch runs.

And PyTorch runs on AMD, through AMD’s own version of CUDA, called ROCm. Install the AMD build of PyTorch and your code runs unchanged.

This isn’t a workaround AMD bolted on. PyTorch ships and maintains AMD builds right alongside its NVIDIA ones, updated with each ROCm release.

AMD has a direct equivalent for each NVIDIA tool, so most of the stack maps across without changes. Its porting guide lists them:

Same jobs, different names, and installing PyTorch for AMD wires all of it up for you.

What does need work: porting custom kernels

A kernel is a code someone writes by hand, to squeeze the last bit of speed out of the GPU, for one specific job. It’s specialist work, usually only teams chasing the last few percent of performance. If your team wrote custom kernels for NVIDIA, those take effort to move.

Everything else, the open-weight models, fine-tuning, standard serving, comes across as-is.

Even a custom kernel isn’t a “from-scratch” rewrite. AMD’s HIPIFY tool converts the routine parts of CUDA to AMD’s HIP automatically, and AMD is upfront that the hand-written parts are where an engineer steps in.

AMD puts a typical port at a few hours to a few weeks, depending on the code.

What moves across, and what doesn’t:

New models land on AMD with day 0 support

The models your team wants to run are already there. The people who build Llama, DeepSeek and the rest, and the teams behind the serving tools vLLM and SGLang, get them running on AMD themselves, usually the day each model ships.

That is day-zero support: a model comes out and runs on AMD the same day it runs on NVIDIA. ROCm 7 has made it routine.

·       Llama 4, served in both FP8 and the newer MXFP4 format

·       DeepSeek V4, day-0 on SGLang

·       Qwen3.6, across MI300X to MI350X

·       Gemma 4, the whole family on vLLM and SGLang

·       OpenAI’s gpt-oss, day-0 on ROCm 7

·       GLM-5.2, which ships with AMD support in its release

When the models you want are ready on launch day, “will it run” stops being a question.

AMD caught up, fast

If AI on AMD works this well, why wasn’t it the obvious choice already? You might ask.

It’s because it only recently became one. A couple of years ago the AMD software was young, and independent testers at SemiAnalysis noted gaps in how models ran on it.

The hardware was strong; the software needed to catch up.

And oh it caught up fast. AMD poured serious engineering into ROCm, and by early 2026 the same testers had seen roughly doubling inference speed on the same chips through software alone – which matched or beat the alternatives on a growing list of workloads.

Progress hasn’t stalled, either. AMD is still investing heavily in R&D to push it further.

The gap was software, and AMD closed most of it in about a year.

So don’t judge AMD on last year’s numbers. Run your own MI350x model on the chips and see how it does on your workload today.

Run your AI workload AMD MI350x, now available on Neysa

Neysa Velocis, our AI cloud, offers AMD chips alongside current NVIDIA’s, so you can run your workload on either, and compare the two.

Velocis is the only SemiAnalysis’s ClusterMAX recommended cloud provider in India – it graded Neysa on what production depends on: reliability, networking, security and operations. 

For a team moving to AMD, that means:

• One platform, notebook to production. Spin up an MI350X in a managed Jupyter or VS Code notebook, with an idle-killer that stops billing the moment you stop working, then serve the same model from a dedicated, single-tenant endpoint. VMs, Kubernetes where you own the control plane, or bare metal, your choice.
  
• Engineers who know AI infra. MLOps and infrastructure specialists who tune kernels, and size clusters available 24×7; with direct-to-engineer escalation and a 30-minute response on critical issues. If a custom kernel needs porting, you have people beside you, not a ticket queue.
  
• Infrastructure built for AI. WEKA parallel storage and an Arista GPU network keep the chips fed with data at full speed, so moving to AMD costs you nothing at the storage or networking layer.
  
• Reliability and governance. A high uptime SLA across compute, storage and network, ISO 27001 and SOC 2, private networking, role-based access, and data that stays in India.
  
• Costs you can see. Per-minute metering and invoices you can split by project, so the savings from a move to AMD show up where finance can trace them.

Your move

The chips are available, your code already runs on them. Talk to a Neysa expert and put your AI on AMD →

FAQs