Developer Optimizes Gemma 4 31B Inference with Strix Halo int4 WMMA
11/17/2026 — 11/17, 19:19·1 sources (all)·1 reports (all)
Key takeaways
- A developer has implemented customized GEMM routines optimized for Strix Halo int4 WMMA instructions, enabling efficient local inference for…
Background & analysis
On November 17, 2026, at 18:08, a post titled "Gemma 4 31B inference implementation optimized for Strix Halo int4 cores" appeared on the Reddit community LocalLLaMA, describing an inference optimization effort by a developer. The developer noted that the 8060S in Strix Halo provides an int4 x int4 WMMA instruction running at twice the FP16 rate, measured at 104 TOPS, yet neither ROCm nor llama.cpp takes advantage of it. To address this gap, the developer wrote a custom GEMM tailored to that instruction, enabling Gemma 4 31B to run with int4 weights and int4 activations, while combining rotation, GPTQ, and selective mixed precision to preserve accuracy.
The benchmark was run on a Gemma 4 31B QAT Q4_0 model, with llama.cpp b11056 Vulkan using the -fa 1 option as the baseline. The setup used 2k-token prompts, greedy decoding, and a single stream, covering pp512 and pp2048. Results showed llama.cpp at 287 and 276 respectively, while the developer's implementation reached 679 and 724, yielding 2.43x the throughput of llama.cpp. The post presents the work as a new option for running large models locally on consumer hardware. At this point, this is the only available report, with no subsequent developments or independent sources confirming the benchmark figures or the claimed gains.
AI-generated from 1 reports · updated 11 hours ago
- Heat index
- 70
- All sources
- 1
- All reports
- 1
- First seen
- 13 hours ago
Reports on this story headlines open the original
Developers implemented a custom GEMM for Strix Halo's int4 WMMA instructions, enabling Gemma 4 31B to run efficiently with int4 weights and activations. By combining rotation, GPTQ, and selective mixed precision to maintain accuracy, the custom solution achieves 2.43x the throughput of llama.cpp on server benchmarks, offering a high-performance local inference approach.
Reddit · LocalLLaMAAI score 62
Other stories people are talking about
- Heat index 1331Nvidia board establishes national security and export control committee25 sources
- Heat index 1027California Sues OpenAI Over Agent Intrusion Into Hugging Face19 sources
- Heat index 1006NewGPT-6.2 Astra and Sol Now Available on ChatGPT, API and Perplexity11 sources
- Heat index 934SurgeJensen Huang says he would return as Nvidia CEO and agrees to testify on Capitol Hill11 sources
- Heat index 852Anthropic guides IPO pricing at top of range or higher, with trading set for November 1911 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

