ML Newsnew | past | best | rss | submit about
Recurrent Neural Networks Hardware Implementation on FPGA (arxiv.org)
54 points by Katydid on Nov 20, 2015 | hide | past | pdf | 16 comments on HN

In plain words: A sequence-reading network with built-in memory was wired directly onto a reprogrammable chip, so its step-by-step math runs in custom hardware. It ran over 21 times faster than the chip's own ARM processor core.

Abstract

Recurrent Neural Networks (RNNs) have the ability to retain memory and learn data sequences. Due to the recurrent nature of RNNs, it is sometimes hard to parallelize all its computations on conventional hardware. CPUs do not currently offer large parallelism, while GPUs offer limited parallelism due to sequential components of RNN models. In this paper we present a hardware implementation of Long-Short Term Memory (LSTM) recurrent network on the programmable logic Zynq 7020 FPGA from Xilinx. We implemented a RNN with $2$ layers and $128$ hidden units in hardware and it has been tested using a character level language model. The implementation is more than $21\times$ faster than the ARM CPU embedded on the Zynq 7020 FPGA. This work can potentially evolve to a RNN co-processor for future mobile devices.

Andre Xian Ming Chang, Berin Martini, Eugenio Culurciello
arXiv:1511.05552 · cs.NE · submitted Nov 17, 2015 · updated Mar 4, 2016
abstract · pdf · html · 7 pages, 8 figures, changed format, added figures, added references, modified introduction

add comment on HN

ilurk on Nov 23, 2015 [–]
> We implemented a RNN with 2 layers and 128 hidden units in hardware and it has been tested using a character level language model. The implementation is more than 21× faster than the ARM CPU embedded on the Zynq 7020 FPGA.

I'm left curious on the performance gain factor when scaling the network in terms of layers and units. Would the performance gap widen as the RNN grows?

reply
leonardt on Nov 24, 2015 | parent [–]
> Figure 8: The execution time is projected to decrease with the increase of number of LSTM cells running in parallel. This can lead to significant performance improvement.

They do say in the text that "Figure 8 shows the expected speed up, assuming the data throughput is high enough to handle the parallel processing" so take it with a grain of salt. There could and most likely will be (as there always is) other factors that prevent ideal scaling.

reply
iaw on Nov 23, 2015 [–]
Is this unexpected? I feel like we're heading towards more hardware implementations of Machine Learning to work around the current bottlenecks.
reply
spooningtamarin on Nov 23, 2015 | parent [–]
There are extremely large amounts of hardware based implementations. Any more complicated TV upscaler (Sony 4K TVs certainly) probably has a neural network embedded in it.

Any algorithms working with images need FPGA implementations to be quick enough, and there are a lot of convnets in use there.

There's a chapter on convnets here: http://www.cambridge.org/us/academic/subjects/computer-scien...

RNNs are nothing special in particular, at least these with small number of layers and nodes.

HMMs or CRFs are easier to handle for sequences, and would probably work well, and there are FPGAs all over the place of these models.

reply
p1esk on Nov 24, 2015 | parent [–]
Any links to TV upscaler chips using NN algorithms? I couldn't find any.

Also, my impression was that FPGAs are much slower than GPUs for neural nets. Unless you're talking about really high end chips like Stratix 10 from Altera, which cost over $30k. Power consumption is a different matter though.

reply
spooningtamarin on Nov 24, 2015 | parent [–]
https://community.sony.co.uk/t5/blog-news-from-sony/inside-4...

They won't tell you it's NNs but it is. Sony distributes a lot of movies, they use their database of movies to train the upscaling models (which is obviously NNs https://github.com/nagadomi/waifu2x ) and then put the chip in the TV.

It's something almost equivalent to storing Pride and Prejudice and Zombies in your TV in 4K, and then reproducing it when they match it with what's playing on TV.

reply
p1esk on Nov 24, 2015 | parent [–]
It does not mention anything about neural networks in that link. How do you know? I looked up that Sony chip, and I can't find anything "neural" about it either.

I don't quite understand what you mean by: "It's something almost equivalent to storing Pride and Prejudice and Zombies in your TV in 4K, and then reproducing it when they match it with what's playing on TV."

Can you explain? What exactly do they have to store in the TV?

reply
spooningtamarin on Nov 24, 2015 | parent [–]
Well, Sony Pictures is huge. Their database of movies is enormous. (Pride and Prejudice and Zombies is one of the movies distributed by Sony)

Imagine they stored all of the movies they distribute in 4K in your TV. Whenever the movie is displayed they just find it in the database and reproduce the 4K version.

Of course, that is very much infeasible, it requires too large of a storage.

What they do is something like Waifu2x I linked, they train a neural network to learn how to properly upscale movies. They put the "algorithm" on the chip, and its fast enough for real time usage.

reply
p1esk on Nov 24, 2015 | parent [–]
It seems like you're speculating. Unless you're familiar with the details of their chip, your guess is as good as mine.
reply
spooningtamarin on Nov 24, 2015 | parent [–]
I don't have to be familiar, you can know state of the art deterministic upscalers and their image quality, and know state of the art statistical upscalers and their image quality.

When you see side-by-side for these upscalers, it's obvious they aren't that good in deterministic upscaling.

reply
lightcatcher on Nov 24, 2015 | parent [–]
One of my favorite named startups ever is working on something likely related to video upscaling: http://www.magicpony.technology/
reply
p1esk on Nov 24, 2015 | parent [–]
That page says they are a software company. Are they working on an upscaler chip?
reply
iaw on Nov 24, 2015 | parent [–]
Wow, I was completely unaware that consumer products like TV upscalars were utilizing neural networks. I assume that the upscalar uses the result of a prior training session and doesn't need periodic retraining.
reply
spooningtamarin on Nov 24, 2015 | parent [–]
Yeah, I was surprised too when I heard that, but if you had a huge database of HD movies, you could train a NN yourself, see how it upscales, use some algorithms or different error functions to correct weird artifacts and put the trained weights on an FPGA.
reply
nuand on Nov 24, 2015 [–]
Seems like an interesting approach. However the 21x speedup seems a little underwhelming considering the speed of the Zynq's programmble logic fabric and how parallizable NNs are. They're quoting a 21x speedup over the ARM processor on the Zynq 7020, which is on par with what powers the RaspberryPi 2. My guess is they didn't pipeline their design enough or appropriately causing one of the datapaths to significantly limit their throughput.
reply
dang on Nov 23, 2015 [–]
Url changed from http://hgpu.org/?p=14968, which points to this.
reply
About | RSS | RSS (all) | HN arXiv