NVIDIA changes how GPUs are shared in Kubernetes

NVIDIA changes how GPUs are shared in Kubernetes. The first notebooks with RTX Spark hit the market on October 16, 2024.

The company introduced Memory-Enforced GPU Sharing technology in Kubernetes, which uses NVIDIA CUDA 13.4 to optimize resource usage. The technology allows AI workloads to use a fraction of a GPU, increasing efficiency. NVIDIA CUDA Tile automates the use of advanced hardware features. The RTX Spark is a superchip for notebooks and compact desktops with a Blackwell GPU and up to 128 GB of unified RAM. The chip integrates an ARM CPU with up to 20 cores, a Blackwell GPU with up to 6,144 CUDA Cores and up to 128 GB of unified RAM. The RTX Spark uses NVLink C2C technology for communication between the CPU and GPU, with bandwidth of up to 600 GB/s. It delivers up to 1 petaflop of performance in AI operations with NVFP4 precision.

Comparison of GPU compute-sharing modes. W1 and W2 alternate in a row along a time axis. SM-sharing: a grid where workloads A — Image: developer.nvidia.com

NVIDIA introduces Memory-Enforced GPU Sharing technology in Kubernetes, which allows AI workloads to use just a fraction of a GPU. The technology depends on a specific configuration for Kubernetes and requires CUDA 13.4, which may limit its adoption in less prepared environments. Users can configure the distribution of memory between CPU and GPU, increasing flexibility for specific tasks.

Specifications

Specifications
Item	Specification
Technology	Memory-Enforced GPU Sharing; another source: CUDA Tile (cuTile)
Processor	Blackwell GPU with up to 6,144 CUDA Cores; another source: ARM CPU with up to 20 cores
Connectivity	NVLink C2C with bandwidth of up to 600 GB/s

New technology optimizes GPU use in Kubernetes

The model is aimed at users who need AI performance and energy efficiency. The RTX Spark platform is available in notebooks and compact desktops, with prices starting at US$ 2,599. The Memory-Enforced GPU Sharing technology does not yet have a launch date in Brazil, but NVIDIA has already confirmed that the RTX Spark has been on the market since October 16, 2024. The new feature complements the coverage of the Kumo platform and the CUDA-X library, which also aim to optimize GPU use in cloud computing environments.

NVIDIA is changing how GPUs are shared in Kubernetes, allowing AI workloads to run efficiently on fractions of resources. This redefines how parallel computing is organized, making GPU use more flexible. The RTX Spark, integrated with this technology, runs NVLink C2C with 600 GB/s of bandwidth, enabling faster communication between CPU and GPU, which is essential for intensive AI workloads.

With information from developer.nvidia.com and canaltech.com.br.