IBM's library is 46x faster than TensorFlow
Moderators: hgm, Rebel, chrisw
-
- Posts: 362
- Joined: Thu Mar 16, 2006 7:39 pm
- Location: Portugal
- Full name: Alvaro Cardoso
-
- Posts: 433
- Joined: Fri Dec 16, 2016 11:04 am
- Location: France
- Full name: Richard Delorme
Re: IBM's library is 46x faster than TensorFlow
IBM machines are too expensive for me, humble amateur chess programmer, to even dare dreaming at. On the other hand, I wonder what is the performance of some affordable development boards with a NPU (Neural Processing Unit) like NVidia jetson tx2(1) ot Huawei HiKey 970 (2).
(1)https://developer.nvidia.com/embedded/buy/jetson-tx2
(2)https://www.xda-developers.com/huawei-l ... n-970-soc/
(1)https://developer.nvidia.com/embedded/buy/jetson-tx2
(2)https://www.xda-developers.com/huawei-l ... n-970-soc/
Richard Delorme
-
- Posts: 12540
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: IBM's library is 46x faster than TensorFlow
If you want the most power per watt, those might be nice.abulmo2 wrote:IBM machines are too expensive for me, humble amateur chess programmer, to even dare dreaming at. On the other hand, I wonder what is the performance of some affordable development boards with a NPU (Neural Processing Unit) like NVidia jetson tx2(1) ot Huawei HiKey 970 (2).
(1)https://developer.nvidia.com/embedded/buy/jetson-tx2
(2)https://www.xda-developers.com/huawei-l ... n-970-soc/
But if you want compute power per dollar, they are not good buys.
Pop a nice GPU in you home computer and you will smoke those units. The Jetson is (for instance) 1.5 TFlop. My 3 year old graphics card at home is 2 TFlop and you can buy an 8 Tflop card for the price of one of those mini-boards.
I just bought some GTX 960 cards at $86 each. Those are 2.3 TFlops
so:
86/2.3 = $37.39 per TFlop.
Dev Board:
$599/1.5 Tflops = $399 per tflop. Less than 10% as cost effective.
Students can get the boards for less.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: IBM's library is 46x faster than TensorFlow
Hehe, proud on my officemate .Cardoso wrote:https://www.theregister.co.uk/2018/03/2 ... ined_fast/
P.S. It can also be executed on Intel and Intel+GPU not only on POWER+GPU