Accelerated Computing · Est. 1993 · Santa Clara

THE ENGINE ROOM OF INTELLIGENCE.

One architecture. From a single die of silicon to planet-scale AI factories — this is a descent through the machine that thinks.

Macro photograph of an NVIDIA Turing TU104 GPU die — billions of transistors etched in silicon, lit from the edge
TU104 · TURING · 13.6 BILLION TRANSISTORS · 12 NM
BLACKWELL GB200 NVL72 1.4 EXAFLOPS FP4 GEFORCE RTX 5090 92B TRANSISTORS CUDA-X 900+ LIBRARIES NVLINK 1.8 TB/S PER GPU DGX SPARK 1 PFLOP ON A DESK TSMC 4NP FOXCONN SIEMENS MERCEDES-BENZ OMNIVERSE OPENUSD NATIVE JETSON THOR 2070 TFLOPS SPECTRUM-X 800 GB/S ETHERNET

Data centers · Blackwell · NVLink

AI FACTORIES

The data center is no longer a warehouse for files. It is a machine with one job: manufacturing intelligence. Raw power in, tokens out — measured not in megabytes, but in reasoning per watt.

NVIDIA HGX B200 NVL8 system — eight Blackwell GPUs behind gold heat spreaders on a dark chassis
HGX B200 NVL8 — eight Blackwell dies, one heartbeat
0GPUs in one GB200 NVL72 rack, acting as a single GPU
0exaflops of FP4 inference per rack
0faster inference than Hopper on trillion-parameter models
0of NVLink bandwidth stitched through one rack spine
  • GB300 NVL72
  • DGX SuperPOD
  • Spectrum-X
  • Quantum-X Photonics
  • Rubin · 2026

Gaming · RTX · Neural rendering

GEFORCE RTX.

Ray tracing was the beginning. Now the frame itself is imagined — neural shaders, DLSS 4 multi-frame generation, and paths of light simulated before your eyes knew to ask.

GeForce RTX 5090 graphics card — triple-fan black shroud with iridescent edge lighting
GeForce RTX 5090 · 92 billion transistors
DLSS 4Multi-frame generation — up to 8× the frame rate, composed by AI
4th-gen RT coresFull path tracing in real time, every photon accounted for
GDDR71.8 TB/s of memory bandwidth feeding 21,760 CUDA cores
Reflex 2Frame-warped latency, measured in heartbeats not milliseconds
COMPUTE IS THE NEW GRAVITY.

Everything built this century bends around it.

Simulation · OpenUSD · Digital twins

OMNI VERSE

Before a factory is built, it is simulated. Before a robot walks, it walks a million times in a world made of light and physics. Omniverse is where the physical world rehearses.

Digital Twins

Factories, warehouses and cities mirrored atom-for-atom in OpenUSD, streamed at planetary scale.

BMW · Siemens · Foxconn
Synthetic Data

Photoreal worlds that generate the edge cases reality is too stingy to provide — on demand, labeled, infinite.

Omniverse Replicator
Cosmos

World foundation models that understand physics — predicting the next frame of reality itself.

NVIDIA Cosmos

Jetson · DRIVE · Isaac · GR00T

MACHINES THAT MOVE

Intelligence is leaving the screen. Jetson thinks at the edge, DRIVE reads the road, Isaac teaches robots to grasp — and GR00T gives humanoids a mind that generalizes.

NVIDIA Jetson AGX Orin developer kit — a compact finned-silver edge AI computer beside its packaging
Jetson AGX Orin · 275 TOPS at the edge
JETSON THOR2070 TFLOPS for humanoid robots, in 130 watts
DRIVE AGX THOROne computer per car: perception, planning, cockpit
ISAAC SIMPhysics-honest rehearsal for every actuator and gripper
GR00T NA foundation model for generalized humanoid skill

CUDA · CUDA-X · Developer universe

THE CUDA UNIVERSE

Beneath everything above — every factory, every frame, every twin — sits a kernel. Millions of threads launched like a choir, each one singing exactly one note, perfectly on time.

0CUDA-X libraries — from cuDNN to cuQuantum
0developers shipping on CUDA worldwide
0CUDA downloads of the toolkit, every year
Start building
View more demos Get $10 off Kimi K3