Hacker News Reader

Comment by ademup | Hacker News Reader

"Zen employs a dynamic predictor known as a hashed perceptron." When will Hollywood tap into this wealth of cool vocabulary?

Is this a gimmicky marketing term, or is it logical\descriptive\rational?

View on HN

(Note: Also replied to another comment, but seems relevant here, too.)

There's a very approachable explanation in the following (around the 1:01:40 mark): https://youtu.be/8I_1TSs695I?t=1h1m40s -- part of Design of Digital Circuits - Lecture 18: Branch Prediction II (ETH Zürich, Spring 2019). Related readings: https://safari.ethz.ch/digitaltechnik/spring2019/doku.php?id....

These lectures are pretty great, by the way, highly recommended to anyone interested in computer architecture: http://people.inf.ethz.ch/omutlu/lecture-videos.html

Incidentally, this year's High-Performance Computer Architecture Test of Time Award has been given to "Dynamic Branch Prediction with Perceptrons" referenced in the lecture (from 2001, https://www.cs.utexas.edu/~lin/papers/hpca01.pdf): https://engineering.tamu.edu/news/2019/02/jimenez-receives-h....

https://www.wikiwand.com/en/Perceptron

Looks like an already studied technical description going back to 2005:

https://github.com/ChampSim/ChampSim/blob/master/branch/hash...

https://www.jilp.org/cbp2014/paper/DanielJimenez.pdf

> Introduced by Tarjan and Skadron 2005

> Basic idea:

> - Hash segments of branch history into different tables

> - Sum weights selected by hash functions, apply threshold to predict

> - Update the weights using perceptron learning

https://ieeexplore.ieee.org/document/4492487

Hashed probably means that some features are mapped to a desired perceptron weight via a hash function. This serves as an implicit regularization and can be much more efficient (no need for a predetermined feature vector, sparse representation etc.). It's called hash trick in ML. Perceptron is a single layer NN. Dynamic predictor, a branch predictor that adapts to program input, not just some predetermined state (like a formula that was shown to be good enough for a collection of programs).

In this case, I'm pretty sure it means that multiple instruction addreses are hashed to the same perceptron (or is what you said effectively the same thing in ML-speak).

Since you don't have branch predictors for each address, you share the same prediction data for multiple addresses (this is one often underlooked cost to heavily branchy code).

A perceptron is a simple type of neural network that is very fast to use when already trained. So it is a real network, I don't know what hashed perceptron is but I'd guess it would have to do something with hashing the inputs to the perceptron?

"Hashed" in this case simply means they did some extra work to deal with the perceptron's XOR problem, since a straight mapping of PC and branch history would led to an inability to distinguish XOR patterns.

Sounds like the main ingredient of Tacos de Sesos.

No, it's not a marketing term. That's literally the name of the branch predictor used. See https://www.cs.utexas.edu/~lin/papers/hpca01.pdf

Hollywood might go for "GameCache" instead.

Reply on news.ycombinator.com