When AMD’s Ryzen family of microprocessors first debuted back in March of this year, they also debuted a multitude of new features and technologies. A group of such technologies are known collectively as AMD SenseMI technology. Information on SenseMI is relatively scarce and can be daunting to understand. For these reasons, I’ve decided to cover each of the five technologies with fairly short but to-the-point descriptions.
Each and every Zen-based processor features over 1,370 sensor points per die from which SenseMI can gather important information about the silicon at any one time. Obtainable readings include temperature, core frequency, voltage and power consumption. The following five technologies rely heavily on these sensors to determine which is applicable to the current workload. Furthermore, AMD SenseMI technology is completely automated and operates at hardware level. As such, it requires drivers to work properly.
We’ll start with the feature you’re most likely to witness at work in the Windows 10 Task Manager. Precision Boost is essentially the direct replacement for Turbo Core technology, found in the likes of the Phenom II X6 and FX processors. As its replacement, Precision Boost can provide additional performance for both single- and multi-core workloads by temporarily increasing the frequency of the processor core(s). Worthy of noting is that the highest Precision Boost frequency given actually applies for up to two cores and four threads, and isn’t an outright single-threaded turbo frequency as with previous architectures.
However, unlike the predecessor, Precision Boost has much finer control over the hardware. When criteria from the sensors are met, frequency is increased by 25 MHz at a time, versus Turbo Core’s 100 MHz.
All Zen-based processors support AMD Precision Boost technology.
Extended Frequency Range (XFR)
Next up is another frequency boost feature. This time, Extended Frequency Range, or XFR, extends on to Precision Boost to provide even more performance under single- and multi-core workloads. The sensors are used to determine available headroom for additional frequency beyond that of Precision Boost.
Where the two technologies differ is when they kick in. While Precision Boost is applied for multiple core utilization scenarios, XFR is only applied in Ryzen 3, 5 and 7 processors for up to two cores simultaneously. Ryzen Threadripper sees an improvement here where XFR is provided for up to four cores at once.
AMD states that XFR requires an adequate cooling solution for it to work, however XFR will work perfectly fine on AMD’s stock Wraith coolers. With beefier air coolers, and under water and liquid nitrogen cooling, XFR will apply its highest frequencies for a longer period of time.
All Zen-based processors support AMD Extended Frequency Range technology, however for models without the “X” in the model number, XFR will only be enabled if said processors are paired with an X370-wielding motherboard. These models also have smaller gains from XFR (50–100 MHz, versus 100–200 MHz).
This technology provides Zen processors with finer control over processor core C-states and package P-states when the raw horsepower of the silicon isn’t required. As such, it replaces the Cool’n’Quiet and PowerNow! technologies found in older desktop and laptop chips, respectively. While a Piledriver-based FX-8370 can underclock down to 1.40 GHz in its lowest power state, all Zen-powered chips can reach down to 550 MHz, with many intermediate states in-between. The 14-nanometer low-power-performance lithography process that AMD’s Zen processors are built upon is excellent for power saving, and is in another league of its own when compared to the power efficiency of the FX series. The Zen architecture is at its peak on the efficiency curve between 2.00 and 3.00 GHz.
All Zen-based processors support AMD Pure Power technology.
Smart Prefetch and Neural Net Prediction
The final two technologies are closely related and refer to the branch prediction unit found in each Zen core. Compared to its predecessors, Zen’s branch prediction unit was thoroughly improved and as a result is also more complex. Instructions sent through a microprocessor are decoded into smaller more manageable segments called microoperations.
Smart Prefetch aims to make use of a microoperations cache to store these sub-instructions and can refer back to them when required. This technique is known as ‘hitting’ the prefetch buffer. This negates the need to decode the same instruction more than once in a parallel workload and can therefore increase instruction throughput. A ‘miss’ from the prefetch buffer will result in the instruction being decoded again, which can waste valuable clock cycles.
Neural Net Prediction, or NNP, allows Zen-based chips to accurately predict the routes for microoperations, based on previous behavioral patterns. It functions very similarly to Smart Prefetch with the exception that while Smart Prefetch only applies to one specific instruction at a time, NNP is capable of predicting which instructions follow after another. Therefore, NNP can predict a sequence of instructions and can load them from the prefetch buffer, ahead of execution time.
All Zen-based processors support AMD Smart Prefetch and Neural Net Prediction technologies.
This frequency is applicable for up to two cores and four threads, for Ryzen 3, 5 and 7 microprocessors.
This frequency is applicable for up to four cores and eight threads, for Ryzen Threadripper microprocessors.