https://lazytensor.com/news/hardware-coherent-900-gb-s-nvlink-c2c-zero-copy-kv-cache-paging-cuts-128k-context-reasoning-ttft-by-64-percent-01a171de993d7203b7950412bf61a39dLazy Tensoren2026-10-25T07:03:06.000ZHardware-Coherent 900-GB/s NVLink-C2C Zero-Copy KV-Cache Paging Cuts 128K-Context Reasoning TTFT by 64 Percenthttps://lazytensor.com/news/co-packaged-1-6-tbps-silicon-photonics-switches-cut-moe-all-to-all-dispatch-tail-latency-by-58-percent-across-1-024-gpu-clusters-01a170ec5c66708391ac2387eebab6edLazy Tensoren2026-10-25T07:02:40.000ZCo-Packaged 1.6-Tbps Silicon Photonics Switches Cut MoE All-to-All Dispatch Tail Latency by 58 Percent Across 1,024-GPU Clustershttps://lazytensor.com/news/co-packaged-silicon-photonics-interposers-deliver-16-tbps-bi-directional-escape-bandwidth-at-1-9-picojoules-per-bit-across-multi-rack-clusters-01a16fd96fc27942a4045a684030f121Lazy Tensoren2026-10-25T07:02:15.000ZCo-Packaged Silicon Photonics Interposers Deliver 16 Tbps Bi-Directional Escape Bandwidth at 1.9 Picojoules per Bit Across Multi-Rack Clustershttps://lazytensor.com/news/ocp-microscaling-mxfp4-tensor-cores-achieve-99-6-percent-bf16-convergence-parity-on-405b-pretraining-via-stochastic-rounding-hardware-01a170c603887b01820c64154ab6603aLazy Tensoren2026-10-25T07:00:53.000ZOCP Microscaling MXFP4 Tensor Cores Achieve 99.6 Percent BF16 Convergence Parity on 405B Pretraining via Stochastic Rounding Hardwarehttps://lazytensor.com/news/hardware-fused-fast-hadamard-transforms-enable-lossless-3-bit-kv-cache-compression-at-128k-context-without-perplexity-drift-01a16e96b61f74f28d86a54480e72238Lazy Tensoren2026-10-24T15:45:44.000ZHardware-Fused Fast Hadamard Transforms Enable Lossless 3-Bit KV-Cache Compression at 128K Context Without Perplexity Drifthttps://lazytensor.com/news/the-end-of-the-pitch-deck-01a16e4645037613b3566790389ed371Lazy Tensoren2026-10-24T14:20:01.000ZThe End of the Pitch Deck?https://lazytensor.com/news/two-phase-micro-channel-copper-coldplates-sustain-1-500w-per-accelerator-socket-at-1-04-facility-pue-01a16e23173271929aa93ce753d2cdc8Lazy Tensoren2026-10-24T13:39:22.000ZTwo-Phase Micro-Channel Copper Coldplates Sustain 1,500W per Accelerator Socket at 1.04 Facility PUEhttps://lazytensor.com/news/optical-cxl-3-1-memory-pooling-sustains-188-ns-load-to-use-latency-across-64-tb-rack-kv-cache-tiers-01a16bb792827c32b45499b612e0212dLazy Tensoren2026-10-24T07:02:37.000ZOptical CXL 3.1 Memory Pooling Sustains 188 ns Load-to-Use Latency Across 64-TB Rack KV-Cache Tiershttps://lazytensor.com/news/dedicated-on-package-sram-draft-coprocessors-push-speculative-decoding-to-4-1-accepted-tokens-per-step-01a16be4be5c7b53ab72a00aef3fc4b1Lazy Tensoren2026-10-24T07:00:38.000ZDedicated On-Package SRAM Draft Coprocessors Push Speculative Decoding to 4.1 Accepted Tokens per Stephttps://lazytensor.com/news/block-scaled-mxfp4-with-stochastic-hadamard-rounding-matches-bf16-convergence-on-70b-pre-training-01a16a66e71678d382d893ff9d8fde46Lazy Tensoren2026-10-23T20:14:59.000ZBlock-Scaled MXFP4 With Stochastic Hadamard Rounding Matches BF16 Convergence on 70B Pre-Traininghttps://lazytensor.com/news/co-packaged-silicon-photonics-interconnects-cut-multi-rack-all-reduce-latency-by-62-at-1-6-tbps-per-port-01a1699755ac7aa19a568801b310edd6Lazy Tensoren2026-10-23T16:28:18.000ZCo-Packaged Silicon Photonics Interconnects Cut Multi-Rack All-Reduce Latency by 62% at 1.6 Tbps per Porthttps://lazytensor.com/news/neuro-symbolic-provers-achieve-84-automated-verification-on-gpu-triton-kernel-safety-invariants-01a1693ce60176d0a75740d2b4c40e1eLazy Tensoren2026-10-23T14:49:51.000ZNeuro-symbolic provers verify 84% of Triton kernel safety invariantshttps://lazytensor.com/news/paul-graham-and-the-changing-economics-of-building-a-startup-01a1685d930b7b638582fa238db3f356Lazy Tensoren2026-10-23T10:45:32.000ZPaul Graham and the Changing Economics of Building a Startup