[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-apex4-makes-int4-inference-work-on-the-right-hardware":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},2970,"apex4-makes-int4-inference-work-on-the-right-hardware","APEX4 Makes INT4 Inference Work — on the Right Hardware","A new quantization system hits 2.09x end-to-end speedup on some GPUs, but the gains depend entirely on your hardware's internal compute balance.","A research team has released APEX4, a pure W4A4 quantization system that finally makes full INT4 inference practical — but only on certain GPU architectures.\n\nThe core problem: running both weights and activations at 4-bit precision (W4A4) should unlock fast INT4 Tensor Core math, but dequantization work on CUDA Cores has historically eaten those gains. APEX4 traces this bottleneck to a single hardware metric — the ratio of Tensor Core to CUDA Core throughput, which they call ρ. On GPUs where ρ is low (RTX 3090 at ρ=16, A40 at ρ=16), the W4A4 kernel delivers 2.0–2.5x speedup at the kernel level. On the A100, where ρ hits 64, the same kernel *degrades* performance to 0.43–0.47x of baseline — meaning it runs more than twice as slow. That is not a rounding error; that is a fundamental mismatch between the workload and the chip.\n\nThe \"why it matters\" is what APEX4 does with that finding. Rather than accepting the A100's penalty, the system adds a mixed-granularity mode that recovers A100 performance to a 1.20–1.40x end-to-end speedup in vLLM. On friendlier hardware, end-to-end gains in unmodified vLLM reach 1.66x on the L40S (ρ=8), 1.78x on the RTX 3090, and 2.09x on the A40. Accuracy holds up: perplexity on LLaMA-2-70B stays within 0.63 of FP16, and zero-shot accuracy beats the competing W4Ax Atom-g128 baseline by 4.0–4.4%.\n\nThe honest caveat is that the biggest speedups land on prosumer and workstation GPUs — the RTX 3090 and A40 — not the data-center A100s that run most production inference today. Teams betting their stack on H100s or A100s should read the ρ analysis carefully before assuming these numbers apply to them.","[\"ai\",\"inference\",\"quantization\",\"hardware\"]","2026-06-30T04:00:00.000Z","2026-06-30T16:09:37.382Z","2026-06-30T16:09:40.827Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The article states the RTX 3090 delivers 'up to 1.78x end-to-end speedup' and correctly attributes 2.09x to the A40, but omits the L40S result (1.66x) entirely while implying completeness — more critically, the kernel-level speedup figures from the source (2.0–2.5x on RTX 3090, 0.43–0.47x on A100) are distinct from the end-to-end vLLM figures, and the article blurs this distinction in a way that misrepresents what APEX4 actually achieves at the kernel level versus in deployment; revise to cleanl","resolved","ai",[30,32,33,34],"inference","quantization","hardware",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.08761",0,{"sections":41},[42,46,51,56,61,65,70,75,80,85,90,94,99,104],{"name":43,"slug":30,"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":34,"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"]