{"slug": "rocm", "name": "ROCm", "category": "dev-libs", "category_name": "Libraries and frameworks", "description": "", "recommenders": 3, "commenters": 8, "warners": 4, "score": -0.608, "rising": 0.02, "first": "2026-04-19", "latest": "2026-09-13", "why": "ROCm runs most things just fine without any ceremony or difficulty.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 4, "neutral": 2, "compare": 0, "warn": 4}, "spark": [0, 0, 0, 0, 0, 1, 0, 0, 2, 1, 0, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "ROCm runs most things just fine without any ceremony or difficulty.", "icon": true, "domain": null, "github_repo": "github.com/rocm/rocm", "official_url": "https://www.amd.com/en/products/software/rocm.html", "pricing": "", "aliases": ["ROCm"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 788644960, "item_id": 49290310, "url": "https://news.ycombinator.com/item?id=49290310", "by": "SwellJoe", "date": "2026-08-13", "stance": "recommend", "quote": "ROCm runs most things just fine without any ceremony or difficulty.", "excerpt": "\"used to\" in your sentence already answers your curiosity. ROCm runs most things just fine without any ceremony or difficulty. I have a couple of the same cards as covered in the article, bought before they got expensive (I'd recommend current Radeon AI Pro 9700 cards over the V620 now, though, as they have increased in price since I bought mine), and they're at the long end of the supported chart for ROCm...nearly E", "context": "runs AMD GPUs and Strix Halo at home", "use_case": "local AI on AMD GPUs", "switched_from": "", "story": "We Have AI at Home Chapter 1: A Box of Scraps"}, {"id": 785274226, "item_id": 49079639, "url": "https://news.ycombinator.com/item?id=49079639", "by": "jauntywundrkind", "date": "2026-07-27", "stance": "recommend", "quote": "Of course, it's not done at all, and that looks bad, and I'm so scared to share it, in that light, but I just want to applaud so much", "excerpt": "I mean, Nvidia says CUDA Tiles says CUDA Tiles MLIR is open source. https://news.ycombinator.com/item?id=46330732 Afaik, AMD, Google, and everyone else is still using some combo of mostly XLA, with StableHLO, IREE, etc? https://openxla.org/ I was going to lightly blow off more steam about AMD's TheRock all-in-one monorepo-ish thing for their graphics drivers being basically unbuildable on anything but a single distro", "context": "", "use_case": "GPU compute", "switched_from": "", "story": "Can AMD Break the CUDA Moat? AMD Advancing AI 2026"}, {"id": 784341457, "item_id": 49021341, "url": "https://news.ycombinator.com/item?id=49021341", "by": "data-ottawa", "date": "2026-07-23", "stance": "recommend", "quote": "I'm able to use ROCm with MTP just fine (Qwen and Gemma).", "excerpt": "I'm able to use ROCm with MTP just fine (Qwen and Gemma). I use Lemonade and have only the ROCm installed through it (ETA: their llama.cpp version is a fork, not sure what is different vs the main trunk).", "context": "", "use_case": "GPU backend for LLM inference", "switched_from": "", "story": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory"}, {"id": 765215280, "item_id": 47825955, "url": "https://news.ycombinator.com/item?id=47825955", "by": "SwellJoe", "date": "2026-04-19", "stance": "recommend", "quote": "Now that ROCm is reliable, I'm using ROCm.", "excerpt": "It can be tricky to get/keep ROCm working, but around 7.2 it became reliable and as fast as or faster than ROCm 6.4. And, I think the first response time of ROCm is pretty consistently faster than Vulkan, even if Vulkan has a slightly higher token rate. Though I don't see that big of a different on token rates, either. Honestly, though, I haven't done enough real testing to know for sure. The benchmarks Donato Capite", "context": "Strix Halo user", "use_case": "running local LLMs on Strix Halo", "switched_from": "Vulkan", "story": "My first impressions on the ROCm and Strix Halo"}]}], "caveats": [{"label": "", "summary": "", "quotes": [{"id": 793858097, "item_id": 49616131, "url": "https://news.ycombinator.com/item?id=49616131", "by": "civvv", "date": "2026-09-08", "stance": "warn", "quote": "The official llama.cpp ROCm build detected the card correctly, but kept crashing with 0xC0000005 while loading the model", "excerpt": "I got Qwen3.8-27B running on my RX 9070 XT by setting up AMD\u2019s ROCm 7.14/TheRock runtime for gfx1201 and adding the required HIP/LLVM paths to Windows. The official llama.cpp ROCm build detected the card correctly, but kept crashing with 0xC0000005 while loading the model, so I switched to the Windows Vulkan build of llama.cpp instead, which worked reliably. I\u2019m running Qwen3.8-27B-Unleashed UD-Q3_K_XL, which is a ~1", "context": "", "use_case": "running llama.cpp on AMD GPU", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 784554784, "item_id": 49034674, "url": "https://news.ycombinator.com/item?id=49034674", "by": "stevefan1999", "date": "2026-07-24", "stance": "warn", "quote": "ROCm, to this day, still sucks.", "excerpt": "but you have to rewrite all your software to ROCm. And ROCm, to this day, still sucks. ZLUDA basically crash on high memory demand, and only accounted for ~70% of CUDA API coverage (and it is still buggy). But hey, at least it does run on Rust-CUDA. I'm one of the few who ported it and fixed a few bugs on ZLUDA. I used it to run a simple SHA256 kernel and it ran sure, but I gave it up because of those fundamental pro", "context": "ported Rust-CUDA and fixed bugs on ZLUDA", "use_case": "AMD GPU compute", "switched_from": "", "story": "AMD's Instinct MI455X: Aiming for the Sun"}, {"id": 784219025, "item_id": 49013689, "url": "https://news.ycombinator.com/item?id=49013689", "by": "icedchai", "date": "2026-07-22", "stance": "warn", "quote": "It definitely wasn't an \"out of the box\" experience, fiddling around with kernel versions and ROCm installs.", "excerpt": "I have a AMD Strix Halo box I use for similar dabbling. It definitely wasn't an \"out of the box\" experience, fiddling around with kernel versions and ROCm installs. These days I mostly wind up using the Vulkan build of llama.cpp for inference.", "context": "", "use_case": "GPU compute on Strix Halo", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}, {"id": 765956688, "item_id": 47872293, "url": "https://news.ycombinator.com/item?id=47872293", "by": "theoli", "date": "2026-04-22", "stance": "warn", "quote": "This is about the GPU clock remaining at max frequency, but that can drive the fan speed to increase.", "excerpt": "There was this ROCm bug I was watching for awhile: https://github.com/ROCm/ROCm/issues/5706 - This is about the GPU clock remaining at max frequency, but that can drive the fan speed to increase. It doesn't happen with Vulkan backends, so that is what I have been using for my two dual R9700 hosts. EDIT: The bug is closed but there were mentions of the issue still occurring after closure, so who knows if it is really ", "context": "runs two dual R9700 hosts", "use_case": "running LLMs on AMD GPUs", "switched_from": "ROCm", "story": "Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model"}]}], "who": [{"by": "SwellJoe", "context": "runs AMD GPUs and Strix Halo at home", "url": "https://news.ycombinator.com/item?id=49290310", "date": "2026-08-13"}], "tags": [], "use_cases": ["amd gpu compute", "running llama.cpp on amd gpu", "local ai on amd gpus", "gpu compute", "gpu backend for llm inference", "gpu compute on strix halo"], "compared": [{"slug": "llama-cpp", "name": "llama.cpp", "co": 7, "switched_from": 0, "switched_to": 1}, {"slug": "vulkan", "name": "Vulkan", "co": 5, "switched_from": 1, "switched_to": 0}, {"slug": "cuda", "name": "CUDA", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "llama-swap", "name": "llama-swap", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "qwen-3-6", "name": "Qwen 3.6", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "zluda", "name": "ZLUDA", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "gemma", "name": "Gemma", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "lemonade", "name": "Lemonade", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=49684356", "title": "CUDA for AMD on Windows", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49611128", "title": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49288293", "title": "We Have AI at Home Chapter 1: A Box of Scraps", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49078762", "title": "Can AMD Break the CUDA Moat? AMD Advancing AI 2026", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49032072", "title": "AMD's Instinct MI455X: Aiming for the Sun", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49019694", "title": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {}, "2026-04": {"recommend": 1, "warn": 1}, "2026-05": {}, "2026-06": {"neutral": 1}, "2026-07": {"warn": 2, "recommend": 2}, "2026-08": {"recommend": 1}, "2026-09": {"warn": 1, "neutral": 1}, "2026-10": {}}, "quotes": [{"id": 794949714, "item_id": 49684357, "url": "https://news.ycombinator.com/item?id=49684357", "by": "chiassedu80", "date": "2026-09-13", "stance": "neutral", "quote": "using ZLUDA + ROCm/HIP", "excerpt": "CUDA for AMD on Windows I\u2019ve been working on a Windows setup that lets CUDA-targeted applications run on AMD GPUs using ZLUDA + ROCm/HIP. Repo: https://github.com/Speedstu/CUDA-for-AMD-Windows So far, it has only been tested on my RX 9060 XT (gfx1200), where I\u2019ve used it with CUDA-enabled LibTorch workloads, including long ai training and use. I also added a GPU scanner / auto-detection system that detects: AMD GPU m", "context": "", "use_case": "AMD GPU compute", "switched_from": "", "story": "CUDA for AMD on Windows"}, {"id": 793858097, "item_id": 49616131, "url": "https://news.ycombinator.com/item?id=49616131", "by": "civvv", "date": "2026-09-08", "stance": "warn", "quote": "The official llama.cpp ROCm build detected the card correctly, but kept crashing with 0xC0000005 while loading the model", "excerpt": "I got Qwen3.8-27B running on my RX 9070 XT by setting up AMD\u2019s ROCm 7.14/TheRock runtime for gfx1201 and adding the required HIP/LLVM paths to Windows. The official llama.cpp ROCm build detected the card correctly, but kept crashing with 0xC0000005 while loading the model, so I switched to the Windows Vulkan build of llama.cpp instead, which worked reliably. I\u2019m running Qwen3.8-27B-Unleashed UD-Q3_K_XL, which is a ~1", "context": "", "use_case": "running llama.cpp on AMD GPU", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 788644960, "item_id": 49290310, "url": "https://news.ycombinator.com/item?id=49290310", "by": "SwellJoe", "date": "2026-08-13", "stance": "recommend", "quote": "ROCm runs most things just fine without any ceremony or difficulty.", "excerpt": "\"used to\" in your sentence already answers your curiosity. ROCm runs most things just fine without any ceremony or difficulty. I have a couple of the same cards as covered in the article, bought before they got expensive (I'd recommend current Radeon AI Pro 9700 cards over the V620 now, though, as they have increased in price since I bought mine), and they're at the long end of the supported chart for ROCm...nearly E", "context": "runs AMD GPUs and Strix Halo at home", "use_case": "local AI on AMD GPUs", "switched_from": "", "story": "We Have AI at Home Chapter 1: A Box of Scraps"}, {"id": 785274226, "item_id": 49079639, "url": "https://news.ycombinator.com/item?id=49079639", "by": "jauntywundrkind", "date": "2026-07-27", "stance": "recommend", "quote": "Of course, it's not done at all, and that looks bad, and I'm so scared to share it, in that light, but I just want to applaud so much", "excerpt": "I mean, Nvidia says CUDA Tiles says CUDA Tiles MLIR is open source. https://news.ycombinator.com/item?id=46330732 Afaik, AMD, Google, and everyone else is still using some combo of mostly XLA, with StableHLO, IREE, etc? https://openxla.org/ I was going to lightly blow off more steam about AMD's TheRock all-in-one monorepo-ish thing for their graphics drivers being basically unbuildable on anything but a single distro", "context": "", "use_case": "GPU compute", "switched_from": "", "story": "Can AMD Break the CUDA Moat? AMD Advancing AI 2026"}, {"id": 784554784, "item_id": 49034674, "url": "https://news.ycombinator.com/item?id=49034674", "by": "stevefan1999", "date": "2026-07-24", "stance": "warn", "quote": "ROCm, to this day, still sucks.", "excerpt": "but you have to rewrite all your software to ROCm. And ROCm, to this day, still sucks. ZLUDA basically crash on high memory demand, and only accounted for ~70% of CUDA API coverage (and it is still buggy). But hey, at least it does run on Rust-CUDA. I'm one of the few who ported it and fixed a few bugs on ZLUDA. I used it to run a simple SHA256 kernel and it ran sure, but I gave it up because of those fundamental pro", "context": "ported Rust-CUDA and fixed bugs on ZLUDA", "use_case": "AMD GPU compute", "switched_from": "", "story": "AMD's Instinct MI455X: Aiming for the Sun"}, {"id": 784341457, "item_id": 49021341, "url": "https://news.ycombinator.com/item?id=49021341", "by": "data-ottawa", "date": "2026-07-23", "stance": "recommend", "quote": "I'm able to use ROCm with MTP just fine (Qwen and Gemma).", "excerpt": "I'm able to use ROCm with MTP just fine (Qwen and Gemma). I use Lemonade and have only the ROCm installed through it (ETA: their llama.cpp version is a fork, not sure what is different vs the main trunk).", "context": "", "use_case": "GPU backend for LLM inference", "switched_from": "", "story": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory"}, {"id": 784219025, "item_id": 49013689, "url": "https://news.ycombinator.com/item?id=49013689", "by": "icedchai", "date": "2026-07-22", "stance": "warn", "quote": "It definitely wasn't an \"out of the box\" experience, fiddling around with kernel versions and ROCm installs.", "excerpt": "I have a AMD Strix Halo box I use for similar dabbling. It definitely wasn't an \"out of the box\" experience, fiddling around with kernel versions and ROCm installs. These days I mostly wind up using the Vulkan build of llama.cpp for inference.", "context": "", "use_case": "GPU compute on Strix Halo", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}, {"id": 779753283, "item_id": 48734580, "url": "https://news.ycombinator.com/item?id=48734580", "by": "data-ottawa", "date": "2026-06-30", "stance": "neutral", "quote": "ROCm is finally usable.", "excerpt": "I have the Framework Desktop with 395+ 128gb RAM Today I am pretty happy with it. LLMs are finally good enough (fast enough with MTP+MoE, but also just much better in capability) that I can fit local ones into real tasks, and I've used image generation with invokeAI to do some genuinely useful things like rendering concepts for a renovation. I mostly use lemonade-server and invokeAI for my workloads, previously I use", "context": "", "use_case": "GPU compute on AMD", "switched_from": "", "story": "DGX Spark vs. Mac Studio and Halo"}, {"id": 765956688, "item_id": 47872293, "url": "https://news.ycombinator.com/item?id=47872293", "by": "theoli", "date": "2026-04-22", "stance": "warn", "quote": "This is about the GPU clock remaining at max frequency, but that can drive the fan speed to increase.", "excerpt": "There was this ROCm bug I was watching for awhile: https://github.com/ROCm/ROCm/issues/5706 - This is about the GPU clock remaining at max frequency, but that can drive the fan speed to increase. It doesn't happen with Vulkan backends, so that is what I have been using for my two dual R9700 hosts. EDIT: The bug is closed but there were mentions of the issue still occurring after closure, so who knows if it is really ", "context": "runs two dual R9700 hosts", "use_case": "running LLMs on AMD GPUs", "switched_from": "ROCm", "story": "Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model"}, {"id": 765215280, "item_id": 47825955, "url": "https://news.ycombinator.com/item?id=47825955", "by": "SwellJoe", "date": "2026-04-19", "stance": "recommend", "quote": "Now that ROCm is reliable, I'm using ROCm.", "excerpt": "It can be tricky to get/keep ROCm working, but around 7.2 it became reliable and as fast as or faster than ROCm 6.4. And, I think the first response time of ROCm is pretty consistently faster than Vulkan, even if Vulkan has a slightly higher token rate. Though I don't see that big of a different on token rates, either. Honestly, though, I haven't done enough real testing to know for sure. The benchmarks Donato Capite", "context": "Strix Halo user", "use_case": "running local LLMs on Strix Halo", "switched_from": "Vulkan", "story": "My first impressions on the ROCm and Strix Halo"}]}