Compare machine learning performance on AMD/ Intel dedicated cards

Hardware Hub GPU
1

I’m interested in learning more about machine learning, Bayes networks, Tensor flow. What should I buy? Should I wait for the next gen Radeon, say… Navi 41? I only care about Linux environments

Has anyone compared the machine learning compatibility and performance for non CUDA dedicated runtimes? Can Tensor flow libraries run on any of them?

ROCm does not make it clear what is being supported beyond the compute parts. At least it’s not clear to me. And compute Radeon Instinct parts are priced at a ridiculous prices.

Is ROCm compatible with Intel Arc 770? Are any of these a good platform hardware to start learning about machine learning?

2

Intel has their own oneAPI solution, as far as I know only AMD supports ROCm at the moment. Not sure about it, but it should be supported on consumer GPUs, too - HOWEVER a lot of tools in the ecosystem are build only for CUDA and thus, if you’re getting serious about machine learning I would strongly urge going for an Nvidia GPU. Especially if you want to go into deep learning with PyTorch, you will want CUDA for all kinds of libraries and tools that are CUDA only and will make your life easy

1 Like
3

Ok, I’ve found a used Zotac RTX 2080 Super at 270$. Not sure if this is from a miner or not, I’m quite suspicious/ anxious.

4

if they insist on paypal gift, tell them you’ll pay extra to cover charges so that paypal will cover any dispute.

Mining is only an issue if they try to hide damage, if it’s been kept dust free, a stable temp will mean it’ll be in better condition than one heavily gamed on.

2 Likes
5

There are CUDA to OpenCL solutions to deal with CUDA specific Tensorflow libraries, your mileage varies wildly in terms of consumer Radeon GPU. The Radeon VII was the closest gaming/creator GPU at an affordable price point which gave impressive compute performance.

Can’t say much about Intel Arc as I don’t own one, however from testing CUDA to OpenCL with Intel Xe IGP(96EU) it barely performs to a GTX 950 for compute so your mileage is going to vary with how Arc scales to GeForce/Radeon and typically the Intel “compute” optimized libraries in my experience are Tensorflow Lite.

1 Like
6

Just for lulz, I’m watching “State of ROCm 5.3 in 2022: 6x Mi210…” with Wendell State of ROCm 5.3 in 2022: 6x Mi210, 1petaflop, in the 2u Supermicro AS -2114GT-DNR - YouTube . He makes a nice impression of the “Shining” (1980) with CUDA and Radeon compute.

I want a spot on scene when AMD breaks the door. If I ever get to do anything with AI, I’m pretty sure it won’t be CUDA. Maybe I will try Nvidia to be able to compare learning experiences, but even as it’s a palace and a fortress, I will still prefer to be in my shabby cabin.

7

If you are interested in learning more about machine learning, Bayes networks, and Tensorflow, you do not necessarily need to buy any hardware at the moment. You can start by learning the basics of machine learning theory and programming on your existing computer. There are plenty of free resources available online, such as online courses, tutorials, and forums, that can help you get started.

Regarding hardware, it depends on your specific needs and budget. Navi 41 is not yet released, and it is not clear how it will perform for machine learning tasks. However, if you are on a budget, you can consider purchasing a consumer-level GPU, such as the NVIDIA GeForce RTX series or the AMD Radeon RX series. These GPUs have good support for machine learning libraries, including Tensorflow, and can be used with Linux environments.

Tensorflow libraries can run on non-CUDA dedicated runtimes, such as ROCm, OpenCL, and Vulkan. ROCm is a platform for running HPC and AI workloads on AMD GPUs and CPUs. It supports various AMD GPUs, including Radeon Instinct, Radeon VII, and Radeon RX 5700 series, and it is compatible with Linux environments. However, compatibility with Intel Arc 770 is not yet clear, as it is a new architecture that has not yet been released.

As for starting with machine learning hardware, a dedicated GPU is recommended for training models faster. However, you can start with a CPU-only setup and move to a GPU when you need more performance. Additionally, cloud computing platforms, such as Amazon Web Services, Google Cloud Platform, or if you are looking for economical options then Ace Cloud GPU offer virtual machines with pre-installed machine learning libraries and GPUs, allowing you to experiment without investing in hardware upfront.

1 Like
8

There are free compute resources on Google Colab, if you check out this tutorial they use it for some of the experiments.

9

ROCm is only compatible with AMD cards
At the moment only ROCm 4.2 is supported main stream by most things
Feature levels are broken down into levels
GFX 8 is pretty much anything before Polaris and generally not supported
GFX 9 is VEGA and instinct cards they have the most support although they dropped Vega “support” they still work
GFX10 is less supported these are mainstream navi and Polaris, you can make them work with a force command

In terms of ease

Cuda just works
ROCm needs a bit of special attention such as hyper specific kernel
Intel I’ve never messed with yet but I believe it’ll take as much work if not more to get working as ROCm

1 Like
10

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.