Loading activity…
@semianalysis_
AI
SemiAnalysis says Nvidia engineers cited 5x from October 2025 to August 2026 with expert human optimization.
AI
SemiAnalysis cites its recent InferenceX work, saying Google engineers and outside companies demonstrated “great TCO.”
AI
SemiAnalysis says AgentX captures multi-turn sessions, long contexts, high prefix reuse and bursts of short-lived sub-agents.
AI
SGLang says it built the mode with NVIDIA and AMD; more low-latency kernels remain in progress.
AI
SemiAnalysis says GB300 FP4’s performance-per-dollar is 1.64 times worse than TPUv7’s for external customers across most of the curve.
Technology
SemiAnalysis says over 24 GW of data centers at 40 campuses are in advanced construction for 2027 delivery.
AI
SemiAnalysis says Inferact led the submission end to end using the TorchTPU vLLM stack.
AI
SemiAnalysis says an upstream vLLM optimization made the 2.8-trillion-parameter model 1.68 times faster on the same hardware.
Technology
SemiAnalysis reports NVIDIA engineers delivered decode optimizations for the 2.8-trillion-parameter Kimi K3 model, reaching 423 tokens per second per user.
AI
SemiAnalysis shares a quoted comparison calling GLM 5.3's cyber capabilities a “night and day” change from 5.2.
AI
SemiAnalysis says vLLM and Inferact achieved the result, four times better than the B300 aggregated configuration.
AI
SemiAnalysis_ jokes that, if GPT-7 is AGI, it could outdo Broadcom at 224G SerDes design, PD and supply-chain management.
AI
SemiAnalysis says Jefferies found a third of the US shopping and travel sites it tested blocked Muse.
AI
SemiAnalysis says hyperscalers aren't affected because of their greater buying power; neocloud executives fear Nvidia retaliation.
Technology
TSMC said Q3 revenue reached $46.72 billion as net income climbed 68.7% year over year.
AI
finkd says the model runs well on new Mac Studios, crediting work with Apple and John Ternus.
AI
SemiAnalysis cites the GLM-5 paper's claim of full-stack optimization across seven Chinese chip platforms.
AI
SemiAnalysis says the submission showed 1.67x better performance per dollar than GB300 NVL72 FP4 in its Agentic Inference Benchmark.
AI
SemiAnalysis says the copper backplane expands the scale-up domain to 72, allowing each GPU to hold only some experts.
AI
SemiAnalysis says sessions with 200k-plus context automatically compact after about 54 idle minutes.
