The AWS Trainium Frontier competition invites ML researchers, systems engineers, and AI-agent builders to co-design models and custom kernels on purpose-built Trainium2 silicon. Teams of one to four participants receive a complete training pipeline, NKI documentation, and profiling tools, then race to maximize model capability within a fixed 30-minute, single-chip training budget in Phase 1.
Phase 1 Requirements Teams must register to qualify and gain access to the private GitHub repository. The challenge requires each team to train a language model on a single Trainium2 chip within a strict 30-minute wall-clock training budget. All submissions must use the provided nanochat-derived training pipeline as a starting point; teams may modify any component, including model architecture, attention mechanisms, learning rate schedules, optimizer configurations, data mixing strategies, and custom NKI kernels, but training must complete within the fixed compute envelope.
Submissions must include: * A trained model checkpoint produced within the 30-minute training window on a single Trainium2 chip. * A reproducible training script that can be re-run by organizers to verify results without manual intervention. * A brief technical write-up (recommended) describing the approach, including any architectural modifications, custom NKI kernels, optimizer changes, or training recipe innovations.
Submissions are ranked on the public leaderboard by model quality as measured by the evaluation harness provided in the starter kit. The score is computed by the evaluate_bpb() function in prepare.py, which recomputes cross-entropy directly from the model's logits.
Advancement and Prizes The top 10 teams at the close of Phase 1 (September 30, 2026) advance to Phase 2, where they receive access to full Trn2 servers for scaled training. The competition features a $40,000 prize pool, and winners have the opportunity to present their findings alongside Annapurna Labs researchers at NeurIPS 2026 in Sydney.