Hi, my name is Joshua-Hicks, and I'm the founder of Hawking Systems.
Hawking has built a native NVIDIA GPU execution stack for Apple Silicon, including a custom DriverKit driver, compiled CUDA kernels, and a model-serving backend. Using an RTX 4000 Ada connected over Thunderbolt, execution of Qwen3 models from 0.6B to 8B parameters have successfully run directly on the NVIDIA card.
Apparently, getting this to work is easier than getting Apple to respond to my requests for distribution (I wonder why). So I'm releasing the runtime as an open source library and reworked this release to limit my dependence on that specific dependency.
Beyond GPU execution, Hawking is working on a C-native harness and experiments in low-bit representation (and hybrid architecture) of an extremely popular 27-billion-parameter model. One of Hawking’s goals is a deeper level of control/specialization and performance compared to competitors like TinyGrad. Another is striving to adopt the cloud experience locally and ameliorate it.
v0.1.1 should be up in a week or two but I would love for you to investigate and customize this offering in its current state.
If you would like to follow Hawking’s progress I would like to say welcome!
P.S. Please contact me if you have a different eGPU setup and would like to help build this library's variety of support outside of offline simulation and RunPod probes.
joshuahicks•12m ago
Hawking has built a native NVIDIA GPU execution stack for Apple Silicon, including a custom DriverKit driver, compiled CUDA kernels, and a model-serving backend. Using an RTX 4000 Ada connected over Thunderbolt, execution of Qwen3 models from 0.6B to 8B parameters have successfully run directly on the NVIDIA card.
Apparently, getting this to work is easier than getting Apple to respond to my requests for distribution (I wonder why). So I'm releasing the runtime as an open source library and reworked this release to limit my dependence on that specific dependency.
Beyond GPU execution, Hawking is working on a C-native harness and experiments in low-bit representation (and hybrid architecture) of an extremely popular 27-billion-parameter model. One of Hawking’s goals is a deeper level of control/specialization and performance compared to competitors like TinyGrad. Another is striving to adopt the cloud experience locally and ameliorate it.
v0.1.1 should be up in a week or two but I would love for you to investigate and customize this offering in its current state.
If you would like to follow Hawking’s progress I would like to say welcome!
https://x.com/hawkingsystems https://hawking.systems/
P.S. Please contact me if you have a different eGPU setup and would like to help build this library's variety of support outside of offline simulation and RunPod probes.