As you read this, AMD's New Horizon event will be kicking off to reveal more details about the company's next generation of high-end CPUs for servers and desktops. We're way ahead of you, though. We got an early sneak peek at Ryzen silicon last week at the AMD Tech Summit in Sonoma, California. Yes, the Zen name that has shepherded this high-end chip through its nascency is no more. Instead, shipping Zen parts for the desktop—formerly code-name Summit Ridge—will carry the name Ryzen. Like the name of the recent ReLive software update, you can pronounce Ryzen a couple different ways, though AMD favored "rye-zen." Like a phoenix, or something.
Before we get into some of the nitty-gritty of Ryzen, we should first take a look at some of the new details AMD is sharing about its baby. The company confirmed that the highest-end Ryzen part will have eight cores and 16 threads running at a base clock of 3.4 GHz. Those cores will have 4MB of L2 and 16MB of L3 cache to play with, and the whole package will have an impressive 95W TDP. AMD wasn't ready to disclose boost clocks for Ryzen just yet, but it seemed confident that there was plenty of headroom in the tank.
We also got to see a sort of check-up on the health of Ryzen silicon. As it did during its preview event at IDF, AMD showed an eight-core, 16-thread Ryzen running a typical desktop workload—in this case, the Handbrake video-transcoding tool. This time, the company set up its Ryzen engineering sample to run at 3.4 GHz with no boost against an unhobbled Core i7-6900K. Recall that that the last time AMD ran a head-to-head test like this, it was against an i7-6900K limited to 3 GHz. It's also fun to note that the i7-6900K is a 140W TDP CPU, even if TDP isn't a universal or cross-comparable figure.
Although we don't know the precise details of either test system, the Ryzen PC finished AMD's sample workload a couple seconds ahead of the i7-6900K. Perhaps more encouragingly, AMD showed some power-draw numbers for this Ryzen sample under a full Blender load, and they were about on par with those of the Broadwell-E chip. That performance suggests that Ryzen's speed won't come with a high power bill attached, and that's heartening news.
AMD also revealed some interesting details about the under-the-hood features of Ryzen. Like other recent AMD chips, Ryzen CPUs will have a network of thermal and voltage sensors scattered across the die that provide a central processor with real-time information about the chip's operating conditions. Both Bristol Ridge APUs and Polaris GPUs already came with these sensor networks on board, but for easy reference, AMD now calls this network of monitoring hardware "SenseMI."
SenseMI will let a given Ryzen chip run at its most optimal point on the dynamic-voltage-and-frequency-scaling curve instead of baking in a predetermined safety margin that doesn't account for chip-to-chip variation. That adaptive tech could allow a chip to run at a given frequency with less voltage, improving efficiency—a feature that AMD will now call "Pure Power." SenseMI could also let a given chip extract its full potential frequency overhead when it dials in boost clocks—something that AMD will refer to as "Precision Boost."
SenseMI also underpins an intriguing new feature called "Extended Frequency Range," or XFR. SenseMI will monitor the effectiveness of the cooling solution that a builder installs on a Ryzen CPU using the Precision Boost feedback loop. Presumably, if one installs a Wraith cooler or similar heatsink, Ryzen chips will be able to hit their standard boost range. Put a monster tower cooler or a closed-loop liquid cooler atop a Ryzen CPU, though, and the chip can automatically exploit the extra thermal headroom to boost above its specified range. The more potent your cooling solution, the higher Precision Boost can push. Simple enough.
In an unusual step for a modern high-performance CPU architecture, AMD also discussed some intimate details of the Ryzen branch predictor. The company says it's using a neural-network-powered prediction algorithm in its latest CPUs. While that description may sound like marketing fluff—effective branch predictors are already learning systems, and neural networks are a hot topic right now—there may be more to it than latching onto a trendy term. AMD senior fellow Mike Clark told The Register that Ryzen uses a hashed-perceptron algorithm at the Hot Chips conference earlier this year. While a perceptron may be a basic neural network, it's still a neural network.
A quick Google suggests the idea of a perceptron-powered branch predictor is nothing new in chip design, but those types of predictors do appear to deliver extremely accurate performance. That's good news for any CPU. My conversations with AMD employees suggest we'll learn more about this topic in future briefings, so we can probably stand down with the pitchforks for the moment. AMD also touts Zen's "smart" data prefetcher, although the company didn't give us any hints as to what it's doing to improve this critical component of CPU performance. I suppose we'll need to wait for further briefings on that, as well.
If Ryzen can deliver on these promises, AMD thinks it has the potential to surf on some favorable trends in the world of gaming PCs. The company projects that the market for gaming hardware is in the midst of a 25% growth spurt from 2015 to 2018, and it expects the market for VR PCs specifically to grow from less than a million this year to over 10 million in 2020. The growth in popularity of eSports titles like Dota 2 and League of Legends, along with the exploding popularity of Twitch streaming, all suggest that new and existing gamers could be looking to upgrade to some new hardware. If those PCs are built around Ryzen CPUs and Radeon graphics cards, AMD could enjoy a much-needed shot in the arm for revenue growth.
Even though we have a little bit of a wait left before we can get our hands on Ryzen hardware, AMD continues to give us reasons for optimism regarding this CPU family and its performance. The company's demonstrations last week showed that Ryzen parts will most likely be competitive with Broadwell-E chips from both a performance and a performance-per-watt standpoint. The fact that AMD achieved that performance with engineering samples that are only running at their 3.4 GHz base clock suggests there may be even more performance yet to be tapped from these chips in less-multithreaded workloads, too. Features like XFR promise perhaps even greater performance rewards for enthusiasts that plan on using Ryzen with potent coolers, too—all without the headache of manual overclocking. We're eager to see just how all these promises play out when Ryzen CPUs debut sometime in the first quarter of next year.