https://ure.us/pillars/compute--workload-performance/Recent content in Compute & Workload Performance on UREHugo -- 0.164.0en-usFri, 19 Jun 2026 00:00:00 +0000https://ure.us/articles/frontier-ai-is-a-system-not-a-model/Fri, 19 Jun 2026 00:00:00 +0000https://ure.us/articles/frontier-ai-is-a-system-not-a-model/Frontier AI is a system, not a model you download. The cheapest layer you can own is GPU vector search. What NVIDIA cuVS actually does at 14 million vectors.https://ure.us/articles/visual-rag-beats-the-vision-model/Fri, 12 Jun 2026 00:00:00 +0000https://ure.us/articles/visual-rag-beats-the-vision-model/A distilled retrieval corpus beat a 3B vision model on accuracy at 180x the speed, on the hardest images online. Why engineering beats brute-force inference.https://ure.us/articles/benchmarking-nvfp4-blackwell/Mon, 08 Jun 2026 00:00:00 +0000https://ure.us/articles/benchmarking-nvfp4-blackwell/An independent 16-arm benchmark of FP8, INT4-AWQ and NVFP4 vs BF16 on a 96 GB Blackwell workstation: quality, throughput, and cross-validation vs NVIDIA.https://ure.us/articles/gpu-fleet-aiops-llm-backend-benchmark/Tue, 07 Apr 2026 00:00:00 +0000https://ure.us/articles/gpu-fleet-aiops-llm-backend-benchmark/Benchmarking seven LLM backends as autonomous operators for an 8,000-GPU cluster with six realistic failure scenarios and deterministic checklist scoring.https://ure.us/articles/gpu-fleet-aiops-the-augmented-operator/Tue, 07 Apr 2026 00:00:00 +0000https://ure.us/articles/gpu-fleet-aiops-the-augmented-operator/Seven LLM backends competed to run an 8,000-GPU cluster. The free local model matched frontier accuracy at one-fifth the latency. The $32 model scored worst.https://ure.us/articles/292x-why-batch-inference-breaks-on-api-pricing/Thu, 02 Apr 2026 00:00:00 +0000https://ure.us/articles/292x-why-batch-inference-breaks-on-api-pricing/One rented B200 GPU processed a million documents in 11 hours for $70. The same workload through API providers costs up to $20,419 and takes 144 days.https://ure.us/articles/a-million-pdfs-wont-organize-themselves/Thu, 02 Apr 2026 00:00:00 +0000https://ure.us/articles/a-million-pdfs-wont-organize-themselves/A million corporate PDFs in a RAG pipeline revealed what tutorials never mention -- the parsing, embedding, and graph labor that no shortcut survives at scale.https://ure.us/articles/best-local-llm-scaling-coding-swarms/Mon, 09 Mar 2026 00:00:00 +0000https://ure.us/articles/best-local-llm-scaling-coding-swarms/Per-task throughput plateaus at four concurrent agents and holds flat through eight. Agents five through eight are free. The contention wall is a floor.https://ure.us/articles/best-local-llm-coding-agent-swarm/Sat, 07 Mar 2026 00:00:00 +0000https://ure.us/articles/best-local-llm-coding-agent-swarm/MoE is 4.9x faster than Dense when four coding agents share one GPU. We ran the concurrent-load benchmark nobody published - single-request numbers lied.https://ure.us/articles/best-local-llm-agentic-coding/Fri, 06 Mar 2026 00:00:00 +0000https://ure.us/articles/best-local-llm-agentic-coding/Real benchmarks on dual RTX 3090: the best local setup for agentic coding is one GPU and an MoE model. 168 tok/s, NVLink optional. Data and recommendations.https://ure.us/articles/why-foreign-ai-specialists-keep-failing/Mon, 02 Feb 2026 00:00:00 +0000https://ure.us/articles/why-foreign-ai-specialists-keep-failing/AI context is now portable - but domain gravity and the human translation layer still decide who wins when building real products in real markets.https://ure.us/articles/security-assurance-engineering-practical-example-ure-chapter-4/Thu, 15 Jan 2026 00:00:00 +0000https://ure.us/articles/security-assurance-engineering-practical-example-ure-chapter-4/URE Case 4/5. How security enables business by arriving early with solutions, not vetoes, and reshaping systems to preserve the mission.https://ure.us/articles/tail-latency-killed-beowulf-cluster-2006/Sun, 04 Jan 2026 00:00:00 +0000https://ure.us/articles/tail-latency-killed-beowulf-cluster-2006/In 2006, I learned that scaling out doesn't work when the interconnect is the bottleneck. Twenty years later, the same physics governs GPU infrastructure.https://ure.us/articles/telemetry-that-lies-gpu-thermal-monitoring/Sat, 27 Dec 2025 00:00:00 +0000https://ure.us/articles/telemetry-that-lies-gpu-thermal-monitoring/Your GPUs report 100% utilization while running slower. Temperatures look fine while racks drift hot. Thermal telemetry is easy to collect and hard to trust.