Juice Remote GPU
We’re building software to enable AI and Graphics workloads to run on remote GPUs.
Our software enables offloading of GPU processing for any CUDA or Vulkan application to a remote host that’s running our agent. We inject our CUDA and Vulkan implementations during runtime, which means no code changes to applications. Our virtualization supports various "manyness" scenarios - e.g. we allow a single client to connect to multiple GPUs or multiple clients to share a single GPU.
Use Cases
You can use Juice to share a single GPU across several workstations, allocate GPUs dynamically to CPU-only machines (eg: for serverless functions), or simplify dev workflows and deployments.
Performance
Below is a benchmark running SDXL, comparing local GPU performance to a remote Juice GPU running on a local network.
We tested Juice Remote GPU with a variety of AI workloads and found that it performs within 5% of a local GPU when running in the same data center.
Compatibility
We support the following APIs:
| API | Version | OS |
|---|---|---|
| CUDA | 11.8 - 12.3 | Windows, Linux |
| Vulkan | 1.3 | Windows, Linux (server-only) |
| DirectX | 11, 12 | Windows, Linux (server-only) |
| OpenGL | 4.6 | Windows, Linux (server-only) |
We test our software with the latest versions of PyTorch and TensorFlow . We support AI inference and fine tuning with current LLMs and image diffusion models.
Looking for more information?
We’re happy to help with any questions. Reach out and we’ll get back to you as soon as we can.
About Us
We are a small team of top engineers from Meta, Intel, and gaming tech. @deanbeeler was a founding engineer at Oculus and built foundational VR technology like ASW. @stevegolik was a senior executive at Turnitin and Surveymonkey. @yuriykagan led VR system UI engineering at Meta.
@david-mccloskey worked on low level software and hardware for a decade at Intel. @charles-w-baker and @ramon-steenson brings years of experience in gaming, graphics, and data science applications to our team.
Partners
