Can I chain additional GPU resources to a Jetson XAVIER to get some better performance?

[danny.moore]

We have a POC we are building using our Gen4 PCIe Switch for GTC. We would like to show our ability to tie multiple NVIDIA resources together.

My understanding is that in a topology where one JETSON is an endpoint to another JETSON, the RC-JETSON will not have any access to the Endpoint-JETSON. So in such an scenerio we would not gain any additional performance.

However, what if I was able to add an additional GPU resource to the JETSON host? Could JETSON detect and utilize these additional resources during video inference?

[dusty_nv]

My understanding is that in a topology where one JETSON is an endpoint to another JETSON, the RC-JETSON will not have any access to the Endpoint-JETSON. So in such an scenerio we would not gain any additional performance.

The two Jetson’s would be able to DMA memory to each other, but they would not be able to directly control the other’s GPU. Each Xavier controls it’s own GPU. Distributing the computational workload would be left up to the user’s application or a multiprocessing framework like MPI.

However, what if I was able to add an additional GPU resource to the JETSON host? Could JETSON detect and utilize these additional resources during video inference?

In a future release of JetPack, you will be able to plug an NVIDIA discrete GPU card into the PCIe slot. Please refer to this FAQ: https://developer.nvidia.com/embedded/faq#devkit-gpus

[SammoGan]

@dusty_nv

My understanding is that in a topology where one JETSON is an endpoint to another JETSON, the RC-JETSON will not have any access to the Endpoint-JETSON. So in such an scenerio we would not gain any additional performance.

The two Jetson’s would be able to DMA memory to each other, but they would not be able to directly control the other’s GPU. Each Xavier controls it’s own GPU. Distributing the computational workload would be left up to the user’s application or a multiprocessing framework like MPI.

However, what if I was able to add an additional GPU resource to the JETSON host? Could JETSON detect and utilize these additional resources during video inference?

In a future release of JetPack, you will be able to plug an NVIDIA discrete GPU card into the PCIe slot. Please refer to this FAQ: https://developer.nvidia.com/embedded/faq#devkit-gpus

Hi,dusty_nv,Can JetPack4.2 support PCIe discrete GPU ?