Calculating CUDA cores

Accelerated Computing CUDA CUDA Programming and Performance
1

I am learning pyCUDA through the book Getting Started with PyCUDA. There is a step where we calculate the amount of CUDA cores per multiprocessor. The number I get is 3840 CUDA cores for the GeForce RTX 2060. However when I search on google, google is telling me that the GeForce RTX 2060 has 1920 CUDA cores (original estimate divided by 2). My question is who is right?

Device: GeForce RTX 2060
Compute Capability: 7.5
Total Memory: 6144 mb
Number of Multiprocessors: 30

To get my original estimate I used pyCUDA to determine the number of steaming multiprocessors as 30. Within each multiprocessor there is a certain number of CUDA cores I found that number with pyCUDA and from this table within the nvidia docs: Programming Guide :: CUDA Toolkit Documentation

After getting those two numbers I multiplied them and got 3,840.

I guess what I am asking is why does google tell me the number of CUDA cores is 1920 instead of 3840 ? Why divide by 2 ?

So I’m not sure what I am doing wrong or not accounting for to justify the divide by 2 at the end.

2

What piece of data specifically did you take from “this table within the nvidia docs”? Table number, description of row?

Here is the relevant information I see (not in a table):

I.6. Compute Capability 7.x
I.6.1. Architecture
An SM consists of:
64 FP32 cores for single-precision arithmetic operations, […]

So, for an RTX 2060 with 30 SMs comprising 64 CUDA cores each: 1920 CUDA cores total

3

Ah ok looks like I read the table wrong. I was looking at table 15 Maximum number of resident grids per device

4

There is no such row in table 15. I think you meant “Maximum number of resident threads per SM”?

Note that “CUDA cores” is a marketing term for “FP32 cores”. Best I can see, the relevant numbers are not part of any table, but appear in the architecture summaries after table 15.