NextG Systems

AI infrastructure

Sharing GPUs and clusters safely between tenants: isolation, scheduling and the systems underneath.

  1. A tensor accelerator in PolarFire fabric, 81× over the board's own CPUs

    Apache TVM's VTA running in PolarFire SoC fabric: 49 GOPS across ResNet-18's convolution layers at 84 % of the array's theoretical peak, driven from Linux on the RISC-V cores with no kernel module. Plus the one PolarFire-specific RTL requirement that porting VTA depends on.

  2. Slicing an NVIDIA GPU between tenants who don't trust each other

    How we divide one NVIDIA GPU's memory and compute between gVisor sandboxes: quotas in the Sentry, a broker in the open kernel driver, and a credit scheduler that a tenant can't game by forking.