AMD received harsh criticism when it announced that its new Ryzen 5000 Mobile U series processors would not all be using its latest core design. In the product announcement, we were told that some of the U series processors would be based on the previous Zen 2 generation, and this was mainly for partners to take advantage of the new naming scheme, but also to reuse projects with the same approximate performance. Several technology enthusiasts (including me, I must say) scoffed at this, as it made the whole system complex. It’s still complex, but we understand that these latest Zen 2-based mobile processors also include a series of updates that make them a better version than they are.
To simplify things, I will call these products by their codenames AMD. Older Zen 2 processors are called Renoir, and newer Zen 2 processors are called Lucienne. Here is a list of the new Ryzen 5000 U-Series, with Lucienne listed in yellow.
Renoir, for all intents and purposes, was a very successful product for AMD. Placed in the Ryzen 4000 Mobile series, it has become the basis of AMD’s mobile portfolio and has been installed in nearly 100 award-winning projects since it was launched on the market. Lucienne, on the other hand, is a secondary player in the latest Ryzen 5000 Mobile series. It does not have the updates that the new Zen 3 cores have, but we have since learned that on the powerful side of things, instead of being a copy of Renoir, it is almost certainly Renoir Plus.
What Lucienne brings to the table about Renoir comes in discrete categories.
Memory controller
The memory controller in Lucienne is now able to decouple its voltage from the cores and enter a lower power state when not in use or for reasons of low bandwidth. Ultimately, this saves energy and AMD has allowed it to bypass specific voltage indicators to help it stay in the low voltage state. In addition to the cores and graphics, the other two energy consumers within a mobile processor are internal communications and external communications, of which the memory controller is under the latter. AMD has also implemented a system by which the memory controller can wake up to a state of full bandwidth faster than before, allowing for a better response from those states of deep sleep.
In addition, the memory controller can now support twice the Renoir’s memory capacity: up to 64 GB of DDR4-3200 or up to 32 GB of LPDDR4X-4267. Using DDR4 means that the system can have more peak memory, in addition to being user adjustable, however, the LPDDR4X switches them for faster bandwidth (68.4 GB / s vs 51.2 GB / s).
Core tension control
In circumstances similar to those of the memory controller, having voltage control of each individual core on a mobile processor is an angle to maximize performance when needed and minimize power loss when idle. In Renoir, all cores can adjust their frequency, but they all had to run at the same voltage. Lucienne changes this so that each core can adjust its voltage independently, allowing for more refined energy management and a more efficient energy system. There are also additional hooks that operating systems can use if they know in advance that high-performance cores are needed.
Preferred core
When we talk about turbo, historically it has been assumed that any core can achieve the highest single-core turbo frequency and that the workload is sometimes shifted between the cores to help with thermal management. When a system uses a preferred core, however, it means that a system can be optimized for that specific core and extracted more performance. AMD introduced its Preferred Core technology on the desktop two generations ago, and now it’s about mobile processors. A core of the eight on the Lucienne silicon will be designated as the best core and, through an operating system driver (standard on Windows), all workloads will be placed preferably on that core.
Frequency ramp
One of the features that bring it all together is the speed with which a core can go from idle to maximum performance and vice versa. If a system takes too long to accelerate or decelerate, responsiveness and energy will be lost. A typical modern system is expected to increase from inactive frequency to peak frequency in two frames at 60 Hz, or 32 milliseconds, however, the latest systems from AMD and Intel did this much faster, usually in 16 ms. AMD’s improved clock control technology now allows Lucienne to reduce this to 1-2 milliseconds. This means that a system can easily increase and decrease between each keystroke on a keyboard, providing immediate response to the user while keeping total energy usage low. In the 16-32 millisecond regime, typing on a keyboard can mean that a core is active almost continuously, however, making this change faster allows for great energy savings through these transitions.
Continuous performance levels
The inherited way for an operating system to command performance is through performance states or P-states. In that case, the operating system would request a specific level of processor power and performance based on the detected workload and the processor would respond. This was originally implemented during a time when the turbo was hitting modern processors and workload analysis was best done through the operating system. Now we can do this level of monitoring on the processor directly and through an operating system driver (already part of Windows), with system support, this level of frequency control can be passed back to the processor. The processor also achieves an effective continuous distribution of performance, instead of discrete P states.
While Renoir had P states, Lucienne gets the benefit of performance requests at the CPU level.
Faster integrated graphics
With additional power control in other parts of the core, how the power delivery works for the integrated graphics has also been adjusted to allow for better regulation and, finally, a lower minimum voltage. Through firmware, AMD has enabled a frequency-sensitive prediction model that allows the GPU to adjust its voltage and frequency based on its dynamic power management. Along with better regulation and the balance of the power budget made between CPU, interconnect, DRAM and GPU, more power budget is available for the GPU. For Lucienne, this means +150 MHz at peak IGP speeds compared to Renoir.
The slide shows Cézanne’s numbers, but it applies to Lucienne too
But I thought Lucienne Silicon was just like Renoir Silicon?
This is the big question. We asked AMD if Lucienne was the same attitude as Renoir, and the answer was not exactly compromised in one direction or the other. The simple answer is yes, however, AMD wants to make it clear that substantial changes have been made to the firmware and manufacturing, which means that while the transistor layout is identical, there are Lucienne features that would never have worked on Renoir without the changes.
So, although it is the same silicon layout and floor plan, some of these features were not possible in Renoir. AMD created these features perhaps knowing that they could not be enabled on Renoir, but sufficient changes and improvements to the manufacturing and firmware stage were made to enable these features to be enabled on Lucienne. Most of the time, these ideas tend to have very limited time windows for implementation and, even if they are designed on the hardware, there is a strict cutoff point where if it doesn’t work as planned, it won’t be activated. Obviously, the best result is to make everything work on time, but building CPUs is more difficult than we think.
Sometimes I wonder how we managed to make these lightning-driven rocks work in the first place.