ARM as desktop PC

[mictadlo]

Hi,

I would like to replace my MacBook Air 2015 with some ARM processor and at least 32 GB of RAM. I would like to do photo and 4K editing. Furthermore, I would like to get into machine learning.

 

I found Jetson AGX Xavier Developer Kit (

https://developer.nvidia.com/embedded/jetson-agx-xavier-developer-kit ).However, I could not find any case for it.

 

Is there any better SBC for the above tasks?

 

I appreciate any help you can provide. 

 

Mic

[Vishera]

Why do you need an ARM processor?

Do you realize that only Linux and few others will work on that machine?

[S.Hamed23]

16 minutes ago, mictadlo said:

Hi,

I would like to replace my MacBook Air 2015 with some ARM processor and at least 32 GB of RAM. I would like to do photo and 4K editing. Furthermore, I would like to get into machine learning.

 

I found Jetson AGX Xavier Developer Kit (

https://developer.nvidia.com/embedded/jetson-agx-xavier-developer-kit ).However, I could not find any case for it.

 

Is there any better SBC for the above tasks?

 

I appreciate any help you can provide. 

 

Mic

get a g14/g15 2021

[SorryBella]

15 minutes ago, mictadlo said:

I would like to do photo and 4K editing. Furthermore, I would like to get into machine learning.

> Photoshop on ARM

> Premiere Pro on ARM

Terrible experience period, unless you can put up with macs. If you want to run windows, get an x86 computers and laptops.

[Vishera]

5 minutes ago, SorryClaire said:

> Photoshop on ARM

You can use Gimp

5 minutes ago, SorryClaire said:

> Premiere Pro on ARM

You can use Kdenlive

[Jurrunio]

Get an iMac with Apple M1 then. Apple's ecosystem is the most optimized for ARM and by a long shot at that.

[SorryBella]

Just now, Vishera said:

You can use Gimp

You can use Kdenlive

You can use alternative, or you can use industry trend so you can actually collaborate with people instead of "being hip".

[Vishera]

3 minutes ago, Jurrunio said:

Get an iMac with Apple M1 then. Apple's ecosystem is the most optimized for ARM and by a long shot at that.

Seems like OP wants to use the machine learning capabilities of the tensor cores.

2 minutes ago, SorryClaire said:

You can use alternative, or you can use industry trend so you can actually collaborate with people instead of "being hip".

Yep,That's a compromise if you work with other people and they use something else.

[Jurrunio]

19 minutes ago, Vishera said:

Seems like OP wants to use the machine learning capabilities of the tensor cores.

ARM just doesnt come with high performance GPUs for that. The Jetson AGX Xavier he mentioned has a 512 core.Volta GPU, in other words only 1/10 of a Titan V, in a package rated for 30w (the whole box) so it cant clock nearly as high. It's still very slow compared to proper machine learning builds on X86.

 

It depends on @mictadlo's priority here. I cant tell since he said to little for me to guess his underlying reason to go for ARM specifically. Does he have to save power or just cant use X86 for his software? Or just watched some "ARM is the future" video and is all over the ARM bandwagon.

[Vishera]

5 minutes ago, Jurrunio said:

ARM just doesnt come with high performance GPUs for that. The Jetson AGX Xavier he mentioned has a 512 core.Volta GPU, in other words only 1/10 of a Titan V, in a package rated for 30w (the whole box) so it cant clock nearly as high. It's still very slow compared to proper machine learning builds on X86.

 

It depends on @mictadlo's priority here. I cant tell since he said to little for me to guess his underlying reason to go for ARM specifically. Does he have to save power or just cant use X86 for his software? Or just watched some "ARM is the future" video and is all over the ARM bandwagon.

You are right,We need more info from OP.

[AMD A10-9600P]

If you want to specifically go for ARM, the only reasonable option for a consumer machine right now is Apple's M1 chip.

[Vishera]

Just now, AMD A10-9600P said:

If you want to specifically go for ARM, the only reasonable option for a consumer machine right now is Apple's M1 chip.

But OP wants to do machine learning with tensor cores.

[AMD A10-9600P]

2 minutes ago, Vishera said:

But OP wants to do machine learning with tensor cores.

OP also said they wanted to do photo and video editing, for those workloads, the only viable option for that is the M1. They won't get both Tensor cores for machine learning and an ARM chip capable of runner consumer software in the same package.

[SorryBella]

18 minutes ago, Jurrunio said:

Or just watched some "ARM is the future" video and is all over the ARM bandwagon.

I mean i get ya wanna keep it respectful, but its DEFINITELY this one.

2 minutes ago, Vishera said:

But OP wants to do machine learning with tensor cores.

And? If he REALLY wanna do that then he shouldve gone with a volta/turing/ampere based GPUs instead. Or atleast just use standard OpenCL/CUDA to accelerate the learning. Its shit, i know, but its better than none.

[Vishera]

1 hour ago, SorryClaire said:

And? If he REALLY wanna do that then he shouldve gone with a volta/turing/ampere based GPUs instead.

The ARM based machine OP suggested to buy,has a Volta GPU in it,It has an NVIDIA SOC.

[Lenovich]

2 hours ago, Vishera said:

You can use Gimp

Sorry mate. No one can use Gimp for photo edit. Photoshop, lightroom apps for mobile phones provide more advanced tools.

[mictadlo]

Thank you for all responses. For my next workstation/laptop, I wanted something that does not use a lot of electricity. Therefore, I thought ARM would be a good way to go.

 

Furthermore, I don't mind running Linux. In the past, I used Darktable, similar to lightroom. 

There is a possibility that for machine learning, I would need a Unix like operating system.

 

Therefore, the below laptops might be not suitable for it unless  I could do dual boot:
* The G14 appears to be loud (

 

*Asus ProArt StudioBook seems not to come in a 13 or 14-inch display. 

I found TensorBook ( https://lambdalabs.com/deep-learning/laptops/tensorbook/customize ), which ships as default with LINUX but it is quite pricely.

HYBRID GRAPHICS: System76 (  https://system76.com/laptops/oryx ) Laptops get a boost of power when you need it or save battery when you don’t. Pop!_OS allows you to toggle between Intel and NVIDIA graphics, and even launch specific applications using your GPU. Is it only available for System 76 laptops? If not would it work with any volta/turing/ampere GPU. 

 

I was looking forward to Apple's M1, but unfortunately, I was disappointed because a maximum of 16 GB of RAM is supported. However, there are rumours of M2 or M1x for July.

 

The problem of missing Tensor Cores in MAC could be solved by edge TPU ( 
https://cloud.google.com/edge-tpu/ ) which could be plug-in via USB-C. What I have not figure out yet is wether it possible to train your own Models on edge-TPU or it necessary to use google cloud for it? Furthermore, whether it would be possible to use multiple edge-TPU in parallel?

[igormp]

9 hours ago, mictadlo said:

Is it only available for System 76 laptops?

No, it works in any laptop with both a Nvidia GPU and integrated graphics.

9 hours ago, mictadlo said:

I was looking forward to Apple's M1, but unfortunately

If you're planning on doing machine learning, the M1 is not up to the task anyway.

9 hours ago, mictadlo said:

The problem of missing Tensor Cores in MAC could be solved by edge TPU ( 
https://cloud.google.com/edge-tpu/ ) which could be plug-in via USB-C

The edge TPU is meant for inference, and its performance for training is on par with modern laptop CPUs (and behind desktop CPUs), so there's no reason to use it outside of embedded scenarios.

9 hours ago, mictadlo said:

What I have not figure out yet is wether it possible to train your own Models on edge-TPU or it necessary to use google cloud for it? Furthermore, whether it would be possible to use multiple edge-TPU in parallel?

Edge TPU is what you use on edge devices (such as raspberries and other small SBCs) for inference. The Cloud TPU is only available at google cloud and is the one meant for training models.