June 23, 2025

CUDA programming and my RTX 3060

My RTX 3060 is a monster with 3 fans made by Gigabyte. It has 12G of VRAM.

It has 3,584 CUDA cores and CUDA 11.x can run on it.

The 3060 uses the Ampere architecture.

"Computer Capability" (CC) for the 3060 is 8.6. This is explained in detail here:

Pulling information from tables in the above, we find that:

Each SM (streaming multiprocessor) has:

128 FP32 cores
64 (?) FP64 cores
64 INT32 cores
4 3rd generation Tensor cores
16 special function units for FP32 transcendentals
4 warp schedulers
read only constant cache
unified data cache and shared memory 128K
Digging into other sources we learn this about the RTX 3060.
(Note that my 12G version uses the GA106-300-A1 chip variant).
13.25 billion transistors
28 SM
3584 CUDA cores
28 RT cores (ray tracing)
112 tensor cores
runs at 1320 Mhz
3M of L2 cache
Wikipedia shows the Core config as 3584 112:48:28:112. What the heck does that mean? This counts shaders-texture:render:RT:tensor.

They say the RTX 3060 has a 192 bit memory path, compared to a 256 path in the 3060-Ti and 3070. So 6 chips with 32 bit path yield 192 width. Each chip is 2G so we get 12G.


Feedback? Questions? Drop me a line!

Tom's Computer Info / tom@mmto.org