Thunder ComputeCloud Computing Technology to virtualize GPUs over TCP.

Thunder Compute has created virtual GPU-over-TCP technology (we call it PCI-v) that enables a new type of cloud instance. We can instantly connect or disconnect any number of GPUs with one command, efficiently switch between tasks, and share a single GPU. This creates one cloud instance that accommodates any development workflow. To elaborate, this one flexible environment can start as an inexpensive CPU-only dev machine and instantly scale to become a massive cluster. Developers can directly access resources they need to eliminate time spent on bureaucratic approval processes and config. Additionally, this technology cuts cost for organizations in a few ways: 1) We can place multiple workloads on a GPU at the same time with dynamic memory sharing. This enables a system similar to vCPUs you see on AWS today, where every user has full access to the entire GPU while allowing us to use any spare capacity on the physical GPU for other users. 2) Developers can interact with CPU-only instances that share pooled GPU resources. This eliminates the concept of reservations and creates a system where a GPU is never tied to a single machine. Something interesting to note, this appears to be the most efficient technology that exists for serverless GPU computing. We have reached effectively instantaneous cold starts by decoupling the GPU environment from the physical GPU, allowing any GPU to instantly switch between any task without having to "cold start" by loading a new environment.