This is literally insane. I've never heard of a neural net branch predictor.
Edit: my layman knowledge about branch prediction and neural networks is showing its gross inadequacy :/
Edit: my layman knowledge about branch prediction and neural networks is showing its gross inadequacy :/
Even a simple dot product can be called a "neural network," albeit a small uninteresting one. Your features could be (say) the state of cache, the number of jumps, the number of recent stalls / stack size / returns, and so on. Put them into a 10-dimensional vector or whatever, take the dot product between that and a set of 10 learned weights, and preload the branch if the result is above some threshold. Even a simple model could perhaps work really well here.
Of course it's not like they have some deep learning convolutional net running on the program input. I also bet the weights are set at the factory so it's not "learning" anything.
Learning on the fly would consume much more power as it back-propagates, so doesn't really make much sense as the potential performance gain would be marginal and would probably cause a performance hit when switching to a different process or thread.
> The perceptrons are trained by an algorithm that increments a weight when the branch outcome agrees with the weight’s correlation and decrements the weight otherwise.
A "perceptron" seems to be a single linear function, akin to a single neuron in a neural network. So complex learning algorithms like back-propagation don't apply here.
The complete process, from the paper, is:
> 1. The branch address is hashed to produce an index "i" into the table of perceptrons.
> 2. The "i"th perceptron is loaded from the table into a vector register "P" of weights.
> 3. The value of "y" is computed as the dot product of "P" and the global history register. [This is a shift register containing values of -1 or 1 for the last N branches seen. -1 means not taken, and 1 means taken.]
> 4. The branch is predicted "not taken" when "y" is negative, or "taken" otherwise.
> 5. Once the actual outcome of the branch becomes known, the training algorithm uses this outcome and the value of "y" to update the weights in "P".
> 6. "P" is written back to the "i"th entry in the table.
The article says that AMD's microarchitecture uses a hashed perceptron system in its branch prediction. So, from the above, it seems that "hashed" means that you hash the branch address to pick which perceptron (model) to use, and "perceptron" is just a dot product. (Actually the wikipedia article[2] suggests that the name "perceptron" refers to the training algorithm, i.e. how you update the weights.)
[1]: Daniel A. Jiménez and Calvin Lin. 2002. Neural Methods for Dynamic Branch Prediction. https://www.cs.utexas.edu/~lin/papers/tocs02.pdf
Search google for "hashed perceptron branch predictor"
Edit: also, Intel branch predictors up to Core2 were well characterized and were not believed to be using NN.
[1] https://hal.inria.fr/hal-01100647/document ,discussed last year at HN.
I'm not as familiar with mobile phone architecture as with that of PCs, but the number and types of operations influenced by the device's various sensors (light sensor, GPS, accelerometer, gyroscope) could conceivably be more than normal, naïve branch prediction can handle effectively.
Also I do not see how the number or quality of sensors could in any way affect the prediction rate.
> AMD's Zen architect Mike Clark confirmed to us his microarchitecture uses a hashed perceptron system in its branch prediction. "Maybe I should have called it a neural net," he added.
Neural networks can be very complicated, but they can also be very simple. What little I know about perceptrons is that they are very simple. If you google "hashed perceptron", many of the papers that come up mention branch prediction in their titles.
OTOH I see this as the optimization of hardware to fix the cap that software, paradigms and devs mean.
This is maybe not the best approach for certain games or benchmarks, but for normal use this has proven to be great.