[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-tilefuse-brings-quantized-llm-inference-to-amd-npus":10,"sections":40},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},3630,"tilefuse-brings-quantized-llm-inference-to-amd-npus","TileFuse Brings Quantized LLM Inference to AMD NPUs","A new kernel library lets AMD's XDNA2 NPU run AWQ-style quantized models natively, cutting prefill latency in half and energy use by more than 64.6%.","AMD's NPU can now run quantized large language models without reshaping the model to fit the hardware.\n\nResearchers have released TileFuse, a mixed-precision kernel library built specifically for AMD's XDNA2 neural processing unit. The library targets the GEMM and GEMV operations that dominate LLM inference and brings W4A16 and W8A16 quantization formats — the kind used by AWQ, a widely adopted weight quantization scheme — directly onto the chip. Rather than forcing developers to re-quantize models for a proprietary NPU format, TileFuse fuses unpacking, dequantization, and matrix computation into a single kernel flow. It also redesigns the data layout to support matrix dimensions up to 32K and makes fuller use of the XDNA2's 4x8 AIE array.\n\nThe performance numbers matter because on-device LLM inference has so far been mostly a CPU-and-iGPU story — NPUs have been marketed heavily but have delivered limited real-world gains for general workloads. TileFuse reports up to 2.0x lower prefill latency and more than 64.6% lower energy consumption compared to iGPU baselines in end-to-end tests on Ryzen AI laptops, which is the kind of result that shifts NPUs from marketing bullet point to genuinely useful compute.\n\nThe broader context: Qualcomm's Hexagon NPU and Apple's Neural Engine both benefit from tight vertical integration — the hardware, software stack, and quantization pipeline are co-designed. AMD is working from a more open position, and TileFuse is a bet that meeting developers where their models already are, rather than demanding format changes, is the faster path to adoption.","[\"ai\",\"hardware\",\"llm\",\"amd\"]","2026-07-03T04:00:00.000Z","2026-07-03T10:00:25.267Z","2026-07-03T10:00:28.074Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The dek claims '65%' energy savings but the source and body both state '>64.6%' — a statistic in the headline element that contradicts the body and cannot be rounded up to 65% without distortion; fix the dek to match the sourced figure.","resolved","hardware",[32,30,33,34],"ai","llm","amd",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.11357",0,{"sections":41},[42,46,51,56,61,65,70,75,80,85,90,94,99,104],{"name":43,"slug":32,"count":44,"latest_published_at":45},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":62,"slug":30,"count":63,"latest_published_at":64},"Hardware",122,"2026-07-14T19:46:26.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":91,"slug":92,"count":88,"latest_published_at":93},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":105,"slug":106,"count":107,"latest_published_at":108},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]