Nvidia will present its latest Blackwell GPU innovations and architectural advances at the Hot Chips 2025 technical conference.
Nvidia Corporation is unveiling major advancements in GPUs, networking, and AI efficiency at the Hot Chips 2025 conference, taking place August 24–26 at Stanford University. The semiconductor design conference will feature the company's latest innovations in rack-scale AI systems, neural rendering, and efficiency-driven training methods. Nvidia executives will deliver six sessions, including a tutorial on designing rack-scale architectures for data centers, with key topics covering the company's Spectrum-XGS networking platform, NVLink Fusion technology, and next-generation Blackwell GPUs.
To address growing demands for AI model training and inference across multiple locations, Nvidia is introducing Spectrum-XGS Ethernet to interconnect multiple data centers into massive AI super-factories. The solution enables near-doubling of performance in multi-site AI workloads, while co-packaged optics switches aim to reduce power usage and latency. Idan Burstein, principal architect at Nvidia, will detail the ConnectX-8 SuperNIC, a key piece of Nvidia's networking stack that powers high-speed GPU-to-GPU communication at scale.
Marc Blackstein, Nvidia's senior director of architecture, will highlight how the company's Blackwell architecture is advancing inference and simulation performance. The GeForce RTX 5090 GPU brings neural rendering enhancements for gaming and visualization, with Nvidia DLSS 4 delivering up to 10x improvement in design cycles and realism. Another breakthrough, the NVFP4 4-bit floating-point format, promises to quadruple AI training efficiency without compromising accuracy, marking a leap forward in large language model training. Andi Skende, senior distinguished engineer at Nvidia, will showcase the DGX Spark desktop AI supercomputer powered by the GB10 Superchip, giving researchers and developers high-performance AI capabilities in a compact form factor.
Nvidia continues expanding support for open-source AI tools, including TensorRT-LLM, PyTorch, and vLLM, as well as NIM microservices for popular open models like GPT-OSS and Meta Platforms, Inc.'s Llama 4. Valued at over $4.34 trillion, Nvidia continues to dominate the AI chip market with its Blackwell platform and growing networking portfolio. UBS and Wedbush analysts remain optimistic, pointing to high demand for Nvidia's B200, GB200, and upcoming GB300 GPUs, along with growing hyperscale spending and large-scale infrastructure projects. The company is set for an earnings report on August 27.