(Credit: Joseph Maldonado/PCMag)
Don't miss out on our latest stories. Add PCMag as a preferred source on Google.
A small AI developer has shown an M3 MacBook Pro using an externally connected Nvidia RTX graphics card for AI calculations. This is not ideal for gaming or even running a display, but it potentially opens the door for AI development on Apple hardware.
Since Apple transitioned from Intel processors in its MacBooks to ARM for its own Apple M-series chips, which feature onboard GPUs, compatibility with Nvidia and AMD GPUs has been nonexistent. As TechRadar points out, developers and enthusiasts have attempted to add support to no avail. Now, though, AI developer TinyCorp has shown it's possible.
TinyCorp previously developed a method for running an external AMD graphics card on Apple silicon using USB 3, but it has now done the same with Nvidia GPUs on the M-Series MacBooks, leveraging USB 4 and Thunderbolt 4. It hasn't given the full details of how it achieved this, but it showcased an image of it in action.
The configuration reportedly works with Nvidia RTX 30, 40, and 50-series GPUs, as well as AMD GPUs like RDNA2, RDNA3, and RDNA4, according to TinyCorp. However, you'll get the best performance from the later models, which feature the latest Tensor Cores and greater quantities of onboard memory. With such a card and the onboard MacBook processing capabilities, developers would be able to run larger language models locally, rather than relying exclusively on cloud deployments of AI, which introduce privacy and latency concerns.
For home developers looking to experiment with AI training and model fine-tuning, this is a significant development that could make Apple hardware more relevant in the developer space. As it stands, although Apple's Neural Engine is capable, it's still very limited compared with the power of a discrete GPU, especially Nvidia's high-end RTX 50-series cards, like the RTX 5090. Having access to that kind of power gives a MacBook an enormous amount of additional compute power to handle AI workloads.
Still, the process would need to be better understood and the made drivers publicly available if this is going to see any kind of wider adoption. It also currently relies on the TinyCorp-developed TinyGrad framework, which may limit future deployment and development.
The proof of concept is impressive and exciting, though, and shows that developers are likely to have a wealth of options for AI development hardware in the future.