Caches - SenseMI and Chipset

20 MB L2+L3 Cache

Another news fact we can now share or at the very least confirm are the RYZEN cache sizes, all we got was the number which is 20 MB for the L2 and L3 cache, and that matches what we have discussed numerous times. Summit Ridge aka RYZEN would in our opinion boils up towards a L1 data cache size is 32 KiB and L2 cache size 512 KiB per core. So the CUs = 2x8MB (L3) + 8x512KB (L2) = 20MB in total for caches, the L2 and L3 caches are identical in size compared towards the Intel Core i7-6900K. AMD has not yet release information about compatible instruction sets and other features. We expect  multiple channels of DDR4 2400 support, AVX2, AES SSE 4.1 and 4.2 instruction sets etc. 

 SenseMI technology

AMD is implementing technology that will make sure that this processor runs application efficient and optimized. SenseMI is a set of sensing and adapting technologies, including an artificial network inside every “Zen” processor to anticipate future decision, preload instructions, and choose the best path through the CPU.
  

 
AMD also is introducing a new interconnected called AMD Infinity Fabric. This is a new and very fast way of connecting various parts within an SoC. Infinity fabric is not only used in the RYZEN processors, but also in the future Vega AMD GPUs and (almost) all other AMD chips in the near future. Inifinity Fabric allows for faster and better secure connections within a chip. The inter- and intra-chip connector will be standardized and used in many AMD product and as such AMD can easily communicate over that very same interconnect. SenseMI technology then, it is based on five parts as shown in the slide above. Pure Power is a technology that allows the Ryzen chips and other Zen variants to work as efficiently as possible.
  

According to AMD, the CPUs are equipped with hundreds of sensors that monitor temperature, voltage and power consumption of all chip segments. Through the Infinity Fabric connections these sensors are all connected to a central control unit, the Infinity System Management Unit. The same network of sensors constantly monitors how much room there is left for extra performance, chip parts can be independently adjusted from each other. Clock frequencies within a plit second can be altered in steps of 25 MHz. 
 
  

Key to understand the euphemism that is SenseMI is that everything inside the processor works heterogeneously and optimized. So you get precise boost frequencies matching towards power management. What in a simple to grasp manner to understand is that SenseMi will do a lot of analyzing and smart prediction based upon the workload at hand. If I recall correctly it was back in the summer in a presentation that AMD already announced that the existing Zen branch predictor (the code that predicts the path there with branches in code probably will be) was mentioned to have been greatly improved. AMD indicated that has improved thanks to AI-like algorithms and protocols. The new branch predictor has a self-learning system on board, making decisions based on good or bad branches and thus it can adapt and improve on performance based upon specific workloads and surrounding dynamics. Branch prediction is essential for the efficient operation of the processor. Based on applications and workloads it can detect what course or path is better to follow, making the result a faster one. It will analyze the workload and adapt and optimize based up-on that. The biggest gain in RYZEN IPC performance is to be found, among other things, right here with much-improved branch prediction.

So decisions are driven by software code execution and anticipate on following decisions, pre-load instructions, choose the best path through the CPU. It’s pretty impressive when you think about it, a processor that optimizes workloads based on algorithms and predictions. So as stated AMD will be able to precisely bin and adjust clock frequencies in steps as small as 25 MHz. RYZEN will get that much smarter prefetecher that anticipates the location of future data accesses by application code . It has a learning algorithm model and learn application data access patterns. Prefetches vital data into local cache so it’s ready for immediate use. Efficiency at its best, and after seeing the first results, AMD might have struck gold here.
  

 
In a much broader sense, SenseMI is a monitoring algorithm that monitors and adapts on many levels. On a more arbitrary level this will also means that SenseMI with its many sensors can recognize thermal characteristics and adapt to it, e.g. a better cooler will optimize the processor to go faster and perhaps clock the Turbo bins upwards a notch. A RYZEN processor with water-cooling will get better temperatures right ? Well the technology will detect that thermal signature and as such it will allow higher boost frequencies with better cooling. Thermal tweaking control at the processor level.

The chipset

A new processor will need a new chipset. AMD sold its chipset division I think two years ago already, a new processor series will need a new chipset as the motherboard needs an infrastructure as well. While most of the chipset functionality is harbored inside the processor, a new generation and architecture processor will always a new motherboard chipset.
  

 
AMD denotes DDR4 support, yet has not disclosed if this is dual-or quad-channel. Considering the above screenshot shows just 4 DIMM slots, you may expect dual-channel. Other then that PCI-Express gen 3.0 is now confirmed as well as support for the NVMe protocol, USB 3.1 Gen 2 (10 Gbps) and M.2. and SATA Express support.

Motherboard and chipset wise, this however was the only thing that AMD has shared. That's it for this small update. We should see some more announcements during CES, and then later on in Q1 2017 AMD is releasing the RYZEN processor series starting with the 8-core / 16 threaded part previously known as Summit Ridge.

ATH +++

- Hilbert out