Yesterday, we unveiled AMD’s Radeon Instinct machine intelligence initiative and now, we’ve got some new details to share about the next-generation Zen desktop processor’s built-in intelligent control technologies called SenseMI.
First things first, the official production name of the Zen desktop processor (Summit Ridge) has been revealed, and it is Ryzen. Yes, AMD loved the “Zen” codename so much that they decided to incorporate it into the final product, inclusive of hipster spelling. The name seems to imply that this will be The chip that puts AMD back on equal ground with rival Intel in the enthusiast CPU market.
AMD has also finally revealed some specs about the chip. You can expect a high-end Ryzen CPU to have 8 cores + 16 threads, a base clock speed of 3.4GHz or better, and 20MB cache (4MB L2, 16MB L3). Ryzen will run on the new AM4 socket as previously announced.
Back in August, we covered many of the Zen microarchitecture’s improvements that go into the Ryzen chip. If you haven't read that yet, you can head over for more technical details. In that article, there was a reference to a Blender rendering benchmark where a Zen CPU matched the performance of an Intel Core i7-6900K. AMD has confirmed that this benchmark was done with an engineering sample Ryzen chip at a base clock of 3.4GHz with no boost enabled. This means that Ryzen, with boost, should outperform its equivalent enthusiast Intel counterpart.
While no further technical specs have been revealed, we now know that Ryzen will feature a group of new sensing and adaptive prediction technologies collectively known as SenseMI (pronounced Sense Em Ai, not Sense Me). These technologies are baked into the chip hardware and will continuously tune and tweak performance of the chip depending on the kind of apps and load that you’re running.
Here are the five major functions of SenseMI:
Pure Power monitors frequency, voltage and temperature under its current load and is supposedly able to optimize the processor to maintain the same level of performance, but at a lower power draw. This is an efficiency function and is designed to work in tandem with Precision Boost.
All the demos we’ve seen till date of Ryzen chips have their boost clocks disabled, but they will definitely have boost clock included; except it will be known a Precision Boost. What’s different is that Precision Boost will make use of the same chip-level monitoring as Pure Power to provide better performance on any workload for the same power draw as an unoptimized processor. At least that’s what it’s supposed to do. In addition, Precision Boost’s on-the-fly frequency adjustments come in smaller 25MHz increments for more fine-tuned control.
Now, it may seem at first glance that Pure Power and Precision Boost cancel each other out since one is designed to reduce voltage while maintaining frequency, while the other is trying to increase frequency while maintaining voltage. AMD hasn’t revealed how SenseMI algorithms will decide what kind of workloads will trigger either or functions, and since we’ve not seen a Ryzen chip with boost clock enabled yet, there are no demos of these two technologies working together. I guess we’ll just have to wait and see.
Instead of selling a special unlocked processor like Intel’s ‘K’ CPUs for overclockers, the XFR feature in SenseMI is basically a built-in unlocker. Ryzen chips with SenseMI is supposed to be able to detect the kind of cooling that you’re using (air, water and even liquid nitrogen), and if the chip’s operating parameters are within AMD’s safe range, you will be able to access frequencies beyond the chip’s Precision Boost limits based on the type of cooling you’re using.
While Pure Power, Precision Boost and XFR are easily understood as all three basically add an additional level of sensors to better automate the monitoring and optimization of frequency, voltage and temperature, Neural Net Prediction and Smart Prefetch fall into grayer territory. Now, AMD has gone all out with its Machine Intelligence agenda, and this is another extension to it.
In a nutshell, a Ryzen chip with Neural Net Prediction and Smart Prefetch is supposed to be able to “learn” from the applications you run based on its code execution, and then optimize data branch prediction, access patterns and prefetch pathways for that particular app to get the most efficient use out of every clock cycle and cache access.
As you went through some of the slides above, you might have noticed something called the Infinity Fabric. AMD hasn’t gone very much into the details of it, but they’ve essentially moved on to a new approach to chip design. This new process is called Infinity Fabric (there is no special code name for the old processes, we asked).
Now even AMD found it a little hard to explain in a sentence what Infinity Fabric is, but from what we understand, it is a move away from monolithic SoC designs and allows AMD more flexibility to take parts of an SoC to put into another, better scale for size (from smaller notebook chips to larger server chips), and easy reconfigure transport layers as necessary. This comprises everything from topology to interconnects (HyperTransport) and extensive subsystem of micro sensors and controllers (network-on-a-chip). The Infinity Fabric is so named because it is a much more flexible and efficient approach to chip design, and while Ryzen is the first consumer product to be created with it, all future planned products from Vega10 GPUs, Naples servers and Raven Ridge notebooks will be based on this Infinity Fabric design.
I’m bringing this up because the takeaway seems to be the new network-on-a-chip design. This allows AMD to place sensors into just about every part of a processor, all of which enable better efficiency in chip operations and offer ultra-fine grain management over power and temperature. Basically, it is this move to the Infinity Fabric design that makes many of the SenseMI features possible on these new Ryzen chips.