Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #6572

NVLink

Related event timeline, sources and context from the news index.

EVENT TIMELINE

6

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA Vera Rubin NVL72 leads MLPerf Endpoints v1.0 agentic inference with >10× throughput/kW vs GB300 (provisional)

Provisional MLPerf Endpoints v1.0 results show NVIDIA's Vera Rubin NVL72 delivering over 10× throughput per kilowatt versus the NVIDIA GB300 NVL72 at a shared minimum end-to-end interactivity of 50 tokens/sec per user; across Kimi K3, DeepSeek-V4.1-Flash and Qwen3.6-35B-A3B the Vera Rubin system achieves 5×–10× higher throughput/kW and 3×–5× higher end-to-end interactivity at matched concurrency. NVIDIA credits hardware (HBM4, Tensor Cores, 6th-gen NVLink) plus context-aware serving (prefix caching, state reuse, Dynamo routing) and model-specific kernels for the gains; the Vera Rubin results are provisional pending MLCommons verification.

8.0

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA launches open-source "NVIDIA Recovery" to automate distributed training recovery

NVIDIA announced NVIDIA Recovery, an open-source workload-level service in the DSX OS portfolio that detects training interruptions via application logs, GPU metrics, and scheduler signals and runs a four-stage loop to attribute failures, isolate nodes, and restart or stop jobs. The service integrates with Slurm via a Helm-deployable Advisor Controller, starts in monitor mode by default, provides an audit trail and Recovery Console, and the GitHub repo and quickstart allow local deployment without physical GPUs; production monitoring reported ~8,550 anomaly detections over 3 months with mean detection and recovery-action latencies of 69s and 94s respectively.

7.0

CODING · 1 SOURCE · NVIDIA Developer

NVIDIA releases CUDA Toolkit 13.5 with GA support for Vera Rubin, MPS v3 and multi‑GPU features

CUDA Toolkit 13.5 is generally available, moving NVIDIA Vera Rubin GPU architecture support to GA and extending Windows on Arm support to NVIDIA DGX Station for Windows. The release adds locality-domain SM partitioning, a major MPS v3 update (weighted timeslicing, SM partition load balancing with borrowing/lending, cgroup routing, per-namespace assignments, polling-based memory accounting), CUDA Tile multi‑GPU unicast/multicast on NVLink/NVSwitch and a SIMT interoperability preview, CUDA Python cuda.core 1.2 enhancements (copy_batch, LaunchConfig), and CCCL 3.5 library updates including tuning APIs and batched Top‑K tied into FlashInfer.

8.0

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA Minerva unifies telemetry into a topology-aware AIOps platform

NVIDIA Minerva ingests telemetry across GPUs, networking, storage, and workloads and builds a dynamic knowledge graph that maps infrastructure to active jobs in real time. Purpose-built ML models detect subtle "gray failures," topology-aware correlation groups alert storms into single incidents, and an AI assistant (Sherlock) plus an MCP server let operators query live evidence and integrate context into external workflows to preserve GPU hours and speed repairs.

7.0

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA cuOpt adds mPDLP multi‑GPU LP solver that scales to 100M+ variables

NVIDIA released mPDLP, a Multi‑GPU Primal‑Dual hybrid gradient solver in cuOpt that shards large linear programs across NVLink‑connected GPUs. mPDLP cuts per‑GPU peak memory (up to 6× vs single‑GPU PDLP), supports problems up to 2.1B nonzeros, and in DGX B200 benchmarks shows PDLP‑step speedups up to 11.4× and typical overall gains of 1.2×–2.5× vs prior D‑PDLP; partner cases reported ~3.3× (Kinaxis, 135M variables) and >5× (PSR, 185M variables). A tutorial and source code are available on GitHub.

7.0

CODING · 1 SOURCE · NVIDIA Developer

NVIDIA documents TensorRT LLM adaptations for confidential Blackwell (B200) inference

NVIDIA published guidance and measured results showing how TensorRT LLM adapts to confidential computing on Blackwell B200 hardware, including CC-aware memory selection, async token readback, using the GPU %globaltimer for autotuning, and communication-path detection when NVLS multicast is unavailable. In their tests, CC-enabled runs retained about 96.1–98.2% of token throughput versus CC-off, with per-token latency (TPOT) within roughly 1.2–4.3% of baseline.

7.0