DALYBIGAS

DALYBIGAS / GAMMA-v2

Public

GAMMA-v2: An end-to-end co-design simulation framework integrating gem5 and MLIR, enabling LLM and operator-level workload modeling, configurable accelerator generation, and system-level evaluation for mapping and architecture exploration.

12
0
100% credibility
Found Mar 31, 2026 at 12 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
C++
AI Summary

GAMMA-v2 is a simulation framework for evaluating AI chip architectures on large language model workloads through compiler-driven planning and detailed performance reporting.

How It Works

1
🔍 Discover GAMMA-v2

You hear about a helpful tool that lets you test if your custom AI chip can handle big language models like chatting assistants.

2
📝 Describe your chip

You simply tell it about your chip's memory size, speed, and computing power using easy checklists.

3
🤖 Pick a model

Choose a popular AI model family like Llama or Qwen, and what task to run, like generating text.

4
⚡ Create smart plan

The tool automatically plans the best way to run your model on the chip, step by step.

5
▶️ Run the test

Press go and watch it simulate how your chip performs the AI work.

📊 See your results

Get clear reports on speed, efficiency, and tips to improve your design.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 12 to 12 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is GAMMA-v2?

GAMMA-v2 is a C++ framework for end-to-end co-design and evaluation of configurable AI accelerators targeting LLM workloads. You describe hardware capabilities like SRAM sizes, DMA bandwidth, and compute arrays in YAML, pair it with LLM specs for prefill/decode modes across models like LLaMA or Qwen, and it generates compiler plans, runtime artifacts, and gem5-style performance reports. It shifts from manual benchmarks to automated architecture exploration, enabling quick what-if analysis on mapping efficiency.

Why is it gaining traction?

Unlike generic simulators, GAMMA-v2 ties MLIR-based compile planning directly to hardware models, spitting out inspectable artifacts like tile/fuse scripts and estimated metrics before full sim runs. Developers dig the repeatable workflow for comparing configs—tweak array shapes or data types and see tokens/sec impacts fast—without hand-coding kernels. The LLM-aware presets for families like DeepSeek or Mixtral make it a sharp tool for modern v2 v1 gamma archer explorations over black-box alternatives.

Who should use this?

AI hardware architects prototyping accelerators for gamma v2 rc plane-scale inference, or researchers in co-design evaluating SRAM/DMA tradeoffs for Qwen2.5-Math workloads. It's for teams needing system-level insights on operator mapping without building full chips, especially those fluent in gem5 and MLIR.

Verdict

Worth forking for accelerator architecture exploration if you're in LLM hardware—early artifacts and reports accelerate iteration. With 12 stars and 1.0% credibility, it's immature (thin docs, benchmark-heavy), so expect setup tweaks; pair with v23ga batterij gamma tests for real validation.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.