Surprisingly Useful AI Article Enhancements
Nvidia has released a wave of technical detail across its next-generation hardware and software platforms, spanning new GPU and CPU architectures, AI training benchmarks, developer tools and simulation capabilities as it positions its infrastructure for the shift toward agentic AI workloads.
The disclosures cover the Rubin GPU, the Vera CPU with its Olympus core architecture, a training performance record on the GB300 NVL72 system, updates to NVLink networking, new tools for model customisation and inference observability, and integrations across gaming, video analytics and 3D simulation.
GB300 NVL72 Sets MoE Pre-Training Record
Nvidia reported that its GB300 NVL72 system achieved 1,648 teraflops (TFLOPs) per GPU when pre-training the DeepSeek-V3 671B mixture-of-experts (MoE) model using 256 GPUs and its Megatron Core framework. The company described this as a world record for MoE pre-training performance.
The result represents roughly three times the per-GPU throughput compared to the earlier GB200 NVL72, and a 1.5 times improvement on the same GB300 NVL72 hardware over a six-month period, driven by software optimisation alone.
MoE architectures have become the dominant approach to frontier model pre-training because they activate only a subset of a model’s parameters for each token. DeepSeek-V3, for example, holds 671 billion parameters but activates approximately 37 billion per token. The tradeoff is increased communication overhead, as tokens must be dispatched to and gathered from expert networks distributed across GPUs.
Nvidia’s open-source framework contributions also showed gains. TorchTitan, PyTorch’s native training stack, achieved roughly six times higher performance through Nvidia optimisations, while JAX performance improved by nearly 10 times over six months. When scaling from 256 to 1,024 GPUs, Megatron Core retained 98.5 per cent of its per-GPU performance, with TorchTitan and JAX each holding at 97 per cent.
Inside The Rubin GPU Architecture
The Rubin GPU is Nvidia’s next-generation accelerator, designed to deliver up to 10 times more agentic throughput per unit of energy than Blackwell, according to the company’s internal benchmarks.
Built from two reticle-limited compute dies unified through a high-speed inter-die link called NV-HBI, the chip contains 336 billion transistors, 224 streaming multiprocessors and 896 Tensor Cores. The third-generation Transformer Engine delivers up to 50 petaflops of NVFP4 inference performance.
On the memory side, Rubin integrates up to 288 GB of HBM4 memory with up to 22 TB/s of peak bandwidth, a 2.8 times increase over Blackwell. NVLink 6 provides 3,600 GB/s of scale-up bandwidth per GPU.
Key architectural changes include inline descriptor update support for the Tensor Memory Accelerator, which simplifies how MoE models manage expert weight movement. Rubin also doubles Tensor Core throughput per clock by processing a larger K dimension, reducing loop iterations for matrix operations.
For long-context inference, Rubin introduces activation sparsity features that compress intermediate attention data into a structured 2:4 sparse form, reducing both compute and data movement in the attention pipeline. Exponential throughput for softmax operations increases by two times for FP32 and four times for BF16/FP16 compared to Blackwell.
At rack scale, the Vera Rubin NVL72 system incorporates Intelligent Power Smoothing with energy storage, which Nvidia claims reduces average power consumption by approximately 10 per cent compared to previous-generation techniques. DSX MaxLPS can enable operators to provision up to 40 per cent more GPUs within the same power budget.
Vera CPU And The Olympus Core
Nvidia’s Vera CPU, built around its custom Olympus core, targets the CPU-side execution demands of agentic AI workloads, where agents run code, invoke tools, retrieve context and interact with databases.
The Olympus core features a 10-wide decode engine, a neural branch predictor, deep out-of-order execution and what Nvidia calls Spatial Multithreading, which partitions core resources more effectively across two hardware threads to reduce interference compared to traditional simultaneous multithreading approaches.
Each Vera CPU includes 88 Olympus cores and 176 SMT threads, paired with SOCAMM2 LPDDR5X memory delivering up to 1.2 TB/s of aggregate bandwidth. The Scalable Coherency Fabric provides up to 3.4 TB/s of bisectional bandwidth with a 164 MB unified L3 cache.
Related: Best Business Laptops for work & school
Related: Best Gaming Laptops
Related: Best Portable Laptop
Nvidia claims the Vera CPU delivers up to 1.8 times higher performance on agentic workloads compared to x86 CPUs, based on internal SPEC CPU 2026 measurements.
The chip uses a monolithic compute die rather than a chiplet design, with each socket presenting as a single NUMA domain. NVLink-C2C provides 1,800 GB/s of coherent CPU-GPU bandwidth, and the platform supports PCIe 6.4 and CXL 3.1.
NVLink Reaches Sixth Generation
Now in its sixth generation, NVLink provides 3.6 TB/s per GPU of bidirectional bandwidth and 260 TB/s of rack-level GPU bandwidth in a 72-GPU domain. Nvidia claims three times lower end-to-end latency and 10 times higher packet rates compared to off-the-shelf Ethernet alternatives, along with 130 TFLOPs of in-network compute for collective operations.
The NVLink 6 Switch introduces management and resiliency features including hot-swappable switch trays, dynamic traffic rerouting, in-service software updates and fine-grained link telemetry.
Nvidia also detailed NVLink Fusion, an interconnect technology that allows hyperscalers and AI-native companies to connect custom silicon to Nvidia’s infrastructure platform. The technology roadmap includes support for scale-up domain sizes up to 1,152 GPUs and connectivity through co-packaged optics.
BlueField-4 For Agentic Infrastructure
Nvidia outlined how its BlueField-4 data processing unit and the new Vera BlueField-4 STX storage processor fit into agentic AI infrastructure.
BlueField-4 integrates up to 800 Gb/s connectivity, a 64-core Grace CPU, LPDDR5X memory and PCIe Gen6. Compared to BlueField-3, it doubles networking bandwidth and delivers up to six times more compute performance.
The Vera BlueField-4 STX powers Nvidia’s CMX context memory storage platform, designed to store, share and retrieve KV cache data across inference workloads without stalling GPU processing. DOCA Memos, a software component, manages KV cache sharing across compute and storage nodes.
Nemotron 3 Customisation Via Hosted Training
Nvidia published a tutorial demonstrating how developers can customise its Nemotron 3 Nano model using reinforcement learning with verifiable rewards through Prime Intellect Lab’s hosted training service.
The workflow requires roughly five minutes of local setup and no local GPU cluster. In the demonstrated example, accuracy on a Python math benchmark increased from 21.9 per cent to 90.6 per cent after 100 training steps, at a total cost of less than $5.
The same workflow applies to the larger Nemotron 3 Super and Nemotron 3 Ultra models by changing the model identifier.
TensorRT Builds Made Observable And Cancelable
Nvidia released a tutorial on implementing IProgressMonitor, a TensorRT API that provides real-time progress reporting and cancellation capabilities during engine builds. The feature addresses a practical problem for developers, end users and AI agents: long-running builds that offer no visibility into progress and no way to abort.
The implementation works in both Python and C++, with the step_complete callback serving as the cancellation mechanism. The tutorial covers integration points for IDE extensions, HTTP services and agent runtimes.
Capcom Brings Path Tracing To RE ENGINE
Capcom’s RE ENGINE team detailed its implementation of path tracing across both Resident Evil Requiem and PRAGMATA, built on Nvidia’s RTX Kit.
The two-year development effort moved direct lighting from shadow maps to path tracing, reducing the visual gap between gameplay and cutscenes. The implementation uses DLSS Ray Reconstruction for denoising and includes real-time light transmission through strand hair.
The RE ENGINE team noted that the relatively contained environments in both titles worked to their advantage, and that RayQuery-based ray tracing allowed continuous validation across a range of devices.
Omniverse Libraries For Sensor Simulation
Nvidia released ovrtx as pre-release software on GitHub, a lightweight C and Python SDK for generating camera, lidar, radar and related sensor outputs from OpenUSD content. The library is now part of Nvidia Agent Toolkit.
PTC’s Onshape Render Studio team is exploring how ovrtx can fit into cloud-native CAD workflows.
“ovrtx gives us a lower-level API path for bringing Nvidia RTX into cloud-native design tools while keeping the application architecture and product experience in our control,” Brendan Hack, Senior Director of Software Development at PTC Onshape, commented.
Video AI Agents And Enterprise Integration
Nvidia detailed how its Video Search and Summarisation blueprint, RAG blueprint and NemoClaw agent orchestration framework can work together to move video analytics from static reports to automated action, including generating Jira tickets from video analysis findings.
In one demonstrated example, a food video analysis pipeline captured user intent through human-in-the-loop prompts, retrieved nutritional guidelines via RAG, analysed the video with that context and automatically created a tracked action item.
Computacenter deployed the pipeline for predictive maintenance, processing drone and thermal inspection footage and auto-drafting Maximo work orders, cutting footage-to-work-order time from 30-45 minutes to roughly 19 seconds.
DeepStream 9.1 And Multi-Camera 3D Tracking
DeepStream 9.1 introduces Multi-View 3D Tracking (MV3DT) and AutoMagicCalib (AMC), which fuse detections from multiple auto-calibrated cameras into a shared 3D coordinate system while maintaining consistent object IDs across views. The release includes 13 agentic skills and support for JetPack 7.2 on Jetson edge platforms.
CUDA 13.3 Adds Carryless Multiplication
CUDA 13.3 introduces clmad, a PTX instruction for hardware-accelerated carryless multiplication available on all Ampere and newer GPUs. On the B200, GHASH throughput reaches approximately 6.3 TB/s, up to 18.8 times faster than the prior bitsliced approach. Sum-check performance over GF(2^128) improved by four to 13 times.
The instruction supports workloads including AES-GCM authenticated encryption, CRC and Reed-Solomon codes, post-quantum cryptographic schemes and binary-field arithmetic used in zero-knowledge proving systems.
OpenUSD Runtimes Via AI Agents
Nvidia published nanousd-labs as part of Omniverse Labs, demonstrating how AI agents can generate USD runtimes directly from the USD Core Specification maintained by the Alliance for OpenUSD. The approach uses the formal specification as a contract that agents build and validate against, rather than adapting an existing codebase.
Last Updated on July 26, 2026 by Nick Ross
