GPU Recommendation for XGBoost with MATLAB R2022b

Hi,

I'm currently planning to use XGBoost with GPU acceleration, mainly for parameter optimization. Right now, I'm performing leave-one-out cross-validation, and my feature dimensions are around 1000.

I haven’t purchased a GPU yet, but I'm considering either:

GeForce RTX 5060 Ti/ 5060 (though I couldn’t find its compute capability listed on the official NVIDIA site: https://developer.nvidia.com/cuda-gpus ), or

GeForce RTX 4070 Ti, which has a compute capability of 8.9.

I'm using MATLAB R2022b, and according to the documentation, "MATLAB® supports NVIDIA® GPU architectures with compute capability 3.5 to 8.x."

Could anyone clarify whether either (or both) of these GPUs would be supported? Or perhaps suggest an alternative GPU that would work well with my setup?

Thanks in advance!

[Walter Roberson]

Note that the 40x0 and 40x0 Ti series all have floating point double precision (FP64) performance at 1/64 of the FP32 performance.

• 4070 Ti Super: 689.0 GFLOPS
• 5070 Ti: 686.6 GFLOPS
• 4070 Ti: 626.4 GFLOPS

Compare this to the end-of-life 1080 Ti, at 354.4 (a 1:32 rating insead of a 1:64 rating), or the 3080 Ti at 532.8 GFLOPS (1:64). This leads to questions about whether it is more cost effective to purchase multiple lower-rated GPUs if your intention is to do double precison work

[Walter Roberson]

I find claims that XGBoost treats all values as single precision floating point, so my concerns about double precision performance are perhaps misplaced here.

• 4070 Ti Super: 44.10 TFLOPS fp32
• 4060 Ti: 22.06 TFLOPS fp32
• 5070 Ti: 43.94 TFLOPS fp32
• 5060 Ti: 23.70 TFLOPS fp32
• 1080 Ti: 11.34 TFLOPS fp32

[JIA]

I understand now—thank you so much for the information and help you provided.

Answer 1

[Joss Knight]

Yes, MATLAB R2022b supports these GPUs.

[JIA]

Sorry. I mean 5060Ti.

[Walter Roberson]

The 5060 Ti is Compute Capacity 12.0 .

This device is not fully supported by MATLAB R2022b (not even by R2025a Pre-release).

There is a forward-support mode that can be enabled that will cause kernels to be automatically re-built on demand. However, problems can be expected, the more so for older versions such as R2022b. It is unclear at the moment whether the problems will be trivial or "a nuisance to work around" or "quite serious".

The 5060 Ti is rated for 370.4 GFLOPS of FP64, which is considerably worse than the FP64 performance of the 4070 Ti.

[Joss Knight]

I think it's fair to say that R2022b isn't going to work well with the RTX 5060 in forward compatibility mode. R2024a, R2024b and R2025a should be okay (although be wary of the caveats in this post).

[Walter Roberson]

By the way, to enable forward compatibility see https://www.mathworks.com/help/parallel-computing/parallel.gpu.enablecudaforwardcompatibility.html

[JIA]

First of all, thank you all for taking the time to respond to my question.

• Due to budget constraints and memory capacity considerations, I’ll most likely go with a 30-series GPU with 12GB of VRAM (in Taiwan, it’s currently difficult to find a 40-series Ti card with more than 8GB).
• I’m also concerned that enabling forward compatibility might lead to incorrect results, so I’ve decided to choose a slightly older GPU model instead.

Wishing everyone the best with their MATLAB work—and hope you all manage to get the GPU you need!

[Walter Roberson]

• RTX 3060 fp32 12.74 TFLOPS
• RTX 3060 Ti fp32 16.20 TFLOPS
• RTX 3070 fp32 20.31 TFLOPS
• RTX 3070 ti fp32 21.75 TFLOPS
• RTX 3080 fp32 29.77 TFLOPS
• RTX 3080 ti fp32 34.10 TFLOPS
• RTX 3090 fp32 35.58 TFLOPS
• RTX 3090 ti fp32 40.00 TFLOPS

[Joss Knight]

Good choice. The 30xx series is still pretty recent and most of the improvements since have been in reduced precision that you are not using.