Edited by 
Short answer
Tensor Cores multiply small matrices in one step, which is what neural networks need; in games they run DLSS upscaling, frame generation and denoising. RT Cores speed up ray tracing by testing which objects each ray of light hits. CUDA cores do general shading and computing. Every RTX streaming multiprocessor has all three: on the RTX 50 series, 128 CUDA cores, 4 Tensor Cores and 1 RT Core per SM. For AI work, only the Tensor Cores and memory matter; RT Cores sit idle.
Tensor Cores vs RT Cores at a glance
| Tensor Cores | RT Cores | |
|---|---|---|
| Job | Matrix multiply-accumulate, D = A × B + C | Ray traversal and ray-triangle intersection tests |
| Used for | AI training and inference, DLSS, denoising | Ray-traced lighting, shadows and reflections, path tracing |
| First appeared | Volta V100 (2017), consumer cards with Turing (2018) | Turing, RTX 20 series (2018) |
| Per SM on RTX 50 | 4 (fifth generation) | 1 (fourth generation) |
| In data-center AI GPUs | Yes, the main source of performance | No: A100, H100, B200 have none |
| NVIDIA's rating unit | AI TOPS or Tensor TFLOPS | RT TFLOPS |
What Tensor Cores do
A neural network is mostly matrix multiplication. A CUDA core multiplies two numbers per clock; a Tensor Core multiplies a whole small tile of numbers in one operation, which gives many times more throughput for the same chip area. Each generation added faster, smaller number formats: FP16 on Volta, TF32 and BF16 on Ampere, FP8 on Hopper and Ada, and FP4 on Blackwell. Read What is a Tensor Core? and Tensor Core generations for the details.
What RT Cores do
Ray tracing simulates light by sending rays into a scene and finding the first triangle each ray hits. A scene has millions of triangles, so GPUs organize them in a tree of boxes called a bounding volume hierarchy (BVH). The RT Core walks that tree and runs the box and triangle intersection tests in fixed-function hardware, leaving the CUDA cores free to calculate the color of each hit. Without RT Cores, the same work in software is many times slower.
How they work together in a game frame
- CUDA cores run shaders that draw the scene at a lower resolution.
- RT Cores trace rays for lighting, shadows and reflections.
- Tensor Cores run DLSS: Ray Reconstruction cleans up noisy ray-traced lighting, Super Resolution upscales the image, and Frame Generation creates extra frames between rendered ones.
This is why NVIDIA pairs the two: ray tracing is expensive, and AI upscaling wins back the frame rate it costs.
Core counts by card
| GPU | SMs | CUDA cores | Tensor Cores | RT Cores |
|---|---|---|---|---|
| GeForce RTX 5090 | 170 | 21,760 | 680 (5th gen) | 170 (4th gen) |
| GeForce RTX 4090 | 128 | 16,384 | 512 (4th gen) | 128 (3rd gen) |
| RTX PRO 6000 Blackwell | 188 | 24,064 | 752 (5th gen) | 188 (4th gen) |
| H100 SXM (data center) | 132 | 16,896 | 528 (4th gen) | none |
Generations
| Architecture | Consumer cards | Tensor Core gen | RT Core gen | Key addition |
|---|---|---|---|---|
| Turing (2018) | RTX 20 | 2nd | 1st | First RT Cores and DLSS |
| Ampere (2020) | RTX 30 | 3rd | 2nd | Sparsity, TF32, BF16 |
| Ada Lovelace (2022) | RTX 40 | 4th | 3rd | FP8, DLSS Frame Generation |
| Blackwell (2025) | RTX 50 | 5th | 4th | FP4, DLSS Multi Frame Generation |
Which cores matter for AI?
For running or training models, Tensor Cores and memory bandwidth do nearly all the work. RT Cores are not used, which is why data-center GPUs such as the H100 and B200 leave them out and spend the space on more Tensor Cores and memory. When choosing a card for local AI, compare VRAM and bandwidth in our GPU specs database and check fit in the VRAM calculator.
Frequently asked questions
Are RT Cores used for AI?
Not by mainstream AI frameworks. Training and inference run on Tensor Cores and CUDA cores. RT Cores only speed up ray-tracing workloads, including some rendering and simulation software.
Do Tensor Cores help in games?
Yes. They run DLSS upscaling, frame generation and ray reconstruction, which can raise frame rates several times in supported games.
Does the GTX 16 series have Tensor Cores or RT Cores?
No. GTX 16 cards use the Turing architecture but without either; Tensor Cores and RT Cores arrived together on consumer cards with the RTX 20 series.
Does AMD have RT Cores and Tensor Cores?
AMD Radeon GPUs have ray accelerators for ray tracing and, on RDNA 3 and RDNA 4, AI accelerators for matrix math. They do a similar job under different names; AMD's Instinct data-center GPUs use Matrix Cores.
