--- schema: 1 type: video title: 'Taming the AI Hardware Cambrian Explosion — Abdul Dakkak, Modular' description: >- Compute is booked out for years and new chips keep arriving. The hard part is software that runs well on all of them. Abdul Dakkak, chief scientist at Modular, explains why today's AI software stack is a mess, with vLLM, SGLang, TensorRT-LLM and llama.cpp each patching around different hardware, and how Modular rebuilt the stack from the ground up. Mojo is a Pythonic systems language for CPUs, GPUs and accelerators that aims to match CUDA's performance; Modular has open-sourced its kernels, including matmul and attention kernels he says beat the vendor libraries and FlashAttention 4, with Mojo 1.0 and an open-source Mojo coming. MAX is an inference framework that rethinks PyTorch-era assumptions, supports thousands of model architectures behind OpenAI-compatible endpoints, and runs across GPUs, nodes and even mixed hardware. Modular Cloud ties it together as a tunable glass box. He shares benchmarks (Kimi K2.5, FLUX.2, AMD MI355 and Gemma 4), sub-second image generation, and how to try it with pip install modular. In this talk: • Why the explosion of AI hardware broke the software stack • Mojo: Pythonic, portable, and as fast as CUDA • MAX: inference across GPUs, nodes and mixed hardware • Benchmarks vs vLLM and torch.compile, and sub-second image generation SPEAKER Abdul Dakkak, Chief Scientist, Modular LinkedIn: https://www.linkedin.com/in/adakkak GitHub: https://github.com/abduld LINKS Modular: https://www.modular.com Modular docs: https://docs.modular.com Modular (GitHub): https://github.com/modular/modular CHAPTERS 0:00 Intro 0:23 Compute crunch, hardware explosion 1:03 A messy software stack 1:48 Rebuilding the stack 3:53 Mojo 4:28 Mojo vs CUDA, Triton and OpenCL 5:23 State-of-the-art kernels 6:13 MAX 7:43 Thousands of models supported 8:58 Multi-GPU, multi-node, mixed hardware 9:33 Modular Cloud 10:58 Why it matters 11:27 Benchmarks: Kimi K2.5 12:07 FLUX.2 12:42 AMD MI355 13:17 Independent results 14:27 Sub-second image generation 14:47 Gemma 4 15:32 Get involved Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #Mojo #AIInference #AIEngineer canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/taming-the-ai-hardware-cambrian-explosion-abdul-dak-f006f8706720 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/taming-the-ai-hardware-cambrian-explosion-abdul-dak-f006f8706720.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:05.201Z' language: null children: [] videoId: 7BcQdtRA3Tc durationSeconds: 1028 publishedAt: '2026-10-11T21:00:22.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 23 endSeconds: 34 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=23' heldReason: rights - ordinal: 1 text: null startSeconds: 81 endSeconds: 116 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=81' heldReason: rights - ordinal: 2 text: null startSeconds: 126 endSeconds: 239 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=126' heldReason: rights - ordinal: 3 text: null startSeconds: 126 endSeconds: 239 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=126' heldReason: rights - ordinal: 4 text: null startSeconds: 283 endSeconds: 328 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=283' heldReason: rights - ordinal: 5 text: null startSeconds: 339 endSeconds: 377 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=339' heldReason: rights - ordinal: 6 text: null startSeconds: 387 endSeconds: 475 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=387' heldReason: rights - ordinal: 7 text: null startSeconds: 486 endSeconds: 546 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=486' heldReason: rights - ordinal: 8 text: null startSeconds: 546 endSeconds: 556 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=546' heldReason: rights - ordinal: 9 text: null startSeconds: 590 endSeconds: 667 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=590' heldReason: rights - ordinal: 10 text: null startSeconds: 590 endSeconds: 667 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=590' heldReason: rights - ordinal: 11 text: null startSeconds: 699 endSeconds: 736 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=699' heldReason: rights - ordinal: 12 text: null startSeconds: 736 endSeconds: 748 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=736' heldReason: rights - ordinal: 13 text: null startSeconds: 748 endSeconds: 769 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=748' heldReason: rights - ordinal: 14 text: null startSeconds: 780 endSeconds: 802 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=780' heldReason: rights - ordinal: 15 text: null startSeconds: 815 endSeconds: 874 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=815' heldReason: rights - ordinal: 16 text: null startSeconds: 815 endSeconds: 874 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=815' heldReason: rights - ordinal: 17 text: null startSeconds: 887 endSeconds: 899 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=887' heldReason: rights - ordinal: 18 text: null startSeconds: 934 endSeconds: 945 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=934' heldReason: rights - ordinal: 19 text: null startSeconds: 945 endSeconds: 1028 watch: 'https://liquidgold.cc/watch/7BcQdtRA3Tc?t=945' heldReason: rights --- # Taming the AI Hardware Cambrian Explosion — Abdul Dakkak, Modular Compute is booked out for years and new chips keep arriving. The hard part is software that runs well on all of them. Abdul Dakkak, chief scientist at Modular, explains why today's AI software stack is a mess, with vLLM, SGLang, TensorRT-LLM and llama.cpp each patching around different hardware, and how Modular rebuilt the stack from the ground up. Mojo is a Pythonic systems language for CPUs, GPUs and accelerators that aims to match CUDA's performance; Modular has open-sourced its kernels, including matmul and attention kernels he says beat the vendor libraries and FlashAttention 4, with Mojo 1.0 and an open-source Mojo coming. MAX is an inference framework that rethinks PyTorch-era assumptions, supports thousands of model architectures behind OpenAI-compatible endpoints, and runs across GPUs, nodes and even mixed hardware. Modular Cloud ties it together as a tunable glass box. He shares benchmarks (Kimi K2.5, FLUX.2, AMD MI355 and Gemma 4), sub-second image generation, and how to try it with pip install modular. In this talk: • Why the explosion of AI hardware broke the software stack • Mojo: Pythonic, portable, and as fast as CUDA • MAX: inference across GPUs, nodes and mixed hardware • Benchmarks vs vLLM and torch.compile, and sub-second image generation SPEAKER Abdul Dakkak, Chief Scientist, Modular LinkedIn: https://www.linkedin.com/in/adakkak GitHub: https://github.com/abduld LINKS Modular: https://www.modular.com Modular docs: https://docs.modular.com Modular (GitHub): https://github.com/modular/modular CHAPTERS 0:00 Intro 0:23 Compute crunch, hardware explosion 1:03 A messy software stack 1:48 Rebuilding the stack 3:53 Mojo 4:28 Mojo vs CUDA, Triton and OpenCL 5:23 State-of-the-art kernels 6:13 MAX 7:43 Thousands of models supported 8:58 Multi-GPU, multi-node, mixed hardware 9:33 Modular Cloud 10:58 Why it matters 11:27 Benchmarks: Kimi K2.5 12:07 FLUX.2 12:42 AMD MI355 13:17 Independent results 14:27 Sub-second image generation 14:47 Gemma 4 15:32 Get involved Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #Mojo #AIInference #AIEngineer Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/taming-the-ai-hardware-cambrian-explosion-abdul-dak-f006f8706720 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.