Intel Loihi

[NumLock21]

Intel has just announced their first neuromorphic processor called the Loihi. Intel Loihi features, 

Quote

 

• Fully asynchronous neuromorphic many core mesh that supports a wide range of sparse, hierarchical and recurrent neural network topologies with each neuron capable of communicating with thousands of other neurons.
• Each neuromorphic core includes a learning engine that can be programmed to adapt network parameters during operation, supporting supervised, unsupervised, reinforcement and other learning paradigms.
• Fabrication on Intel’s 14 nm process technology.
• A total of 130,000 neurons and 130 million synapses.
• Development and testing of several algorithms with high algorithmic efficiency for problems including path planning, constraint satisfaction, sparse coding, dictionary learning, and dynamic pattern learning and adaptation.

 

Quote

 

The Loihi research test chip includes digital circuits that mimic the brain’s basic mechanics, making machine learning faster and more efficient while requiring lower compute power. Neuromorphic chip models draw inspiration from how neurons communicate and learn, using spikes and plastic synapses that can be modulated based on timing. This could help computers self-organize and make decisions based on patterns and associations.

The Loihi test chip offers highly flexible on-chip learning and combines training and inference on a single chip. This allows machines to be autonomous and to adapt in real time instead of waiting for the next update from the cloud. Researchers have demonstrated learning at a rate that is a 1 million times improvement compared with other typical spiking neural nets as measured by total operations to achieve a given accuracy when solving MNIST digit recognition problems. Compared to technologies such as convolutional neural networks and deep learning neural networks, the Loihi test chip uses many fewer resources on the same task.

Further, it is up to 1,000 times more energy-efficient than general purpose computing required for typical training systems.

 

Selected universities and research institutions will get to try them out in the first half of 2018, and it probably won't be a few more years until consumers can get their hands on one.

 

Spoiler

 

https://newsroom.intel.com/editorials/intels-new-self-learning-chip-promises-accelerate-artificial-intelligence/

 

[themctipers]

ah shit i thought this was after icelake or some shit 

 

ivy -> devils cannon / haswell -> skylake = kaby lake = coffee lake +2 cores = cannon lake but for mobile only (and 10nm) 

[NumLock21]

4 minutes ago, CUDA_Cores said:

here it comes...

 

 

But can it run duck life 4?!

Does not run it, it controls it.

[Damascus]

Thought this was another weird post about loli/lolicon, I'm a bit disappointed it's a Herald of the end times instead. 

[NumLock21]

4 minutes ago, CUDA_Cores said:

Duck life 4 seemed slightly more entertaining than that overused joke so I changed it. 

Can't Duck life 4, when there is no Duck Life 3.

 

[Taf the Ghost]

The funny bit about all of the AI & Machine Learning stuff is it let's marketing spew forth a blizzard of buzzwords.

 

So it's really just a semi-custom/refined FPGA? Though it should be, for the specific tasks, something of a big cost saver.

[Ryujin2003]

4 hours ago, VegetableStu said:

*squints* core... i5.. processor...???

No, it's a delidded Xeon! /s

[Sniperfox47]

5 hours ago, NumLock21 said:

~snip~

Looks interesting. Can't wait to see how this compares to TrueNorth. We may be finally at the stage where we see some innovation in consumer level Microarchitectures again!

5 hours ago, Taf the Ghost said:

The funny bit about all of the AI & Machine Learning stuff is it let's marketing spew forth a blizzard of buzzwords.

 

So it's really just a semi-custom/refined FPGA? Though it should be, for the specific tasks, something of a big cost saver.

What buzzwords are you talking about? Neuromorphic chips have been a field of research since the 1980s and it's a very specific style of processor design.

[Notional]

Sweet. But can it think up Crysis?

[BuckGup]

So we got Tensor cores and now this. Oh man self aware AI is going to happen in 2018

[NumLock21]

2 hours ago, Ryujin2003 said:

No, it's a delidded Xeon! /s

No it's definitely a core i5. Need to squint harder. 

[The Benjamins]

15 minutes ago, NumLock21 said:

No it's definitely a core i5. Need to squint harder. 

I thought it was a sandy bridge xeon clocked to 2.5Ghz, confirmed by LTT User.

[Guest]

Studied this a bit in my masters program. Cool seeing follow up news/stories from where the research was even a couple of years ago. 

[NumLock21]

1 hour ago, The Benjamins said:

I thought it was a sandy bridge xeon clocked to 2.5Ghz, confirmed by LTT User.

Nope it's actually a core i5. Not a xeon.

[The Benjamins]

2 minutes ago, NumLock21 said:

Nope it's actually a core i5. Not a xeon.

I was making a joke in reference to what Patrick said the ENG samples of zen were.

[NumLock21]

Just now, The Benjamins said:

I was making a joke in reference to what Patrick said the ENG samples of zen were.

And i was telling you that neuromorphic processor is actually a core i5. Just look at pic. 

[Stefan1024]

An other big player jumps onto the neuronal network hype tran. So we have Google, ARM, nvidia, AMD, Apple and a dozen start ups. Looks like the entire industry thinks this is the next "big thing" in computing.

 

While it's powerful for some specific problems, you can't do everthing with NN or there is a much better and faster solutions than a NN.

[Sniperfox47]

41 minutes ago, Stefan1024 said:

An other big player jumps onto the neuronal network hype tran. So we have Google, ARM, nvidia, AMD, Apple and a dozen start ups. Looks like the entire industry thinks this is the next "big thing" in computing.

 

While it's powerful for some specific problems, you can't do everthing with NN or there is a much better and faster solutions than a NN.

I think you misunderstood. Neuromorphic chips are significantly different than a traditional machine learning "neural net" though, even if the result is the same.

 

They basically do what a Neural Net does in software in hardware instead and are far *far* more efficient for inferencing, consuming hundreds of times less power than a GPU/TPU at only 100ish milliwatts. For comparison, a low power arm chipset can be 2.5-5W

 

The only two companies who have semi-finalized and public neuromorphic chip designs afaik at the moment are Intel with their new Loihi chip and IBM with their slightly older TrueNorth chip.

 

Neural Net inferencing is definitely the future, and making sure that devices can do it even without an internet connection is a big push right now. You're right that it doesn't solve every problem, but a lot of problems that were previously intractable can be handled quite neatly by machine learning.

[Stefan1024]

49 minutes ago, Sniperfox47 said:

I think you misunderstood. Neuromorphic chips are significantly different than a traditional machine learning "neural net" though, even if the result is the same.

 

They basically do what a Neural Net does in software in hardware instead and are far *far* more efficient for inferencing, consuming hundreds of times less power than a GPU/TPU at only 100ish milliwatts. For comparison, a low power arm chipset can be 2.5-5W

 

The only two companies who have semi-finalized and public neuromorphic chip designs afaik at the moment are Intel with their new Loihi chip and IBM with their slightly older TrueNorth chip.

 

Neural Net inferencing is definitely the future, and making sure that devices can do it even without an internet connection is a big push right now. You're right that it doesn't solve every problem, but a lot of problems that were previously intractable can be handled quite neatly by machine learning.

I used a NN at 20 mW (Cortex M4 @ 64 MHz), so much for low power

To be fair, it wasn't very deep, but enougth to solve our problem, as a lot of "tradtional" agorithmic was used before the net. Also an ASIC would use even less power.

 

If I'm not mistaken, all the new NN ASICs like the TPU are for inferencing and you don't need an internet connection (only to update / retrain the net).

The avantage of intel is having a chip that can also train the net what is more compute intensive. On the other hand this makes it more complex and power intensive. While using less power than a GPU, it looks like 10s of watts given the heatspreader and the die size shown. But we all know that power consumtion is not equal to efficiency.

[NumLock21]

The Ai cpu in that pic is actually a Core i5!

[Sniperfox47]

8 minutes ago, Stefan1024 said:

I used a NN at 20 mW (Cortex M4 @ 64 MHz), so much for low power

To be fair, it wasn't very deep, but enougth to solve our problem, as a lot of "tradtional" agorithmic was used before the net. Also an ASIC would use even less power.

The difference being that this is at full load. The full TPD of TrueNorth for the whole SoC at max draw was 70mW, and it performed far better than most GPU setups.

 

Escaping the Von-Neuman architecture limitations is a big deal.

 

9 minutes ago, Stefan1024 said:

If I'm not mistaken, all the new NN ASICs like the TPU are for inferencing and you don't need an internet connection (only to update / retrain the net).

The avantage of intel is having a chip that can also train the net what is more compute intensive. On the other hand this makes it more complex and power intensive. While using less power than a GPU, it looks like 10s of watts given the heatspreader and the die size shown. But we all know that power consumtion is not equal to efficiency.

Yeah but a TPU is still Von-Neuman and is basically just a traditional processor designed for existing software neural networks solutions (a la Tensorflow and Cafe).

 

Also the picture was a stock Intel picture, not a picture of the actual chip...

 

[Stefan1024]

3 minutes ago, Sniperfox47 said:

The difference being that this is at full load. The full TPD of TrueNorth for the whole SoC at max draw was 70mW, and it performed far better than most GPU setups.

 

Escaping the Von-Neuman architecture limitations is a big deal.

 

Yeah but a TPU is still Von-Neuman and is basically just a traditional processor designed for existing software neural networks solutions (a la Tensorflow and Cafe).

 

Also the picture was a stock Intel picture, not a picture of the actual chip...

 

I was reading a bit more on the thematic. It's more an analog computer rather than a traditional one. This kinde of NN implementation is actually very neat and interesting. And also different to what most of the other guys doing.

[shermantanker]

Man... what if the Anti-Christ is an all powerful AI supercomputer overlord, and the mark of the beast is a microchip to network everyone to it.

[Guest]

5 hours ago, Sniperfox47 said:

They basically do what a Neural Net does in software in hardware instead and are far *far* more efficient for inferencing, consuming hundreds of times less power than a GPU/TPU at only 100ish milliwatts. For comparison, a low power arm chipset can be 2.5-5W

Would love to see how this neuromorphic processor compares to the so called NN found in smartphone SOCs (Isn't the A11 Bionic supposed to have one now too?) Maybe we will see several of these on a PCIe card or alternative in the next few years. (Or even in a prototype laptop to allow developers more time with it)

[Nicnac]

how do you even pronounce that?