AMD AI DevMaster Hackathon

Join the global online competition for developers, researchers, students, AI practitioners, and open-source contributors. Build innovative AI applications accelerated by AMD Radeon™ GPUs and the ROCm™ software stack and compete for a share of the $30,000 USD prize pool.

Scroll down for FAQs :backhand_index_pointing_down:

Why Join? :gpu:

  • $30,000 USD in prizes
  • Free AMD Radeon™ GPU access for eligible participants
  • Three innovation tracks: Agentic AI, Multimodal AI, and Physical AI
  • Compete solo or in teams of up to 3 members

Bring your ideas to life, showcase what you can build with AMD and compete with innovators from around the world.

:cat: Register today and start building!


❓Frequently Asked Questionts

How to access Radeon Cloud resources​:red_question_mark:

  • Please make sure logging in to Radeon Cloud using the same email address you registered for AMD AI Developer Program. This will ensure your access can be properly matched and activated.
  • Detailed GPU cloud setup and login instructions are available in the project repository. Please review the guidance before getting started. If you encounter any access issues, please reach out through the hackathon support channels.

Will I receive a certificate for participating in the hackathon​:red_question_mark:

  • Based on the final evaluation results, participating teams will have the opportunity to receive the following certification certificates:
    Gold Award Certificate
    Silver Award Certificate
    Bronze Award Certificate
    Excellent Award Certificate
  • In addition, all participants who successfully complete the competition and submit a valid project will receive a certificate of completion.
  • These certificates will be issued by AMD to recognize participants’ professional expertise and innovative spirit in AI, GPU computing, and innovative application development.

Is the Demo video required, what format do you accept​:red_question_mark:

  • Yes it’s mandatory, any format is fine. Bilibili, YouTube, or cloud drive links are all accepted as long as the link opens and can be played.

What must be included in the final submission​:red_question_mark:

  • Every submission needs to contain the project documentation saved as a PDF file, the complete source code together with a clear README, and a demo video that is between 3 to 5 minutes long. The demo video is mandatory. You may also include a PPT or poster if you wish, but it is optional.

What is the required PR title format​:red_question_mark:

  • When you submit your project, please follow this exact title format for your pull request: start with the track number (Track x), followed by your team name, and then the name of your application. This helps the organizers quickly identify and organize all submissions.

Does the platform support Model APIs / OpenAI-compatible endpoints​:red_question_mark:

  • Under Deploy Type, choose vLLM or SGLang to deploy an OpenAI-compatible endpoint with one click. There is also a free shared API (Qwen / DeepSeek) that does not consume your credits. See the User Guide for details.

Can the GPU only be used for inference, or can I also do training​:red_question_mark:

  • Both inference and training are supported. You are encouraged to try it yourself.

What models can run on 48GB VRAM​:red_question_mark:

  • In FP16: Qwen3-27B, Qwen3-35B-A3B, Gemma 4 26B-A4B, etc.
  • With INT4 quantization: Even 70B models are feasible.
  • ComfyUI and Stable Diffusion run very well. Feel free to test and see what works for your use case.

I haven’t received my credits yet, what to do​:red_question_mark:

  • Credits are issued based on your Luma registration email. Make sure your Radeon Cloud login email matches your Luma email. If you still haven’t received credits after confirming this, contact the organizer.

What should I do after I finish using my instance​:red_question_mark:

  • Remember to destroy your instance when you are done.

What are the recommended settings when creating an instance on Radeon Cloud​:red_question_mark:

  • Image: Choose GH-proxy-stable (recommended, works out of the box)
  • Storage: Choose Persistent (PVC) — your data will survive even if the instance is destroyed
  • SSH: Add your public key in your Profile first, then enable the SSH toggle when creating the template. Some images require you to install sshd manually (see User Guide).
6 Likes

Shared my Track-2 submission - RepoMind (on-prem coding agent + CUDA→ROCm, measured on W7900). Full write-up + repo here: GitHub - SRKRZ23/repomind-amd-devmaster: On-prem agentic coding assistant + CUDA to ROCm migration, 100% on AMD Radeon (ROCm). AMD AI DevMaster Track 2. · GitHub

Happy to answer questions.

Hello , I just tried to reddem these credits but it is not available anyone help me out: Credit redemption temporarily unavailable System maintenance is in progress. Credit redemption is temporarily unavailable. Please contact the administrator if you need credits.

Registered last week, got email saying I registered with a link to July 9th event (which is past) and no other links to resources, no other email telling me how to go forward. HELP!

Txs,

A

Hey there, if the issue persists, please reach out to devcloudrequests@amd.com

You should receive a link from ai_dev_contests@amd.com, please send me a DM if you don’t get your account setup email.

1 Like

Hi @Ayllee_AMD I haven’t received email for account setup up, it is very confusing!!

I have NOT received email from ai_dev_contests@amd.com,

Please help. More than a week has gone by without access, seems like there’s no actual way to participate in this hackathon.

1 Like

I have not received it either. Please help

2 Likes

Where to share the repository link, so that they review it and respond. Have you submitted this work to the AMD AI DevMaster Hackathon Team?

Hi everyone, we’ve added setup and access instructions at the top of this thread, do check them out!

The Radeon Cloud website is inaccessible from Canada, even when using any VPN. I cannot open the website to claim these credits.

1 Like

I will help you later I send the link for the instructions

Here are the instructions for redeeming your points Radeon-hackathon-2026-07/Radeon-Cloud-User Guide/README.md at main · AMD-DEV-CONTEST/Radeon-hackathon-2026-07 · GitHub

Looking for Hackathon tips straight from our experts? On 31 July, 2026 we will be doing a live session with speakers from AMD, Hugging Face / LeRobot and Dify. Mark the date!

Hi, it could be a local firewall issue. Some participants were able to resolve the issue by allowing the Radeon Cloud site through their firewall. We also suggest trying the following:

  • Clear browser cache and cookies
  • Try a different browser (Edge/Google Chrome…)
  • Switch to a different VPN endpoint/region (or temporarily disable the VPN)
  • Try a different network connection

If you need further help, feel free do send me a DM.

Four ways I got GPU power measurement wrong on Radeon Cloud (and how I fixed them)

I spent this hackathon measuring the real energy cost of local LLM inference on a W7900. Every one of these bugs produced plausible-looking numbers, which is what made them dangerous. Posting them here in case they save someone a GPU hour.

1. The Radeon Cloud host has 8 GPUs, and sysfs exposes all of them.

/sys/class/drm/card*/device/hwmon/hwmon*/power1_average will happily give you readings from a card that isn’t yours. I was measuring another user’s GPU for a full session before noticing.

Fix — resolve your assigned card by PCI bus first:

rocm-smi --showbus          # e.g. 0000:43:00.0
# then match PCI_SLOT_NAME in /sys/class/drm/card*/device/uevent

2. power1_average lags about 18 seconds.

It’s a long-window moving average. My first benchmark ran 2–5 second bursts and reported 14 W under full load — physically impossible, but it looked like a plausible idle-ish number rather than an obvious error.

Fix: sustained runs of 90 s, discarding the first 25 s of samples. You can see the ramp yourself:

# while a load is running
watch -n1 'cat /sys/class/drm/cardN/device/hwmon/hwmonM/power1_average'

3. llama.cpp divides -c among parallel slots.

-c 8192 -np 16 gives each request 512 tokens, not 8192. My agent kept getting truncated mid-JSON and I blamed the model for three hours. For agent workloads use few slots and a large context (-np 4 -c 32768); for throughput benchmarks the per-slot context doesn’t matter.

4. A liveness check is not a model check.

I benchmarked a 14B model and got numbers identical to the 7B to one decimal place. Cause: the previous llama-server still held port 8081, the new one failed to bind, and my readiness check (curl /v1/models returns 200) confirmed that something was answering — never what.

Fix — free the port, then verify the served model name:

pkill -f "llama-server.*--port 8081"
# wait for the port to actually free, then:
curl -s localhost:8081/v1/models | grep -q "$EXPECTED_MODEL" || exit 1

A load-generator note that isn’t a bug but changes results: firing N requests and waiting for all of them leaves the GPU partly idle during the straggler tail, and that tail grows with batch size — so it deflates power exactly at high concurrency, where the interesting behaviour is. Closed-loop load (keep exactly N in flight) plus recording the achieved concurrency fixes it and makes the run auditable.


Some measured results, if useful as a reference point. W7900, ROCm 7.2.1, llama.cpp HIP, Qwen2.5-Coder Q4_K_M:

7B 14B
Peak sustained throughput 696 tok/s 412 tok/s
Throughput gain, conc 1 → 16 6.75× 7.19×
Energy efficiency gain 8.70× 8.80×

Two things I did not expect: the relative efficiency gain from batching is the same on both models within 1%, and power draw is not monotonic — both peak at concurrency 4 and draw less at 16 while doing ~6.8× the work.

Code, raw results, logs and the rejected run: GitHub - Artkill24/costbench: Measured energy cost of on-prem LLM inference on AMD ROCm: 8.7× tokens-per-joule from batching, with a cost-governance gateway and agent built on the measurements. · GitHub (MIT)

Happy to answer questions on the measurement setup — and if anyone has bandwidth counters working on RDNA3, I’d like to hear from you. I could not confirm the memory-bound hypothesis: mclk sits pinned at 1124 MHz at every concurrency level, so clocks say nothing about traffic.

guys where do i like submit my project ??

1 Like

Here - you need to fork the repository and submit a pull request, rest of the submission requirements are also in the repo.