F5F Stay Refreshed Hardware Desktop Can I use PyTorch with Lunar Lake's iGPU and NPU?

Can I use PyTorch with Lunar Lake's iGPU and NPU?

Can I use PyTorch with Lunar Lake's iGPU and NPU?

Z
zeliotL
Member
211
07-21-2021, 11:11 PM
#1
The iGPU at Lunar Lake delivers a minimum of 53 TOPS. Including the NPU brings the performance up to around 93, nearly matching a RTX 3060. With that level, I’m curious about running a machine learning framework such as PyTorch.
Z
zeliotL
07-21-2021, 11:11 PM #1

The iGPU at Lunar Lake delivers a minimum of 53 TOPS. Including the NPU brings the performance up to around 93, nearly matching a RTX 3060. With that level, I’m curious about running a machine learning framework such as PyTorch.

A
audi497mks
Senior Member
601
07-29-2021, 02:02 AM
#2
TOPS isn't related to running frameworks; it focuses on hardware capabilities. It's not suitable for training models with INT4. The NPU is designed for inference, not for training. Comparing these to a 3060 isn't feasible due to significant differences in the components. You won't be able to use both the iGPU and NPU at the same time. Intel offers their own PyTorch extension that supports their GPUs (including iGPUs): https://github.com/intel/intel-extension-for-pytorch They also provide a library for interacting with the NPU, enabling conversion of some PyTorch models to run on it (note most won't work): https://github.com/intel/intel-npu-acceleration-library Most of these solutions don't function well on Windows.
A
audi497mks
07-29-2021, 02:02 AM #2

TOPS isn't related to running frameworks; it focuses on hardware capabilities. It's not suitable for training models with INT4. The NPU is designed for inference, not for training. Comparing these to a 3060 isn't feasible due to significant differences in the components. You won't be able to use both the iGPU and NPU at the same time. Intel offers their own PyTorch extension that supports their GPUs (including iGPUs): https://github.com/intel/intel-extension-for-pytorch They also provide a library for interacting with the NPU, enabling conversion of some PyTorch models to run on it (note most won't work): https://github.com/intel/intel-npu-acceleration-library Most of these solutions don't function well on Windows.

S
SubsNewgab
Junior Member
21
07-29-2021, 03:24 AM
#3
You're not relying on Windows, are you? Then building a 3060 PC sounds like a solid plan!
S
SubsNewgab
07-29-2021, 03:24 AM #3

You're not relying on Windows, are you? Then building a 3060 PC sounds like a solid plan!

F
Fugo612
Junior Member
10
07-29-2021, 11:55 PM
#4
No problem at all. I don’t know what you’re aiming for.
F
Fugo612
07-29-2021, 11:55 PM #4

No problem at all. I don’t know what you’re aiming for.

B
BlackElemental
Junior Member
8
07-30-2021, 12:26 PM
#5
Sure, I understand. I was considering purchasing a laptop with at least 93 TOPS, but the 53 is significantly lower than a 3050 with 6GB RAM, making the comparison difficult.
B
BlackElemental
07-30-2021, 12:26 PM #5

Sure, I understand. I was considering purchasing a laptop with at least 93 TOPS, but the 53 is significantly lower than a 3050 with 6GB RAM, making the comparison difficult.

W
wedcake
Junior Member
35
07-30-2021, 02:55 PM
#6
Great! I’m still unclear about your goal with the 3060. Training and inference are distinct processes, so more context would help.
W
wedcake
07-30-2021, 02:55 PM #6

Great! I’m still unclear about your goal with the 3060. Training and inference are distinct processes, so more context would help.

S
StreetHobo
Senior Member
568
08-04-2021, 04:21 PM
#7
I plan to combine both training and inference, focusing more on training.
S
StreetHobo
08-04-2021, 04:21 PM #7

I plan to combine both training and inference, focusing more on training.

A
Ayisfly
Junior Member
11
08-04-2021, 07:57 PM
#8
This situation really supports using a dedicated GPU with sufficient VRAM for your models, right?
A
Ayisfly
08-04-2021, 07:57 PM #8

This situation really supports using a dedicated GPU with sufficient VRAM for your models, right?