📜 ⬆️ ⬇️

Deep inside the Exynos chip in the Galaxy S7 is a neural network to predict transitions

14 nanometer FinFET Exynos 8890 crystal system

Yesterday at a specialized conference on microelectronics Hot Chips in California, Samsung engineers for the first time showed at the presentation drawings of the mysterious M1 processor cores (codenamed Mongoose), which work on S7 and S7 Edge smartphones.

In the international versions of these Android smartphones, a 14-nanometer system is installed on a FinFET Exynos 8890 chip. It has four standard ARM Cortex-A53 cores (1.6 GHz) and four proprietary M1 cores operating at 2.3 GHz and 2.6 GHz.

The American development team developed the M1 from scratch in three years on its own project . In the benchmarks, the Exynos 8890 is inferior to the iPhone 6S A9 chip in performance on a single core, but wins in multi-core tasks.

Core plan M1. Image: Samsung

One of the components of the M1 core is the built-in branch predictor (branch predictor), which predicts whether a conditional branch will be executed in an executable program. Branch prediction allows to reduce pipeline downtime due to preloading and execution of instructions that should be executed after the conditional branch instruction has been executed. Branch prediction is critical because it allows optimal use of the processor’s computational resources.

Branch prediction through a predictor of transitions is a quite standard feature in processors. Already in the first SPARC and MIPS processors, a primitive static branch prediction method was used, when instructions placed after the conditional branch instruction are always loaded into the pipeline. In modern processors, more advanced methods of dynamic prediction of transitions are implemented:

So, it turns out, in the M1 core, Samsung engineers implemented branch prediction using a neural network . This is a relatively new type of predictor of transitions. For the first time, it was proposed to be used in theoretical work by Professor Lucian Vintan in 1999 (scientific article " Towards a High Performance Neural Branch Predictor "). Two years later, the first predictor of transitions on perceptrons was developed, which theoretically could be implemented in hardware (scientific article " Fast Path-Based Neural Branch Prediction ", by Professor Daniel A. Jiménez from Rutgers University in the USA).

A key advantage of a predictor on a neural network is a linear growth of resources with an increase in the analysis of the transition history (in classical predictors, resource consumption grows exponentially with an increase in history). For this reason, the predictor on the neural network is more efficient. Already the first neural network from the work of Daniel Jimenez showed an efficiency gain of 5.7% compared to the hybrid predictor Scott MacFarling .

In subsequent years, the same Daniel Jimenez and other researchers worked to eliminate the shortcomings of the predictor of transitions, including a large delay in calculations.

The newest ideas for predictors of transitions are proposed by researchers at the Championship Branch Prediction competition, the last of which was held in June 2016 at the ISCA 2016 Computer Architecture Symposium in Seoul.

Despite the vigorous scientific research in this area, until now it has not been known about a single mass processor in which the predictor of transitions on perceptrons is implemented. The point is not that FinFET Exynos 8890 with M1 cores in S7 and S7 Edge smartphones is the first such processor. Just the developers of Samsung and other companies keep the information secret, and this can be understood. Conversion predictors are one of the most guarded secrets in the semiconductor industry. Manufacturers often do not even patent their conversion predictors, so as not to give out a secret to competitors, and also because later it will be difficult to prove the fact of patent infringement, taking into account the most complicated logic in modern proprietary processors.

Slide from Samsung presentation at Hot Chips conference. It depicts a transition predictor module using a neural network. Image: Samsung

Samsung is the first company that officially announced the use of a neural network in its transition predictor. Experts in the microelectronic industry say that a similar technology seems to be used in Jaguar and Bobcat predictors in AMD chips. Not surprisingly, the current Samsung Vice President and Samsung Processors Director, Samsung Research Center in Austin (TX), Brad Burgess, previously headed the Bobcat microarchitecture project at AMD.

It is likely that Intel and AMD are quietly using perceptron predictors in desktop and server processors. As already mentioned, this information is kept secret and is not even patented.

Samsung first broke the vow of silence. Perhaps the developers of Exynos microarchitecture just decided to brag.

Source: https://habr.com/ru/post/397075/

All Articles