If you haven’t had you fill of AMD news this week, today we have a Zen reveal, or should we say Ryzen reveal. Yes that is the new name for the Zen brand, and it has a new logo too.
SemiAccurate was briefed on some new details surrounding Zen/Ryzen last week, not too much technical but a lot of higher level bits to share. To start out with, you will be seeing Ryzen in Q1 for desktops, Q2 for servers, and 2H/2017 for laptops. Before the conspiracy theories start-up, if you look at the product refresh cycles, laptops get new chassis for fall/back to school and not really any other time. That should explain why Ryzen is set for 2H in mobile form, it isn’t any deeper than that.
Ryzen logo, two more letters and color
As you can see from the logo, there are two new letters and it is now in color. While we prefer the old black and white logo, the new one is a lot better than most, SemiAccurate almost likes it. So what are you getting, other than the spiffy logo? Effectively a Zepplin so 8-cores, 16-threads, 3.4+GHz, and 20MB of cache are the official bullet points. The unofficial specs are 512KB of L2 per core and 2x 8MB of L3 shared across groups of four cores per block. There are four slices per 8MB block that are picked by low-order address interleaving with the same average latency per core Each Ryzen has 32 PCIe3 lanes, 8 SATA/NVMe lanes, and 4x 10GbE with the caveat ‘up to’ likely because some of the phys and pins are shared. There are also two channels of DDR4 per socket.
AMD Zen core complex with cache
Backing this all up is something AMD calls SenseMI technology or Sense Machine Intelligence. This overarching umbrella has five sub-groups called Pure Power, Precision Boost, Extended Frequency Range, Neural Net Prediction, and Smart Prefetch. Lets take a look at each in as much depth as we can right now, today’s disclosure wasn’t the deep dive, that comes later.
Pure Power in one graph
Pure Power is the new term for DVFS with a bunch of new features thrown in, mostly for going beyond the core itself. For those familiar with AMD’s PowerTune 2.0, think of this as an extension of the ideas from that generation. The new Pure Power is a closed loop that involves Infinity Fabric now too, the umbrella term for the interconnect on Ryzen and other AMD IP. Since the IF touches just about every part of the system in a very granular way, a closed loop power management based on it also touches just about everything on the CPU. More on this technology when we can, it too is pretty neato and encompasses words like Hypertransport+.
Looking at the Pure Power curve above, it allows Ryzen to use less energy at the same frequency as ‘unoptimized’ silicon, IE more for less. Precision Boost does the opposite or at least takes a 90 degree turn (Turn the orange arrow counterclockwise) and doesn’t drop the power, it raises the frequency. As is the norm with modern silicon, PB does all its clock scaling on the fly but instead of the normal x86 100MHz increments, it is able to adjust in 25MHz steps. More granular clocks are a good thing and Ryzen is more granular. Another name for this is turbo but call it what you will.
Extended Frequency Range (XFR) is a nod to the overclocking set and acts very similarly to the way Intel does it’s overboosting where thermal headroom permits. The short version of how this works is that the chassis and heatsink of a CPU has a thermal storage capacity. If you are running under TDP, the heatsink can briefly take more energy than it can shed until a thermal equilibrium is reached. At that point the heat source has to lower its output or temperature will rise above tolerable levels.
This is the long way of saying the heatsink will absorb more energy than it can shed, but only for a little while. If you know the energy storage capacity of your cooler, you can use this to go above recommended power outputs until you ‘fill the battery’. XFR is this overboost in marketing form, but more importantly it allows the fringe overclockers to not be hobbled by power management. If you have a thermal solution that can pull off lots of energy, think LN2 pots or other exotic coolers, XFR should allow users to boost way beyond official turbo numbers. How well this works in practice remains to be seen, but it is a refreshing change from Intel’s knifing of the enthusiast community. Fingers crossed on this one people, it could be good.
AMD Neural Net Prediction
Neural Net Prediction and Smart Prefetch are similar in function with the former choosing optimal execution resources and possibly warming up cold blocks to avoid waits, the latter fetches data needed by the CPU before it is needed. Neither seem to be anything that is not currently in use by almost every modern high-end core, but there could be a lot of nuance that matters. Until AMD gives out a decoder ring for marketing terms to technical features, we will just have to guess at the specs.
On the fetch side, AMD lists a few neat tricks starting with decoupled branch prediction. The TLB is in the Branch Prediction (BP) pipe, 8 entry for L0, 64 for L1, and 512 for L2, plus each BTB entry can have two branches associated. There is also an Indirect Target Array and micro-tags for the instruction counter and op cache. That said Smart Prefetch sounds better but the technical version is much more informative.
The prediction side things is a little more vague. AMD calls it Scary Smart Prediction with an AI controlled decision-making process for execution. There isn’t much to say about this feature other than it sounds interesting but doesn’t map to any of the details SemiAccurate has heard about the Zen core. When AMD connects the dots fully, we will bring it to you. Since most predictors are well into the 99%+ range already, we are talking small differences but with tens to hundreds of cycles penalty for a miss, fractions of a percent matter. Time will again tell.
So that is about it for now, the big news being the specs and the Ryzen name. The CPUs SemiAccurate got to play with last week did not have some features turned on, specifically turbo wasn’t working and power management wasn’t close to final. Even with that, a pre-release 8-core Ryzen CPU performed on par with an 8-core Broadwell-E CPU running at the same 3.4GHz frequency. Better yet it did so at a lower energy draw for the system, but that was only on one benchmark. It will be interesting to see how a more polished, optimized Ryzen does when it debuts in a few months, competition is a good thing. At the moment, it looks like AMD has caught up to Intel with Ryzen, or at least are close.S|A
Latest posts by Charlie Demerjian (see all)
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018
- Intel can’t supply 14nm Xeons, HPE directly recommends AMD Epyc - Sep 7, 2018
- AMD reintroduces the Athlon name with two CPUs - Sep 6, 2018