Juice

Juice, by Juice Labs, introduces a revolutionary computing paradigm: GPU-over-IP. It is a sophisticated software solution designed to create a new computing substrate where accelerated compute power flows over networks, much like electricity over a power grid. This technology fundamentally changes how organizations utilize and manage their expensive GPU resources. By decoupling GPUs from their physical hosts, Juice allows any application on any machine to tap into GPU power from anywhere on the network—be it on-premises, in the cloud, or in a hybrid environment.

The platform is engineered for seamless integration into existing infrastructures. It requires no changes to your applications, hardware, or underlying operating systems. Juice works by virtualizing GPU resources and making them available over standard IP networks. This means a data scientist's workstation or a CPU-only server can instantly attach to a powerful remote GPU to run a demanding AI model, and then release it just as quickly, making the resource available for the next workload. This dynamic allocation eliminates GPU idling, dramatically increases utilization rates, and maximizes the return on investment in high-performance computing hardware.

How to use Juice

Integrating Juice into your workflow is designed to be straightforward and non-disruptive. Here’s a typical process:

1. Deployment: Juice is a software-only solution. You deploy the Juice agent on the machines that host the GPUs (GPU servers) and on the machines that need to access them (client nodes). This can be done on bare metal, in virtual machines, or within containers across Windows and Linux environments.
2. Access: Users can access the remote, virtualized GPUs through several interfaces. The most common method is via the Command Line Interface (CLI) by simply running a command like pip install juice. A user-friendly desktop application is also available for Windows, and an SDK is in development for deeper, native application integration.
3. Orchestration: Administrators can group physical GPUs into logical pools. They can then set access policies to control which users or systems can access specific pools. Using tags, workloads can be intelligently matched to the most appropriate GPU resources based on requirements like VRAM size or GPU model. This entire orchestration layer operates without the complexity of Kubernetes, although it can coexist with it.
4. Execution: Once connected, you run your applications as you normally would. Juice intercepts the GPU API calls (like CUDA, DirectX, OpenGL) from your application and routes them over the network to the remote GPU. The computation happens on the remote GPU, and the results are sent back. To the application, it appears as if a local GPU is present.

Core Features of Juice

• GPU-over-IP Technology: The core innovation that enables remote GPU access over standard networks with near-native performance.
• Dynamic Pooling & Sharing: Aggregate GPUs from across your entire infrastructure into centralized pools for on-demand allocation, eliminating silos and underutilization.
• GPU Fractionation: Split a single physical GPU into multiple virtual GPUs that can be used by different applications or users simultaneously, further boosting efficiency.
• No Code or Hardware Changes: Drops into your existing environment without requiring any modifications to your application code, drivers, or physical hardware setup.
• Cross-Platform & Cloud-Agnostic: Supports Windows and Linux, and can run on-prem, in any cloud, or bridge multiple environments (hybrid/multi-cloud).
• Broad API Compatibility: Fully compatible with major machine learning frameworks (PyTorch, TensorFlow, ONNX) via the CUDA API, and graphics APIs (DirectX, OpenGL, Vulkan).
• Simplified Orchestration: Manage your entire GPU fleet, set access controls, and monitor usage through intuitive tools without needing Kubernetes.

Use Cases for Juice

Juice is being adopted by leading organizations to solve critical infrastructure challenges:

• Academic Supercomputing: Top technical universities use Juice for dynamic GPU scheduling and pooling in superclusters, allowing researchers to share resources efficiently without complex orchestration.
• Cloud Service Providers: Digital infrastructure companies leverage Juice to route GPU capacity from regions of abundance to regions of scarcity, effectively creating a global, fungible GPU cloud.
• Enterprise AI & Data Science: Fortune 100 companies empower their data science teams to share a central pool of GPUs, allowing them to run AI/ML workloads dynamically from their own workstations or CPU servers.
• Game Development & VFX: Studios like Ubisoft use Juice to connect developers to remote GPUs, enabling massive cost savings and innovative new development workflows by sharing high-end graphics resources.

Advantages of Juice

The primary advantage of Juice is a massive increase in efficiency. By achieving GPU utilization rates far beyond the industry average, it delivers significant cost savings. Its flexibility allows organizations to adapt their compute resources dynamically to changing needs. The ability to turn any CPU-only node into a GPU node on the fly democratizes access to accelerated computing. Furthermore, its high-performance, low-latency connection ensures that even demanding workloads run smoothly, as proven by benchmarks with models like LLaMa, BERT, and YOLO.

Pricing and Plans

Juice offers a tiered pricing model to suit different needs:

• Juice Free: $0/Month. Ideal for individuals or small-scale testing. Includes up to 10 hours of total GPU time per month, pooling of up to 3 GPU agents, and support for Nvidia GPUs with up to 40GB of VRAM.
• Juice Pro: $30/Month. Designed for professionals and small teams. Offers unlimited GPU time per month, pooling of up to 3 GPU agents, and support for Nvidia GPUs with up to 40GB of VRAM.
• Juice Enterprise: Custom Pricing. A tailored solution for large organizations. Provides unlimited GPU time, unlimited GPU agents, and support for any size of Nvidia GPU, along with dedicated support and advanced features.