This is a project I've been working on for a while now. Ultimately the core issue Mimir aims to solve is how disconnected GPU programming is.
For example in machine learning or scientific computation a lot of training or exploratory processes are performed on high-grade NVIDIA servers. However, when deploying an edge device is almost always a possible canidate but rarely considered due to key libraries being exclusively written using CUDA for NVIDIA GPUs.
If kernels were written for a high-level standard like Mimir then the same libraries can be used everywhere. Essentially the promise of the JVM for computational GPU programming, "Write once run everywhere". However, this promise isn't quite ready to be fulfilled since the only implemented runtime uses Vulkan, this means a couple thousand different consumer devices are supported but virtually no server GPUs are currently supported with only Vulkan.
Mimir is still very much a work in progress but I plan on building a machine learning library in Python akin to Pytorch where everything centers around Mimir for a unified acceleration approach.
I'm in desperate need of more contributors and would love any and all feedback. I'm currently an 18 year old Sophomore at university, this is my largest project by far and I'd love any and all support.
This is a project I've been working on for a while now. Ultimately the core issue Mimir aims to solve is how disconnected GPU programming is.
For example in machine learning or scientific computation a lot of training or exploratory processes are performed on high-grade NVIDIA servers. However, when deploying an edge device is almost always a possible canidate but rarely considered due to key libraries being exclusively written using CUDA for NVIDIA GPUs.
If kernels were written for a high-level standard like Mimir then the same libraries can be used everywhere. Essentially the promise of the JVM for computational GPU programming, "Write once run everywhere". However, this promise isn't quite ready to be fulfilled since the only implemented runtime uses Vulkan, this means a couple thousand different consumer devices are supported but virtually no server GPUs are currently supported with only Vulkan.
Mimir is still very much a work in progress but I plan on building a machine learning library in Python akin to Pytorch where everything centers around Mimir for a unified acceleration approach.
I'm in desperate need of more contributors and would love any and all feedback. I'm currently an 18 year old Sophomore at university, this is my largest project by far and I'd love any and all support.