The AMD Ryzen CPUs represent the first flush of the red team’s Zen processor architecture. They’re also the first all-new CPUs AMD have released in many a year and, with the computational performance on show, are delivering genuine competition to Intel’s CPU dominance.
The new Zen architecture could change the CPU landscape when it launches this year, but right now these are the best CPUs for gaming.
We've spent a long while with the new AMD platform and in our Ryzen 7 1800X review we've been impressed by the number-crunching capabilities of the flagship Zen CPU, but not so much by its gaming performance.
The fact AMD have closed the gaming gap on Intel is important though, because even if they haven't gained parity, when we get to the mainstream Ryzen 5 processors things could really shake up in the PC gaming world.
With any new platform though there will be issues. Aside from gaming the memory performance isn't great - the platform lacks quad-channel support and struggles above 2,667MHz on many kits. It's also not a particularly inspiring overclocker, partly thanks to the fact it's a hot and power-hungry processor.
But do you know your branch prediction from your retire queue? Your FP register file from your integer physical register file? Nope, us neither, but here's what the new AMD Zen architecture will mean for their next generation of processor and potentially for your next gaming PC.
In short, the AMD Zen CPU architecture represents the best chance for the red team to close the gap on Intel, offering PC users a genuine computational alternative for our desktop rigs.
If you want to jump right to the good stuff click on the quick links below to get there.
The main goal for the upcoming AMD Zen CPUs is to offer a credible alternative to the Intel hegemony existing around high-end processors right now, and they absolutely have to be priced accordingly.
The top-tier eight core Ryzen CPUs has been confirmed now. We're looking at less than $500 for the Ryzen R7 1800X, the new eight-core, 16-thread AMD superchip.
The eight-core, 16-thread R7 1800X has a suggested retail price of $499, the slightly slower R7 1700X is $399 and the bottom R7 1700 is $329.
Interestingly the two 'X' chips are also listed as 'WOF' parts, which translates as 'without fan' while the standard R7 1700 is listed as 'PIB' or 'processor in box.' That would indicate the R7 1800X and R7 1700X are the more overclocking-focused versions of the Ryzen CPU range as they're expected to be paired with third-party coolers, while the 65W R7 1700 comes with a standard AMD Wraith Spire chip chiller because it's not expected to be pushed beyond its normal frequency boundaries.
Though all Ryzen chips are overclockable if you've got them dropped into a compatible motherboard, thanks to the universally unlocked multipliers. There are also reportedly 'X' series processors in all three tiers of Ryzen chip, delivering overclockable CPUs right down to the bottom of their stack.
That's an impressively low initial price if that's in any way true, with the cut-price eight-core chip being a great way of offering an incentive to people on an old Core i5 looking for an upgrade.
AMD have then gone back to the drawing board for their new x86 AMD processor architecture with Zen offering the promise of clawing back performance, and potentially market share, from their big blue Intel rivals. This time though their new CPU design represents more of a balancing act; nailing improved single-core throughput for existing game engines as well as catering to the multi-threaded approach newer graphics APIs, like DirectX 12 and Vulkan, are finally offering.
AMD’s CEO, Lisa Su, introduced the ‘guts of Ryzen’ at the New Horizon event in December, detailing the five key new features that make it the fastest processors architecture they’ve ever created. The five new features in question are: Neural Net Prediction, Smart Prefetch, Pure Power, Precision Boost and Extended Frequency Range. Altogether they’re known as AMD SenseMI Technology.
These allow the Ryzen CPU to “sense and adapt to its environment, which makes it an extremely intelligent processor, capable of learning on the fly.” According to Su that gives the Ryzen silicon ‘true machine intelligence’ which accounts for around 25% of the performance uplift the Zen architecture is promising over its forebears.
The Neural Net Prediction part of the chip allows it to self train and preload the right instructions to run faster. Once it figures out what the right instructions are the Smart Prefetch feature learns to anticipate the data an app needs and aims to have it ready before it’s actually needed.
The Ryzen processors also house hundreds of sensors to track and locally optimise the silicon to enable it to run at its fastest while still maintaining the lowest possible power draw. These sensors feed into the Pure Power and Precision Boost features, which deliver a huge amount of control to the CPU itself.
Pure Power and Precision Boost work together to “dial up the frequency and dial down the power on each part of the chip, independently, in milliseconds.” This performance optimisation is aimed at making the changes to power and frequency as quickly and as unobtrusively as possible.
This was something AMD’s Mike Clark, Zen’s lead architect, spoke about before. “We had the power focus from the beginning of the project. We do aggressive clock gating with multi-level regions to make sure we’re gating off logic when we’re not using it."
The final piece of the SenseMI puzzle is the Extended Frequency Range. This is essentially analogous to the GPU boost you get on the latest Nvidia graphics cards where the upper limits of the clockspeed aren’t rigidly defined parameters but are almost entirely based on the thermal performance of the machine’s cooling setup. If you’re rocking liquid cooling you’ll get a higher upper limit to the operating frequency of your Ryzen CPU.
It’s a fully automated solution and basically allows for processor overclocking without any user input outside of deciding what cooling to drop onto their chip. It looks like this is going to be a feature of all the Ryzen processors, which makes sense given AMD have confirmed that there’s going to be no locked multipliers on any of their new CPUs, opening them all up to overclocking - should you have the right motherboard anyway.
Outside of the new features there are some silicon advances which are also delivering performance boosts to the new Zen architecture. Before last year's Hot Chips symposium in Silicon Valley AMD gave us a sneak peak at what's under the Zen hood, and it's a mighty complex beast. Mike Clark, the lead architect on the Zen x86 core, took us through a deep dive on the heart of the new microarchitecture detailing exactly how they managed to nail their target of a 40% instructions per clock increase over their last generation of processor.
“We’ve taken a brand new core engine - tuned for performance - combined it with a high-bandwidth, low-latency cache system," explains Clark. "We've thrown SMT on top of it so we can get high throughput when we need to and the used the new 14nm FinFET technology to also get low-power and scale the design across a number of SoCs from client to servers.”
It's those three pillars of improved core design, better cache and a focus on lower power which have allowed AMD to hit their target. They've been working on the first two for decades, but targeting power contraints normally isn't something that comes until later on in AMD's silicon designs.
"We knew, for this design, we needed to have power analysis going on in the microarchitecture from day one and be choosing the right features to still deliver the IPC increase," says Clark, "but be able to maintain equivalent power to our previous generation."
One of the main things which has given the core engine its boost is the move towards an SMT-based core, giving it an approach to multi-threading which is much closer to Intel's HyperThreading. They've also improved the brand prediction allowing for more instructions to be completed per clock cycle.
"We’re both a wider machine - we can send more instructions into the machine per cycle, 6 vs. 4," says Clark. "And then we’re a deeper machine, meaning we can hold more instructions and we can look further into the program for parallel work that we can execute."
That means the Zen architecture is able to both chew through more processing tasks as well as dig deeper into the queue of tasks it has yet to finish, picking off the ones it can tackle concurrently.
But all that counts for naught if they can't keep the high-performance system fed with data. To do that you need a good cache system and AMD have doubled the bandwidth available to its level one and level 2 cache while improving the level 3 cache bandwidth by five times what was available on its previous cache design.
"The combination of that better core engine and that better cache system allow us to get the 40% IPC uplift, explains Clark.
"Overall we’ve been increasing our power design methodology on the physical design side, all through the ‘Dozer generation," says Clark. "And we were able to leverage all that learning - with new learning - and all that in combination has lead to us being able to achieve the IPC uplift and hold power steady or neutral with the previous generation.”
Lisa Su revealed at their recent investor earnings call the 'multi-generational roadmap' AMD was working towards for the Ryzen successors. Currently they're just being referred to as Zen 2 and Zen 3, but Su explained that they would involve the architecture shrinking down to the 7nm node.
"From our standpoint process technology, we ramped 16nm and 14nm really well last year and into this year," explained Su. "We are actually in the process of developing now in 7nm and we think the 7nm foundry roadmaps that are available are very competitive and will ensure that we have a strong multi-generational roadmap."
The initial AMD Ryzen CPUs are the eight-core, 16-thread beasts. These are the top-end processors previously known as Summit Ridge, though the Ryzen range of chips will cater for many different budgets and needs.
"Zen, for AMD, is not just one product or two products," explained AMD CEO, Lisa Su. "It's really a family of products that will roll out."
The chief architect for Zen, Mike Clark, echoed his CEO's statement. “Zen is not a destination," he said. "it’s a starting point. AMD is back, but we’re also back to stay. We’re going to continue to push high-performance compute going forward in our next generation - we’re totally committed to high performance compute.”
On the Q4 2016 earnings call Su discussed the 'multi-generational roadmap' that AMD is working to for the Zen family, including the 'Zen 2 and Zen 3 follow on.' As well as updating the archiecture the Ryzen successors will also support production improvements too.
"From our standpoint process technology, we ramped 16nm and 14nm really well last year and into this year," explained Su on the call. "We are actually in the process of developing now in 7nm and we think the 7nm foundry roadmaps that are available are very competitive and will ensure that we have a strong multi-generational roadmap."
We’re expecting hex and quad-core parts eventually to go along with the octa-core processors. They will also support Simultaneous Multi-Threading (SMT), a technology analogous to Intel’s HyperThreading, allowing each discrete CPU core to process two threads concurrently.
There are rumours, however, that AMD will be releasing a lower-tier quad-core Ryzen processor without their SMT feature enabled, leaving it as a four-core, four-thread chip. That would make sense given that it looks unlikely that AMD will be making any Ryzen processors with fewer than four cores.
The Zen architecture has been built in four-core complexes, so the octa-core chip will sport a pair of these core complexes.
It’s a different situation to AMD’s previous architecture in which their Bulldozer CPUs (as well as the later, iterative architectures toting the similarly heavy-plant-based codenames of Piledriver and Excavator) were built using nominally dual-core modules, where each ‘core’ shares an unhealthy chunk of logic between them. The quad-core modules in Zen have discrete logic in each core, sharing only access to the L3 cache.
A quad-core Bulldozer chip uses a dual-module design and an octa-core uses four modules. AMD’s eight-core chips, like the FX-8370, are effectively quad-core CPUs using a slightly hobbled form of HyperThreading. With the Summit Ridge CPUs, on the other hand, each promised Zen core is going to be fully discrete with its very own, jealously-guarded silicon.
In terms of the process AMD is using for their new architecture they are going with the same 14nm FinFET lithography already used in their Polaris-based graphics cards. The FinFET design uses essentially the same Tri-Gate transistor layout Intel has been building into their processors since the 22nm shift for their Ivy Bridge platform. It’s a 3D transistor design used to get around the electrical leakage which occurs in planar transistors when components are shrunk down to such microscopic levels.
In terms of the power draw the top octa-core Ryzen is a 95W chip with the R7 1700 only pulling a 65 TDP. Their previous eight-core processors came in at 125W, though they also had a couple of 220W CPUs too, but we doubt they'll want to go down that route again with Zen...unless they want to get a 5-6GHz chip out down the line.
But even more interesting on that front is the idle power performance of these engineering samples. They are reported to be able to deliver idle power of just 2.5W and 5W for the quad and octa-core CPUs respectively; both able to scale down their clockspeeds to an operating frequency of just 550MHz.
And we thought the Skylake architecture was impressive when the 6700K showed it could drop down to 800MHz with a 3.5W idle power draw.
In terms of the platform the Zen processors drop into, AMD is simplifying things by unifying both their CPU and APU sockets. The new AM4 socket means that, with a single motherboard, you can be assured of compatibility with all of AMD’s Zen silicon going forward.
There is a much greater pin density on the new socket compared to the outgoing AM3+ design. The last connection had just 942 pins while the AM4 socket offers a full 1331 potential connections. Not all AM4 chips will use the full complement of pins with the Bristol Ridge APUs not needing all the PCIe lanes or memory options on offer.
This does mean the Zen chips look practically identical to existing AMD processors. Speaking as someone who regularly received loose AMD CPUs in barely padded envelopes, the prospect of bent CPU pins is looming large again with Zen.
There are different levels of AM4 motherboard, however, catering for both the high-performance Zen crowd as well as the more mainstream APU guys. There will be three discrete chipsets for the new AMD platform, with the X370 catering for the high-end systems, with their eight-core, 16-thread Zen processors.
The X370 is supports 'overclocking+' which brings the deepest level of system control to the Zen platform, while also taking advantage of the touted 36 PCIe 3.0 lanes available in the Summit Ridge processors to deliver full x16 bandwidth to a twin CrossFireX or SLI multi-GPU array.
Below that are the mainstream-oriented B350 and entry-level A320. Broadly speaking the three tiered range ought to be roughly analogous to Intel's Z270, H270 and H110 chipsets for the new Kaby Lake CPU lineup.
The mainstream B350 still supports a slightly lower level of CPU overclocking, but it doesn't look like it will be natively supporting multi-GPU systems at launch. The low-end A320 then is most basic of AM4 chipsets and won't come with either dedicated overclocking or multi-GPU support.
There will also be a range of dedicated AM4 small form factor chipsets; the X300 will cater to the micro machine fans, allowing them to craft powerful rigs with a tiny footprint. There were a whole load of AM4 boards on display out and about at the CES event at the start of the year.
Specs are one thing, but it’s the performance of the Zen CPU architecture that will make or break the new design, and potentially even AMD itself. They’ve tried to be fancy and second guess the future of the PC hardware market before - introducing an affordable, multi-threaded architecture with Bulldozer which was left behind while the rest of the industry stuck resolutely to relying on the performance of a single processing thread.
You may say gaming is more about graphics power these days, and to some extent you would be correct, but the fact remains you’ll get more performance from your graphics card if it’s plumbed into an Intel-based system than an AMD one, no matter who makes your GPU. Sadly AMD CPUs hold back your graphics hardware, which means high-end gaming systems will always steer clear of the red team’s existing processor designs.
Sadly that's still the case with our initial testing of the Ryzen 7 1800X. The gaming gap has been closed, but there is still a performance delta between a GPU running via an Intel chip and the same one plumbed into a Ryzen CPU.
That performance difference is mainly evident in older games, however, though it's still shown to a lesser extent in more modern DirectX 12 titles too. It's possible performance optimisation for next-gen games might be able to take advantage of the multi-threaded performance on offer within the Ryzen CPU family, but improvements like that are tougher to get with software updates on processors than graphics cards.
But the multi-threaded performance is seriously impressive. So if you're a content producer more than a gamer, and one who doesn't mind losing a little GPU performance in exchange for a great-value number-cruncher, then Ryzen could well be for you.
Follow the quick links below to get to the relevant section.