27 comments

[ 3.3 ms ] story [ 67.9 ms ] thread
Super nice, is it an asic designed for ML algorithms? Which ones? Well, I wouldn't call it a startup though, 100 employees, 10 years of development.
There isn't much technical in the article, but it does remind me a bit of the adapteva epiphany.
Here's some spec on the chip itself (KnuEdge seems to be a larger holdings company of sorts, the chip is more associated with subsidiary KNUPATH):

https://www.knupath.com/products/hermosa-processors/

i do wonder what the ARM license is being used for - acting as an interface to the cores themselves?
Looks like an FPGA with only the DSPs and no logic units.

The SoC is probably used in the same way (high level tasks like networking etc.) as it is in FPGA land.

Agreed. I've seen a few similar mockups that used a processor IP as the chip-level "traffic cop".
Very interesting.

It's still quite vague but this seems like a better summary: "We combine distributed computing cores with an ultra-low latency fabric in a massive-bandwidth “push” model, which delivers higher performance from small to ultra-large deployments."

Basically, it seems like the chip is aiming to provide more memory bandwidth for parallel computing, more relative to GPGPU computation, which indeed provide a lot of parallel computation but has a somewhat limited memory bandwidth.

The "neural" part seems to be just emulating some aspect of neural network architecture for the memory architecture and the chips seem to be suited for general purpose parallel computation (indeed, apparently aiming to be better suited than existing chips).

And I think it's good that the chips apparently aren't like IBMs "brain on a chip" project, which as far as I can tell is a complete mistake; given that's just emulating one neural network architecture, it prematurely commits to that approach when the proper nn architecture is still a question in considerable flux.

I hope this is something that could compete with Google's Tensor Flow chips and even be more general-purpose. If it is, that would be a great thing.

Wonder how similar to http://www.artificialbrains.com/spinnaker this is
From the article and their website, it doesn't seem like KnuEdge's compute units are spiking neurons. Spinnaker and IBM 's TruNorth using spiking neurons to compute.
Well, not quite. Spinnaker is OPTIMIZED for spiking neurons, but in theory you could run standard computations on it. It is just really messy (speaking form experience).

The main idea behind Spinnaker as far as I understood it is "cheap cores, lots of them, architecture for message passing and load distrubtion vaguely inspired by human brain". This rang similar on a first read

> “I thought, holy smokes,” he said. “It’s going to be too expensive. It’s not propulsion. It’s not environmental control. It’s not power. This software business is a very big problem, and that nation couldn’t afford it.”

I think the difficult thing about space travel and exploration is that it is all of the above.

This part of the article struck me as very interesting:

KnuVerse solutions are based on patented authentication techniques using the human voice – even in extremely noisy environments – as one of the most secure forms of biometrics.

Of course that requires a very significant [[citation needed]]. I would like to see their performance in that task or at least the specific patents that allow them to provide the most secure authentication method even in extremely noisy environments.

EDIT: Google Patent shows nothing for KnuEdge, KnuPath or KnuVerse. They must have a very-differently-named shell company for the patents.

The company changed names somewhat recently from Intellisis (mentioned in the article). Try searching for that.
I find it amusing that they're touting it for an application where I find it quite likely that similar NN algorithms will excel at generating speech in someone else's voice. (c.f., "neural style transfer" in images, but applied to speech.) We're already getting pretty decent at this for video -- see, for example, https://www.youtube.com/watch?v=ohmajJTcpNk (Face2Face: Real-time Face Capture and Reenactment)
was just trying to find papers for that yesterday - neural style for voice/audio. Conceptually it sounds it should be doable, but looking at the actual implementation, I'm not sure it's doable at all with a CNN for audio.
Unsure if it's CNN-based, but check out http://www.wowtune.net/
very impressive, even if a bit out of tune on the Autumn Leaves demo!

They have a very good team with actual audio industry experience. Look forward to seeing more demos.

It sounds like a phoneme based speech/singing synthesizer, similar to Yamaha Vocaloid. I wonder how much training data is required to extract the phonemes to create a "voice"

I wonder what the clock network topology looks like on a chip with 256 cores.
If they're implementing neural nets they don't necessarily need to do so cycle synchronously. Think of it as a specialized cluster on a chip.
I think the commenter is talking about the KnuPath chip.

>And it (each core) will run C++ software, something that is already very popular. Programmers can program each one of the cores with a different algorithm to run simultaneously

Running C++ almost certainly has to be done synchronously

https://www.knupath.com/products/hermosa-processors/

Is a diagram of that chip. You don't actually run 'C++ software', you run a binary and if you have a cluster on a chip you are perfectly capable of running each node in the cluster on its own clock net, with high speed links connecting the nodes.

I must admit to being reminded of the Cell CPU hype, which is perhaps another case of a solution looking for a problem.