NVIDIA AI Updates: November 6, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. NVIDIA Publishes Agent Benchmarks for the Vera CPU, Including Daytona Bringing 1,000 Sandboxes Online in 10.8 Seconds
NVIDIA published agent-workload results for its Vera CPU, arguing that a single part can replace the mix of fast and dense x86 CPUs that providers use to balance agent speed against capacity in power-limited AI factories. In a 100-megawatt data center comparison, NVIDIA says Vera runs 1.6x more agents per watt than the fastest x86 CPU and delivers 2x the agent speed of the most energy-efficient x86 CPU, without naming either part. Partner results: Perplexity ran agentic coding tasks 1.5x faster than on traditional CPUs and plans to deploy Vera in an upcoming production system; DeepInfra measured up to 2.2x the speed of the best x86 processor on its production agent orchestration workload, the fastest of four architectures tested; ClickHouse’s standard binary, without Vera-specific tuning, scaled near-linearly across 176 cores and topped the ClickBench hardware leaderboard; DuckLabs reached about 1.5x the median TPC-H SF1000 performance of an Intel Xeon 6 baseline; Daytona brought 1,000 agent sandboxes online in 10.8 seconds against 88.4 seconds on x86 and completed 12.7 jobs per second at 2,000 concurrent sandboxes; and Modal reported up to 2x sandbox throughput per core against comparable x86 hardware, with the gap growing under heavy load. NVIDIA’s technical blog lists Vera at 88 Olympus cores with 1.8x faster single-threaded performance and 40% lower peak loaded latency than x86 CPUs, up to 14 GB/s of memory bandwidth per core and up to 1.5 TB of LPDDR5X. Source
2. NVIDIA Details Five Memory Levers for Agentic Inference, From NVFP4 KV Cache to a Pod-Shared CMX Context Tier
NVIDIA laid out how it stretches GPU memory for agentic inference as HBM supply tightens, noting that HBM capacity per GPU stays at 288 GB from GB300 to Rubin (its main blog puts the increase at roughly 3%) while bandwidth rises to 19.2 TB/s. It credits a combination of five techniques for gains it measured on SemiAnalysis AgentX, which replays recorded coding-agent sessions: up to 30x the throughput per megawatt of GB300 NVL72 for Vera Rubin NVL72 at 160 tokens per second per user, and up to 30x the throughput per dollar at 125 tokens per second per user on a DeepSeek V4 Pro workload. The five techniques: NVFP4 (4-bit E2M1 values, an FP8 scale per 16-value block and a per-tensor FP32 scale), which shrinks weights about 3.5x against FP16 and 1.8x against FP8 and, applied to the KV cache, cuts its footprint by up to 50% against FP8 while keeping over 99% of BF16 baseline scores across 10 DeepSeek-V4-Pro benchmarks; wide expert parallelism in TensorRT-LLM, where EP32 on GB200 NVL72 delivers up to 1.8x the output tokens per GPU of EP8; Dynamo’s KV-aware routing, disaggregated serving and KV Cache Runner, which tracks blocks across GPU, pinned CPU, local SSD and network filesystem tiers through NIXL, with a KVCR-Guard feature that preserves committed host-memory cache across an engine failure; parking idle sessions’ KV cache in Vera’s LPDDR5X during tool calls and bringing it back over 1.8 TB/s NVLink-C2C; and CMX context memory storage, a pod-shared “G3.5” tier of Ethernet-attached flash built on BlueField-4 STX storage processors, Spectrum-X and DOCA Memos, which NVIDIA says delivers up to 5x the tokens per second and power efficiency of traditional storage for long-context inference. NVIDIA also lists models released with NVFP4 checkpoints, including Black Forest Labs’ FLUX.2 [klein], Thinking Machines Lab’s Inkling, Reflection’s Beam 1.0, Meta’s Muse Spark, StepFun’s Step 5 Preview and Nemotron 3.5 Super VL. Source NVIDIA Blog
3. NVIDIA Packages AICR, Cluster Readiness Engine and NVSentinel as DSX OS for Continuous GPU Cluster Validation
NVIDIA described DSX OS, the open source software layer of its DSX AI Factory Platform, as three Apache 2.0 projects that cover a Kubernetes GPU cluster’s lifecycle: AI Cluster Runtime (AICR) to configure, the NVIDIA Cluster Readiness Engine (NVCRE) to certify, and NVSentinel to operate. AICR emits version-pinned recipes with a dependency graph and per-component health assertions (for example aicr recipe --service eks --accelerator h100 --os ubuntu --intent inference --platform dynamo), renders them for Helm, Argo CD, Flux or Helmfile, and aicr diff --fail-on-drift turns drift detection into a CI gate. NVCRE runs a Certification resource of NCCL communication tests, DCGM level 4 diagnostics and NeMo pretraining runs on 8B and 56B-parameter models, or a user’s own containerized test through a WorkloadRun, and bisects failing node groups to name suspect nodes; performance thresholds such as busBandwidthGBps: "value >= 300" are not set by default, so without them a slow node can still pass. NVSentinel installs in monitoring-only mode and, once enabled, cordons or taints nodes through CEL-based quarantine rules with a circuit breaker that caps how many GPU nodes can be quarantined in a sliding window. With lifecycle validation enabled, it creates a ValidationRequest when a node joins or returns from remediation, gates scheduling until NVCRE tests pass and keeps nodes that still fail cordoned; in-place driver, operator and fabric changes do not trigger revalidation automatically, so operators must cordon the nodes and request validation themselves. NVIDIA keeps attribution (NVCRE) and quarantine (NVSentinel) separate to avoid double cordons, so platform teams can swap in their own node lifecycle tooling. Source