MIG
A hardware-level GPU partitioning technology introduced with the NVIDIA Ampere architecture. It divides a single physical GPU into up to seven fully independent instances, each with dedicated memory, cache, and compute engines, providing isolation approaching that of separate physical cards.
Core Facts
Timeline (last 90 days)
MIG and the CUDA 13.4 memory enforcement mechanism can be used together, but they serve different purposes: MIG provides physical-level determinism and performance isolation, while memory enforcement draws VRAM boundaries between processes sharing the same physical device.
Modern scheduling systems can partition compute along two dimensions, time and space: in the spatial dimension, GPU memory partitioning (such as MIG, MPS) or multi-process sharing allows smaller-scale experiments to share the same card; in the temporal dimension, the scheduler exploits the peaks and troughs of the sampling and update phases in RL training to interleave tasks from different experiments, reducing GPU idle time.
Platforms such as Modal typically combine multiple approaches to achieve GPU isolation: enabling MIG on supported hardware and, for hardware that does not support it, doing finer-grained packing and preemptive reclamation at the scheduling layer.
NVIDIA's MIG technology provides near-native isolation by partitioning models such as the A100 into multiple independent instances at the hardware level, but not all GPU models support it and the partition granularity is fixed.
The H100 supports MIG technology, which can partition a single GPU into up to 7 independent instances, each with its own dedicated memory, cache, and compute units.
MIG instances have limited compute capability, and some instance specifications do not support full NVLink bandwidth, so they are better suited to multi-tenant or task-isolation scenarios rather than pursuing extreme performance for a single task.
Compute Capability is a version number NVIDIA defines for its GPU architectures, expressed as "major.minor" (e.g., 8.x for Ampere, 8.9 for Ada Lovelace, 12.0 for Blackwell), encoded directly in the GPU hardware identity and read by the driver via PCI configuration space and GPU registers; it does not depend on upper-layer software or platform labels and cannot be forged by editing product descriptions or the virtualization layer (MIG partitioning only affects the visibility of memory and compute units, not the architecture version number).
MIG has relatively coarse granularity and fixed partitioning schemes, making it unable to fully replace software-level time-slicing scheduling.
NVIDIA's MIG is only available on data-center-grade cards (such as A100 and H100) and is not supported on consumer-grade RTX cards.
Green Contexts share similar goals with MIG and MPS (improving GPU utilization and isolation) but differ in positioning: MIG is a hardware physical partition with the strongest isolation but lower flexibility, MPS targets multi-process sharing, and Green Contexts provides finer-grained programmatic control within a single process.
4 more timeline events
All Facts (18)
NVIDIA's MIG technology enables hardware-level resource isolation on high-end GPUs such as the A100/H100.
80%UnverifiedNVIDIA MIG仅支持A100/H100等高端卡且切分粒度有限，vGPU方案需要额外购买商业许可证
85%UnverifiedNVIDIA's MIG technology allows a single GPU to be partitioned into multiple independent instances, capable of raising resource utilization from the 30–50% of traditional approaches to higher levels.
70%UnverifiedMIG and the CUDA 13.4 memory enforcement mechanism can be used together, but they serve different purposes: MIG provides physical-level determinism and performance isolation, while memory enforcement draws VRAM boundaries between processes sharing the same physical device.
50%UnverifiedModern scheduling systems can partition compute along two dimensions, time and space: in the spatial dimension, GPU memory partitioning (such as MIG, MPS) or multi-process sharing allows smaller-scale experiments to share the same card; in the temporal dimension, the scheduler exploits the peaks and troughs of the sampling and update phases in RL training to interleave tasks from different experiments, reducing GPU idle time.
50%UnverifiedPlatforms such as Modal typically combine multiple approaches to achieve GPU isolation: enabling MIG on supported hardware and, for hardware that does not support it, doing finer-grained packing and preemptive reclamation at the scheduling layer.
50%UnverifiedNVIDIA's MIG technology provides near-native isolation by partitioning models such as the A100 into multiple independent instances at the hardware level, but not all GPU models support it and the partition granularity is fixed.
50%UnverifiedMIG instances have limited compute capability, and some instance specifications do not support full NVLink bandwidth, so they are better suited to multi-tenant or task-isolation scenarios rather than pursuing extreme performance for a single task.
50%UnverifiedThe H100 supports MIG technology, which can partition a single GPU into up to 7 independent instances, each with its own dedicated memory, cache, and compute units.
50%UnverifiedCompute Capability is a version number NVIDIA defines for its GPU architectures, expressed as "major.minor" (e.g., 8.x for Ampere, 8.9 for Ada Lovelace, 12.0 for Blackwell), encoded directly in the GPU hardware identity and read by the driver via PCI configuration space and GPU registers; it does not depend on upper-layer software or platform labels and cannot be forged by editing product descriptions or the virtualization layer (MIG partitioning only affects the visibility of memory and compute units, not the architecture version number).
50%UnverifiedMIG has relatively coarse granularity and fixed partitioning schemes, making it unable to fully replace software-level time-slicing scheduling.
50%UnverifiedNVIDIA's MIG is only available on data-center-grade cards (such as A100 and H100) and is not supported on consumer-grade RTX cards.
50%UnverifiedGreen Contexts share similar goals with MIG and MPS (improving GPU utilization and isolation) but differ in positioning: MIG is a hardware physical partition with the strongest isolation but lower flexibility, MPS targets multi-process sharing, and Green Contexts provides finer-grained programmatic control within a single process.
50%UnverifiedMore reliable methods for reducing GPU idle power consumption are setting a power limit via nvidia-smi, using MIG partitioning, or unloading the model when there are no requests, rather than relying on environment variables of uncertain origin.
50%UnverifiedNVIDIA introduced MIG technology in the Ampere architecture in 2020, allowing a single physical GPU to be partitioned into up to 7 independent instances.
50%UnverifiedKubernetes原生不具备感知GPU的能力，GPU设备插件以DaemonSet形式运行在每个GPU节点上，用户可通过nvidia.com/gpu声明申请GPU；GPU不支持超卖，除非使用MIG或MPS等虚拟化技术
50%UnverifiedGPU不支持超卖，一张卡要么整卡分配要么不分配，除非使用MIG或MPS等虚拟化技术
50%UnverifiedMIG是NVIDIA在Ampere架构（A100，2020年）中引入的特性，提供真正硬件级隔离
50%