New Street Big Ideas in Semiconductors Conference presentation
Logotype for NVIDIA Corporation

NVIDIA (NVDA) New Street Big Ideas in Semiconductors Conference presentation summary

Event summary combining transcript, slides, and related documents.

Logotype for NVIDIA Corporation

New Street Big Ideas in Semiconductors Conference presentation summary

17 Nov, 2026

AI model scaling

  • Language-model complexity was described as doubling every 2 months; the presentation projected a 100T-parameter single model by 2023.

  • Training GPT-3’s 175B-parameter model required 314 zettaFLOP, stated as 3,640 PFLOP/s-days; model growth was framed as constrained by economics.

  • Few-shot learning was presented as enabling generalized capabilities such as translation from a small number of examples; GPT-3 training cost was cited at $12M.

Megatron and computing infrastructure

  • Megatron training used PyTorch and combined CUDA, CUDA-X AI, NVSwitch, DGX SuperPOD, NCCL, cuBLAS and cuDNN.

  • A 1T-parameter model configuration was listed with 25,600 hidden size, 128 layers, 512-way model parallelism and 3,072 GPUs.

  • Training sustained 502 PFLOP/s across 3,072 GPUs, equal to 52% of tensor-core peak; per-GPU performance was 163 TFLOP/s.

  • Language models were identified as useful for products and services, with a proposed $1B model framed as needing to reinvent an entire company.

Deep learning applications

  • DLSS was presented as applying deep learning to ray tracing and image upscaling, using 128-megapixel 16K supercomputer-rendered ground truth for 4K output.

  • The presentation positioned AI as a way to transform and grow core business and to provide tools for other organizations to apply advanced AI.

  • A self-driving-car platform diagram linked data and HD-map generation, synthetic data, perception and driving AI training, DRIVE AV and Drive Sim.

Partial view of Summaries dataset, powered by Quartr API
AI can get things wrong. Verify important information.
All investor relations material. One API.
Learn more

Latest events from NVIDIA