Nvidia Jetson Nano

#1

Juilia already runs on the Raspberry Pi. Would this be a good system for prototyping Flux.jl applications?
Maybe a bit short on memory.

4 Likes
#2

My Jetson Nano arrived this morning. ARM 8 64 bit architecture.
Master version of Julia has been compiling all day!

1 Like
#3

If prototyping is recognized to be the phase where developer time is more important than hardware cost, then I guess

answers your question.

#4

Does the prebuilt AArch64 binary work?

#5

It is ideal for low-cost DIY projects :slight_smile:

image

“For the price, the Jetson Nano TensorRT inference performance is looking very good.”
“The Jetson Nano did come out much faster than the ODROID-XU4 for the multi-threaded Rust benchmarks.”

“Overall this is arguably the best sub-$100 Arm developer board we’ve seen to date depending upon your use-cases. The Jetson Nano will certainly open up NVIDIA Tegra SoCs to appearing in more low-cost DIY projects and other hobbyist use-cases as well as opening up GPU/CUDA acceleration that until now has really not been possible on the low cost boards.”

#6

Simon, I built version 1.2.0-DEV from GitHub yesterday. The build worked perfectly and we tested CUarrays yesterday.
I do not have the board at work today. I will report back regarding the prebuilt AArch64 binary this evening UK time. I See no reason why it would not work.

2 Likes
#7

I confirm the Jetson Nano runs the AARCH64 build for the Raspberry PI
Version 1.0.3 TLS downloaded from the Julialang.org site.

Can anyone suggest benchmarks or tests I could run?

5 Likes
#8

If you are still looking for a benchmark, maybe you can test the speed of inference of the pre-trained vgg19 model (https://github.com/FluxML/Metalhead.jl) ?

Something as mundane as a matrix multiplication would already be quite interesting to me:

using CuArrays
using BenchmarkTools

A = randn(Float32,1000,10000); cuA = CuArray(A)
B = @btime Array(cuA*cuA');

(I get 3.214 ms on a GeForce GTX 1080, but the GPU is actually busy with other work too)

#9

CuArray test Note 128 core GPU onboard
A is 64x64 array Time 426 usec
A is 128x128 array Time 644 usec
A is 256x256 array Time 2.538 msec
A is 512x512 array Time 4.262 msec
A is 1024x1024 array Time 15.633 msec

A is 1000x1000 array Time 15.703 msec

2 Likes
#10

I have access to a boat load of rackmount GPU servers if you would like some specific tests run :slight_smile:

Sadly I have Julia 1.1.0 installed on my pet 1080 server - and CUArrays has a conflict with the SIUnits package. Grrr…

#11

Just got mine yesterday, glad to know that Julia will run

#12

One question, the prebuild Flux model vgg19 would not run as it seems to need more than 4Gbytes of RAM.
Is there an easy way to estimate how much RAM a model will consume?