Old Nuc Vs Pi 400

Not really a fan of Intel but find even a Pi struggles with more state of art TTS so had a looking at 2nd user NUCs as I am a fan of reuse.

Above is a Pi400 vs an old £55 NUC picked up off ebay.
The AVX instructions of the NUC i3-3217U do have the Pi400 beat and don’t get a celeron that is without as for NN frameworks such as Onnx or tensflow they add a big boost where as a general purpose processor the 1.8Ghz i3-3217U just has a small edge on a Pi400 (Pi4 1.8Ghz OC)

Its the Mini PCIe slots in the NUC where things get interesting as my recent experiments with a Jetson nano 2gb gave results where the 0.5Tops of the GPU was just a tad faster than the CPU but https://coral.ai/products/pcie-accelerator and 4Tops should be a big boost.

Hopefully maybe the Pi4A+ might come to the rescue with a PCIe lane as for offline use it still produces a sluggish brain.

The fan and heatsink got a clean and the thermal paste which took some cleaning from its concrete like state got replaced but maybe it will last another decade to the one its already had?
There are quite a number of 4/8gb with a 32/120gb SSD for £50-100 quid the above in the benchmark was £55 which is £20 less then a Pi4-8gb and comes with PSU and SSD so no need for additional PSU & SD.

Stay clear of the celerons as presume its the AVX of the I3 that give the Pi400 such a beating but even the old 1st gens with a bit of a refurb can make a decent brain.

Updated and added a Jetson nano which is surprising as cpu only and beats the pi400

Added a haswell Nuc as well the Jetson nano is on ebay as really didn’t like the bloated Ubuntu 18.04 image and had a lot of problems with 2gb that don’t seem to have with a Pi4.

Another benchmark using https://github.com/TensorSpeech/TensorFlowTTS as I think high quality TTS is the biggest load in a voice AI system.
Its not Larynx and think is faster but its just what I have been playing with of late and know how to install manually.

ARMv8 Cortex-A72 - vc4drmfb - BCM2835 Raspberry Pi (pi400)

INFO: TfLiteFlexDelegate delegate: 2 nodes delegated out of 964 nodes with 1 partitions.

--- 9.258699893951416 seconds ---
--- 7.028157711029053 seconds ---
--- 9.003497838973999 seconds ---
--- 6.985812187194824 seconds ---
--- 8.801349401473999 seconds ---
--- 7.019275426864624 seconds ---

1st resultant wav is 11.111 secs so x 1.234 realtime
2nd resultant wav is 8.87 secs so x 1.263 realtime

Intel Core i3-3217U - Intel 3rd Gen Core - Intel

INFO: Created TensorFlow Lite delegate for select TF ops.
INFO: TfLiteFlexDelegate delegate: 2 nodes delegated out of 964 nodes with 1 partitions.

--- 7.083249092102051 seconds ---
--- 5.474043369293213 seconds ---
--- 7.031285047531128 seconds ---
--- 5.50117039680481 seconds ---
--- 7.054577827453613 seconds ---
--- 5.499629735946655 seconds ---

1st resultant wav is 11.111 secs so x 1.575 realtime
2nd resultant wav is 8.87 secs so x 1.62 realtime

Intel Core i5-4250U - Intel Haswell-ULT IGP - Intel

INFO: Created TensorFlow Lite delegate for select TF ops.
INFO: TfLiteFlexDelegate delegate: 2 nodes delegated out of 964 nodes with 1 partitions.

--- 3.398526668548584 seconds ---
--- 2.6226658821105957 seconds ---
--- 3.322057008743286 seconds ---
--- 2.608353614807129 seconds ---
--- 3.3249073028564453 seconds ---
--- 2.601407289505005 seconds ---

1st resultant wav is 11.111 secs so x 3.341 realtime
2nd resultant wav is 8.87 secs so x 3.40 realtime

PS the Haswell nuc was £85 with a M$ wireless keyboard and the bump to 2.5Ghz and also think the Avx instructions in Haswell also got upgraded, which accounts for the perf increase.
The Jetson Nano 2gb that is not featured was just so hardwork to get up and running and then complained the GPU had no ram and is now on ebay.
The 2Gb with just a small amount allocated on a Pi4-2gb is no problem but the 2gb Jetson sharing memory with a Maxwell GPU is really hardwork and 4gb prob better but can get a 2nd user Nuc for less… !

Haven’t got the Coral Edge working with anything than recognising parrots exceptionally quick converting models is much more complex than I 1st thought.

wav1
https://drive.google.com/open?id=1MFZNT1lr2Q7aa5gggPg326ZzMSzBWmvs
wav2
https://drive.google.com/open?id=1XnZq3E7dn3nc-6n8xQXnqIhdmlouC5ga