Using Green Context in CUDA on Jetson Devices with Ampere Architecture

[juamaros]

Hello,

I’ve been looking into the new versions of CUDA and noticed the Green Context function for assigning specific SMs to each process.

My question is whether this technology is compatible with Jetson devices that have the Ampere chip architecture.

Knowing this, I was also wondering if this technology could be used to simulate MIG functionality on these devices, since they are not natively compatible.

Thanks a lot, and I look forward to your response.

[Greg]

Green Context is within the primary context. The GR (graphics engine that includes the compute pipe) can only execute one context at a time. By default a CUDA context is owned by a single process.

Multi-Process Server (MPS) allows multiple processes to execute in a single GPU context. MPS server to my knowledge is not supported on Jetson.

Assigning work to a Green Context is used to control resource assignment to multiple streams of work in the same process unless using with MPS. By using Green Context on Jetson I don’t think you are able to simultaneously execute multiple processes on the GR engine.

Moreover, Green Context is a work distribution scheme. In the case of MIG the GPU has multiple-instance of the compute pipe and the hardware can be physically partitioned into gpu_instances and compute_instances providing QoS for most units through physical partitioning. Green Context is a SW mechanism for partitioning resources but it does not provide the same process isolation and quality of service guarantees.

[juamaros]

Thank you very much, I will try testing with other alternatives.

Best regards.