That is a* very, very important paper. The answer is yes.
(The guy who posted the link is Dan Luu who used to work for a company that made x86 CPUs. His blog is worth looking at. He also used to work for Twitter -- but his predictions about the Doom and Disaster after Musk's takeover hasn't panned out at all. Good at some things (CPU architecture and blog posts). Not so good at Elon Musk predictions ;) )
---
Edit: put in the missing "a". My fingers and my brain don't always agree on what to write.
Modern branch predictors are based on relatively simple state machines and already have >95% accuracy. Thus, even if machine learning-based predictors can sometimes beat them, it is not clear that they can beat them enough for the much more complicated circuitry they need to be worth it.
There are several multipliers and adders (possible, floating point ones), a register file for perceptron weights and perceptron selection logic. The scheme also need to keep information somewhere for training - the decision and backpropagation stimulus are separeted in time.
So the circuitry is complicated despite superficial simplicity of the model.
The version in the paper only is ever multiplying by 1 or -1, so you do not need a full multiplier circuit. The weights are also stored as signed integers, not floats, so no complicated floating point circuity. I am not sure what current state of the art is, but considering the cost of multipliers/floating point circuity I would be surprised if it changed to those if signed integers work.
All branch predictors need some way of storing their state and selection logic, and the way a perceptron branch predictor stores its data is just a big table indexed by some hash of the program counter of the branch, which is pretty standard for branch predictors. Also, all branch predictors have a sort of "backpropogation" in that pipelined processors produce the actual result of the branch (possibly many) cycles later, so this also is not as much of a factor. Since the training is a function of the weights you do not need to store extra data beyond the threshold, but that is already being computed as the prediction anyways.
bot> A 5% miss rate in branch prediction leads to a 14.7% improvement in misprediction rates on a trace of SPEC2000 benchmarks compared to the gshare predictor. The use of machine learning-based predictors has the potential to improve these results further.
To be clear, all hardware branch predictors are "relatively simple state machines"; they need storage in the branch prediction tables which must super-fast to access, which means they can only store a few (sometimes dozens, but certainly not hundreds) bits per branch to reach the access latency goal. With input to the predictor encoded as binary, the weights quantized and small and encoded into binary, and the history small, even perceptrons are "relatively simple state machines". After all, their implementation is just going to become some combinatorial logic in the end.
One thing to point out is that the threshold of predictor complexity is dependent on the execution pipeline. A very speculative and deep architecture has a bigger need for better predictors, since it has a massive penalty when there is a misprediction.
As someone who works on design verification for modern CPUs, this is not true. Most modern CPUs use some form of TAGE branch prediction, if not something more complex.
Well, yes, but the complexity of TAGE is nowhere near the complexity of a modern deep neural network where you essentially have no chance of figuring out how its predictions are made. Branch predictors based on multi-layer networks, RNNs, LSTMs, or networks with more than a handful of neurons is a hot research topic, but has not, to the best of my knowledge, found their way into any modern cpu.
Yes modern CPUs aren't using neural nets in hardware or anything like that for branch prediction. I was referring to TAGE not being a simple state machine. That makes it sound like modern CPUs are just using a basic two-bit branch predictor.
With many architectural caches, it's typically valuable to think of miss rate instead. A 5% miss rate is very high if the miss path is 20x the cost of the hit path - misses account for 1/2 of the execution time, so reducing that to 4% gives a 10% improvement in runtime.
system> What are the microarchitectural tricks that allow prediction to take place in one clock cycle?
bot> The microarchitectural tricks include using the branch address to hash and select a perceptron from the table, calculating the dot product of the perceptron and the global history register, using the training algorithm to update the weights in the perceptron, and writing the updated perceptron back to the table.
Modern CPUs use TAGE-like[1] and Perceptron-based branch predictors, achieving a 99.7% prediction accuracy. Zen 2 in particular uses TAGE[2].
I recall implementing one in software in a computer architectures class; it was pretty gnarly but the prediction accuracy (and therefore performance) compared to a simple two-bit saturating counter is immense.
I implemented this in a CPU simulator for a computer architecture class back in grad school, and I believe I couldn't reproduce their prediction success rate.
essentially yes, except that on modern CPUs the fetch/decode stages (where the branch predictors live) are so far away (in space and time) from where the loop counters live that it's not practical to connect them directly - instead the front end maintains history bitmaps of recent branch activity (whether branches are taken or not) which in effect contain the history of how many times you've been around a loop
26 comments
[ 3.0 ms ] story [ 62.6 ms ] threadDid this go anywhere? Is it worth revisiting in 2023?
(The guy who posted the link is Dan Luu who used to work for a company that made x86 CPUs. His blog is worth looking at. He also used to work for Twitter -- but his predictions about the Doom and Disaster after Musk's takeover hasn't panned out at all. Good at some things (CPU architecture and blog posts). Not so good at Elon Musk predictions ;) )
--- Edit: put in the missing "a". My fingers and my brain don't always agree on what to write.
So the circuitry is complicated despite superficial simplicity of the model.
All branch predictors need some way of storing their state and selection logic, and the way a perceptron branch predictor stores its data is just a big table indexed by some hash of the program counter of the branch, which is pretty standard for branch predictors. Also, all branch predictors have a sort of "backpropogation" in that pipelined processors produce the actual result of the branch (possibly many) cycles later, so this also is not as much of a factor. Since the training is a function of the weights you do not need to store extra data beyond the threshold, but that is already being computed as the prediction anyways.
bot> A 5% miss rate in branch prediction leads to a 14.7% improvement in misprediction rates on a trace of SPEC2000 benchmarks compared to the gshare predictor. The use of machine learning-based predictors has the potential to improve these results further.
Some updates to it were big point of "what we did in Zen" presentations when first Ryzen and EPYC CPUs landed.
To be clear, all hardware branch predictors are "relatively simple state machines"; they need storage in the branch prediction tables which must super-fast to access, which means they can only store a few (sometimes dozens, but certainly not hundreds) bits per branch to reach the access latency goal. With input to the predictor encoded as binary, the weights quantized and small and encoded into binary, and the history small, even perceptrons are "relatively simple state machines". After all, their implementation is just going to become some combinatorial logic in the end.
system> What are the microarchitectural tricks that allow prediction to take place in one clock cycle?
bot> The microarchitectural tricks include using the branch address to hash and select a perceptron from the table, calculating the dot product of the perceptron and the global history register, using the training algorithm to update the weights in the perceptron, and writing the updated perceptron back to the table.
I recall implementing one in software in a computer architectures class; it was pretty gnarly but the prediction accuracy (and therefore performance) compared to a simple two-bit saturating counter is immense.
[1]: https://doi.org/10.1145/3226098 [2]: https://en.wikichip.org/wiki/amd/microarchitectures/zen_2#Br...