[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-gpgpu-graphics-chips-general-compute-engines-en":3,"article-related-gpgpu-graphics-chips-general-compute-engines-en":30,"series-industry-b998330a-abd3-48a8-9a49-d9e247832cd4":75},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"b998330a-abd3-48a8-9a49-d9e247832cd4","gpgpu-graphics-chips-general-compute-engines-en","GPGPU turned graphics chips into general compute engines","\u003Cp data-speakable=\"summary\">GPGPU uses graphics chips to run general-purpose workloads once reserved for CPUs.\u003C\u002Fp>\u003Cp>Graphics processors started as display hardware, but by the mid-2000s they were already doing serious compute work. \u003Ca href=\"\u002Ftag\u002Fnvidia\">Nvidia\u003C\u002Fa>’s \u003Ca href=\"https:\u002F\u002Fdeveloper.nvidia.com\u002Fcuda-zone\" target=\"_blank\" rel=\"noopener\">CUDA\u003C\u002Fa> arrived in 2006, \u003Ca href=\"https:\u002F\u002Fwww.khronos.org\u002Fopencl\u002F\" target=\"_blank\" rel=\"noopener\">OpenCL\u003C\u002Fa> became the open standard, and AMD’s \u003Ca href=\"https:\u002F\u002Frocm.docs.amd.com\u002Fen\u002Flatest\u002F\" target=\"_blank\" rel=\"noopener\">ROCm\u003C\u002Fa> later gave developers an open-source path on Radeon and Instinct hardware. The shift was not subtle: the article notes that many supercomputers now rely on GPUs, and that some optimized workloads have seen speedups of several hundred times over CPU-only code.\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>Milestone\u003C\u002Fth>\u003Cth>Year\u003C\u002Fth>\u003Cth>Why it mattered\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Conway's Game of Life on a blitter\u003C\u002Ftd>\u003Ctd>1987\u003C\u002Ftd>\u003Ctd>Early proof that non-graphics compute could run on graphics-style hardware\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>GPU compute becomes practical\u003C\u002Ftd>\u003Ctd>2001\u003C\u002Ftd>\u003Ctd>Programmable shaders and floating-point support made general compute easier\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>GPU linear algebra breakthroughs\u003C\u002Ftd>\u003Ctd>2003\u003C\u002Ftd>\u003Ctd>Research groups showed GPUs could beat CPUs on some matrix problems\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>CUDA launch\u003C\u002Ftd>\u003Ctd>2006\u003C\u002Ftd>\u003Ctd>Made GPU programming far less tied to graphics concepts\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>ROCm launch\u003C\u002Ftd>\u003Ctd>2016\u003C\u002Ftd>\u003Ctd>AMD’s open-source alternative to CUDA arrived\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>How GPUs escaped the graphics box\u003C\u002Fh2>\u003Cp>The core idea behind GPGPU is simple: a GPU has far more processing elements than a CPU, even if each core runs at a lower clock speed. That tradeoff makes sense for workloads that repeat the same operation across huge blocks of data, which is exactly what graphics already do.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784208782505-n41b.png\" alt=\"GPGPU turned graphics chips into general compute engines\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>Early GPGPU work had an awkward constraint. Developers had to translate math into graphics primitives and route it through APIs like \u003Ca href=\"https:\u002F\u002Fwww.khronos.org\u002Fopengl\u002F\" target=\"_blank\" rel=\"noopener\">OpenGL\u003C\u002Fa> and \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fwindows\u002Fwin32\u002Fdirect3d11\u002Fdirect3d-11-graphics\" target=\"_blank\" rel=\"noopener\">Direct3D\u003C\u002Fa>. That was powerful, but it was also clumsy. The article describes this translation as cumbersome, and that is the right word.\u003C\u002Fp>\u003Cp>Things got much easier when general-purpose programming models arrived. Tools such as Sh\u002FRapidMind, Brook, and Accelerator helped abstract away the graphics layer. Then \u003Ca href=\"\u002Ftag\u002Fcuda\">CUDA\u003C\u002Fa> made GPU code feel closer to ordinary high-performance computing, which is why it became the dominant proprietary framework.\u003C\u002Fp>\u003Cul>\u003Cli>GPUs excel when the same operation runs across many data points.\u003C\u002Fli>\u003Cli>CPU code often handles branching and serial logic better.\u003C\u002Fli>\u003Cli>Modern GPU workloads depend heavily on fast memory transfers in both directions.\u003C\u002Fli>\u003Cli>Scientific computing, image processing, and computer vision fit the model especially well.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>The software stack that made GPGPU usable\u003C\u002Fh2>\u003Cp>GPGPU did not take off because of hardware alone. It took off because the software stack finally caught up. OpenCL gave developers a vendor-neutral API, while CUDA gave Nvidia users a mature toolchain with strong documentation and a large ecosystem.\u003C\u002Fp>\u003Cp>That split still matters. OpenCL is designed to run across Intel, AMD, Nvidia, and ARM platforms. CUDA remains the best-known proprietary stack, and it is still the default choice in a lot of \u003Ca href=\"\u002Ftag\u002Fmachine-learning\">machine learning\u003C\u002Fa> and scientific code. AMD’s ROCm is the open-source answer, but the Wikipedia article notes that it still trails CUDA in consumer support.\u003C\u002Fp>\u003Cblockquote>\u003Cp>“CUDA is a general purpose parallel computing language and programming model that leverages the parallel compute engine in Nvidia GPUs to solve many complex computational problems in a more efficient way than on a CPU.” — \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fabout-nvidia\u002Fcuda-zone\u002F\" target=\"_blank\" rel=\"noopener\">Nvidia\u003C\u002Fa>\u003C\u002Fp>\u003C\u002Fblockquote>\u003Cp>That quote gets to the heart of the pitch. CUDA was not about drawing triangles faster. It was about making GPUs useful for problems that had nothing to do with rendering, from dense matrix math to simulation.\u003C\u002Fp>\u003Cp>There were also ecosystem-specific options. \u003Ca href=\"\u002Ftag\u002Fmicrosoft\">Microsoft\u003C\u002Fa> shipped \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fwindows\u002Fwin32\u002Fdirect3d11\u002Fdirectcompute-intro\" target=\"_blank\" rel=\"noopener\">DirectCompute\u003C\u002Fa> with Direct3D 11, \u003Ca href=\"\u002Ftag\u002Fapple\">Apple\u003C\u002Fa> pushed \u003Ca href=\"https:\u002F\u002Fdeveloper.apple.com\u002Fmetal\u002F\" target=\"_blank\" rel=\"noopener\">Metal\u003C\u002Fa>, and Android experimented with RenderScript before moving toward compute shaders and Vulkan Compute. The common thread is obvious: once mobile GPUs got powerful enough, compute moved onto phones too.\u003C\u002Fp>\u003Ch2>Why GPUs can outrun CPUs on some workloads\u003C\u002Fh2>\u003Cp>CPU and GPU design priorities are different. CPUs spend a lot of silicon on caches, branch prediction, and flexible control flow. GPUs spend more on execution resources that can do the same work across many threads at once. That is why a CPU is still better for messy logic, while a GPU can be much faster for tightly parallel tasks.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784208786600-rlz2.png\" alt=\"GPGPU turned graphics chips into general compute engines\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The article also points out an important memory detail. GPUs usually have smaller, faster local memory, while CPUs depend on larger pools of RAM and often much slower access paths. If you can move the active part of a dataset into GPU memory and keep the computation parallel, the speedup can be dramatic.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>CPU strength:\u003C\u002Fstrong> low-latency control flow, branching, and general-purpose task handling.\u003C\u002Fli>\u003Cli>\u003Cstrong>GPU strength:\u003C\u002Fstrong> massive data parallelism, matrix math, and image-like data.\u003C\u002Fli>\u003Cli>\u003Cstrong>Best-fit workloads:\u003C\u002Fstrong> genome mapping, protein analysis, molecular dynamics, and computer vision.\u003C\u002Fli>\u003Cli>\u003Cstrong>Reported gains:\u003C\u002Fstrong> some optimized pipelines have reached several hundred times the CPU baseline on one task.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That last point is where GPGPU gets interesting in practice. The wins are real, but they are highly workload-specific. A GPU will not magically speed up every program. It shines when the problem can be rewritten so that thousands of threads do similar work at once.\u003C\u002Fp>\u003Cp>This is why the most successful GPU software tends to live in scientific computing and data-heavy media pipelines. NVIDIA’s own software for genome analysis is a good example, and physics engines such as \u003Ca href=\"https:\u002F\u002Fwww.havok.com\u002F\" target=\"_blank\" rel=\"noopener\">Havok\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Ftechnologies\u002Fphysx\u002F\" target=\"_blank\" rel=\"noopener\">PhysX\u003C\u002Fa>, and \u003Ca href=\"https:\u002F\u002Fwww.chaos.com\u002Ffx\" target=\"_blank\" rel=\"noopener\">FX\u003C\u002Fa> show how the same compute model also fits games and simulation.\u003C\u002Fp>\u003Ch2>Where GPGPU goes from here\u003C\u002Fh2>\u003Cp>GPGPU is no longer a niche trick for graphics engineers. It is part of the default toolkit for high-performance computing, AI training, simulation, and media processing. The interesting question now is not whether GPUs can do general compute, but which workloads still belong on CPUs because the data movement cost is too high.\u003C\u002Fp>\u003Cp>That tradeoff will keep shaping software design. If your workload is parallel, data-heavy, and predictable, a GPU is often the first place to look. If it is branchy, latency-sensitive, or small enough that transfer overhead dominates, the CPU still wins. The real skill is knowing when to rewrite the problem for the GPU and when to leave it alone.\u003C\u002Fp>\u003Cp>For developers, that means GPGPU is less about hype and more about fit. The next wave of performance gains will probably come from better compilers, higher-level APIs like \u003Ca href=\"https:\u002F\u002Fwww.khronos.org\u002Fsycl\u002F\" target=\"_blank\" rel=\"noopener\">SYCL\u003C\u002Fa>, and tighter integration between CPUs and GPUs inside the same system. The open question is which toolchains will make that split easier without hiding the cost of moving data back and forth.\u003C\u002Fp>\u003Cp>If you are building for compute-heavy workloads today, the practical takeaway is simple: profile first, then decide whether your bottleneck is arithmetic, memory, or transfer overhead. That answer tells you whether the GPU is your best accelerator, or just an expensive detour.\u003C\u002Fp>","GPGPU lets GPUs run non-graphics workloads, and CUDA, OpenCL, and ROCm made that practical.","en.wikipedia.org","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FGeneral-purpose_computing_on_graphics_processing_units",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784208782505-n41b.png","industry","en","c02c3eac-077c-416c-9abe-f16ca7e87a98",[17,18,19,20,21],"GPGPU","CUDA","OpenCL","ROCm","GPU computing",[23,24,25],"GPUs became useful for general compute once shaders and floating-point support arrived.","CUDA made GPU programming much easier by hiding graphics-specific details.","OpenCL and ROCm gave developers vendor-neutral and open-source options.",0,"2026-07-16T13:32:36.87856+00:00","2026-07-16T13:32:36.869+00:00","254076b2-b9bb-4af1-a9c3-4791f375c74d",{"tags":31,"relatedLang":34,"relatedPosts":38},[32],{"name":18,"slug":33},"cuda",{"id":15,"slug":35,"title":36,"language":37},"gpgpu-graphics-chips-general-compute-engines-zh","GPGPU 讓 GPU 變成通用算力引擎","zh",[39,45,51,57,63,69],{"id":40,"slug":41,"title":42,"cover_image":43,"image_url":43,"created_at":44,"category":13},"bedbd2b7-1fa5-4fe1-8900-f001fa256b9a","anthropic-eerie-new-ad-sparks-backlash-en","Anthropic’s eerie new ad sparks backlash","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784228577377-8c7b.png","2026-07-16T19:02:33.597573+00:00",{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"5f316b7b-1e9d-4224-9d8b-eb3ef2001078","openai-hires-google-cloud-channel-chief-philip-larson-en","OpenAI hires Google Cloud channel chief Philip Larson","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784226769880-sa8x.png","2026-07-16T18:32:27.863303+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"0a7af24b-abb3-48da-ab48-95b3f3cba55e","openai-partner-network-clear-path-for-partners-en","OpenAI’s partner network now has a clear path","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784224965797-tmy2.png","2026-07-16T18:02:22.288434+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"99a5c6db-a711-44bb-bdf1-23d0a31b1f49","ai-firms-are-filling-london-offices-fast-en","AI Firms Are Filling London Offices Fast","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784205184233-ej7x.png","2026-07-16T12:32:31.50179+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"add9368d-a09a-4637-8565-a12c48bda72c","anthropics-contrarian-playbook-turns-enemies-into-growth-en","Anthropic’s contrarian playbook turns enemies into growth","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784203446979-a7k7.png","2026-07-16T12:03:37.464909+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"f3b6f173-7795-4817-aa84-7c350501ee38","warner-ai-agent-act-platform-control-en","Warner’s AI AGENT Act targets platform AI control","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784188980648-g2x4.png","2026-07-16T08:02:31.520043+00:00",[76,81,86,91,96,101,106,111,116,121],{"id":77,"slug":78,"title":79,"created_at":80},"d35a1bd9-e709-412e-a2df-392df1dc572a","ai-impact-2026-developments-market-en","AI's Impact in 2026: Key Developments and Market Shifts","2026-03-25T16:20:33.205823+00:00",{"id":82,"slug":83,"title":84,"created_at":85},"5ed27921-5fd6-492e-8c59-78393bf37710","trumps-ai-legislative-framework-en","Trump's AI Legislative Framework: What's Inside?","2026-03-25T16:22:20.005325+00:00",{"id":87,"slug":88,"title":89,"created_at":90},"e454a642-f03c-4794-b185-5f651aebbaca","nvidia-gtc-2026-key-highlights-innovations-en","NVIDIA GTC 2026: Key Highlights and Innovations","2026-03-25T16:22:47.882615+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"0ebb5b16-774a-4922-945d-5f2ce1df5a6d","claude-usage-diversifies-learning-curves-en","Claude Usage Diversifies, Learning Curves Emerge","2026-03-25T16:25:50.770376+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"69934e86-2fc5-4280-8223-7b917a48ace8","openclaw-ai-commoditization-concerns-en","OpenClaw's Rise Raises Concerns of AI Model Commoditization","2026-03-25T16:26:30.582047+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"b4b2575b-2ac8-46b2-b90e-ab1d7c060797","google-gemini-ai-rollout-2026-en","Google's Gemini AI Rollout Extended to 2026","2026-03-25T16:28:14.808842+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"6e18bc65-42ae-4ad0-b564-67d7f66b979e","meta-llama4-fabricated-results-scandal-en","Meta's Llama 4 Scandal: Fabricated AI Test Results Unveiled","2026-03-25T16:29:15.482836+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"bf888e9d-08be-4f47-996c-7b24b5ab3500","accenture-mistral-ai-deployment-en","Accenture and Mistral AI Team Up for AI Deployment","2026-03-25T16:31:01.894655+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"5382b536-fad2-49c6-ac85-9eb2bae49f35","mistral-ai-high-stakes-2026-en","Mistral AI: Facing High Stakes in 2026","2026-03-25T16:31:39.941974+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"9da3d2d6-b669-4971-ba1d-17fdb3548ed5","cursors-meteoric-rise-pressures-en","Cursor's Meteoric Rise Faces Industry Pressures","2026-03-25T16:32:21.899217+00:00"]