Intel Arc A310 + AMD 5500 performance report

[ProsjektX]

Switched to a new low-cost AMD platform a few weeks back, and thought if anybody else would be interested, here are the inference speeds on Mobilenet and yolonas. The A310 performs really well and power consumption is quite low (I'd estimate around 5W or so based on a power plug). It is cool to touch the heat sink and the fan very rarely spins. 4 UHD HEVC streams are barely noticeable on the GPU video codec engine (~1% according to intel_gpu_top). The limiting factor for the number of streams will be the GPU memory. Each allocates around 0.5GB of GPU mem, from 4GB available, so about 8 UHD streams, depending. Tested some time back with a GTX 1050 also, but it got quite hot. Fyi, the dual video engine in the Arc Alchemist series are the same across all GPUs, so A310 and A770 performs the same. One advantage with the A310 is it pulls power only from the PCIe slot.

AMD 5500 (zen3 6+6 core) at $70 + Intel Arc A310 GPU at $80 (recent offer)
Overall CPU load with 4 streams about 3%.

Inference times:

Openvino CPU:
Mobilenet v2: 5.3ms
YoloNAS Frigate+ 320x320: 17.5ms
YoloNAS Frigate+ 640x640: 64ms

Openvino GPU:
Mobilenet v2: 4.0ms
YoloNAS Frigate+ 320x320: 7.5ms
YoloNAS Frigate+ 640x640: 12.9ms

[kadamik]

I am thinking of a very similar build, but I had a question for you. You have 4 UHD cameras, but what resolution is the Detect Stream? Also do you know what happens if you exceed the 8 UHD cameras you estimated would fit in memory?

[ProsjektX]

I use detect streams of 2x 720p and 2x 1080p. If exceeding or approaching memory limits, don't know.. not sure if there's a dynamic aspect to the memory allocation here either, sorry.

[NexGen-3D-Printing]

What encoding are you running? h264 or h265?

I seem to get 10 to 13ms with my A310 running 9 x 1440p @25fps in h265 and a detect resolution of 720p @ 10fps

The card indeed runs super cool and inference time is about on par with a coral TPU, not sure if my additional latency is due to encoding, fps or possibly my platform as I'm running an older system with dual Xeon's, but I'm liking this card, was a little concerned that I should have gotten the A380 instead.

[ProsjektX]

I assume your inference times are with a yolonas model, and openvino is configured for GPU (or AUTO)? Not entirely comparable with the smaller mobiledet model used for coral.

Either way, my streams are H265, but that shouldn't matter for the inference time stats. I saw slightly higher inference times, like yours, on an older platform using an AMD 2400G CPU with the same RAM and RAM speed. PCIe bus for both 3.0 x8. Maybe there are parts of the openvino detector that runs more efficient on newer CPUs with new instruction sets..?

btw, resizable BAR is supported and enabled? for more effecient data transfers