latch
A compiler for training megakernels on NVIDIA H100. MIT-licensed research preview.
Get picks like this daily. Today's launch numbers, the top launches, AI and tech news, and a daily opportunity pick, straight to your inbox.
Visit website ↗ Source listing ↗
GITHUB · launched Oct 25, 2026 · 14 upvotes (top 81%)
- Horizontal
- Model & infra
- Developer
- Not AI
- Hobby / open-source project
Similar launches
See all 3867 similar launches →Could you build this?
No A compiler that fuses full training forward/backward/optimizer passes into H100 megakernels is novel GPU systems research requiring deep CUDA, kernel scheduling, and memory-hierarchy expertise.
What it would actually take: It requires a PyTorch graph frontend, an IR for training graphs, and a code generator emitting CUDA/PTX with manual shared-memory management, synchronization, and kernel fusion. Building and tuning it needs expert GPU systems engineers and H100 hardware access for benchmarking, and the work sits at the research frontier.
Competitors
Other products that read as similar to this one : 3867 launches clear the similarity bar, closest 8 shown.
Attention rank: #2996 of 3868 (itself plus its competitors, highest first, normalized so YC and Product Hunt are compared fairly).
Launched 361 days after the earliest competitor.
- solidcore · github · 2026-10-12 · 73 upvotes · similarity 0.71
- OpenDLSS-NR · github · 2026-09-20 · 371 upvotes · similarity 0.67
- tpu-megakernels · github · 2026-09-23 · 119 upvotes · similarity 0.67
- WarpLite · github · 2026-09-23 · 224 upvotes · similarity 0.65
- nvidia · github · 2026-10-01 · 14 upvotes · similarity 0.64
- TIRx-harness · github · 2026-09-29 · 102 upvotes · similarity 0.64
- KohakUnlock · github · 2026-10-25 · 366 upvotes · similarity 0.64
- rune · github · 2026-10-22 · 78 upvotes · similarity 0.64
Other launches for this product
- No other launches for this product.