AMD’s Zen CPU is now called Ryzen, and it might actually challenge Intel
- ️Tue Dec 13 2016
Proper SMT, faster caches, 40% higher IPC. Does AMD have a monster on its hands?
The AMD Zen CPU architecture is now called Ryzen (pronounced rye-zen, not rizen). Perhaps more importantly, though, as we creep towards Ryzen's promised Q1 2017 release date, AMD has finally revealed some solid specs for a chip based on Ryzen cores.
The first Ryzen-based part will be the Summit Ridge family of high-end desktop chips. Summit Ridge chips will feature an 8-core 16-thread CPU—which is a true 8-core chip using simultaneous multithreading instead of the much maligned clustered multithreading of Bulldozer—with the top-end part sporting a base clock of 3.4GHz. While that isn't as speedy as Intel's latest quad-core desktop parts (the i7-6700K has a base clock of 4GHz), it is more than competitive with Intel's 8C/16T Broadwell-E processor, which has a base clock of 3.2GHz.
Of course, clock speed isn't everything (otherwise, we'd still all be rocking Pentium 4 processors), but there was a concern that Ryzen's clock speeds would be way off the mark, somewhat negating the promised 40 percent improvement to instructions-per-clock (IPC). That AMD is hitting good clock speeds—and is doing it at under 100W TDP, far below the 140W TDP of Broadwell-E—is extremely promising.
Moreover, AMD claims to have plenty of processing power left on tap, thanks to a host of efficiency tweaks it's dubbed SenseMI (Machine Intelligence). The first is "Pure Power," which is a set of temperature, clock speed, and voltage sensors that promise more efficient power delivery to the CPU. "Precision Boost" takes that information and uses it to adjust the clock speed on-the-fly "without halts or queue drains" in small 25MHz increments.
This ties into "Extended Frequency Range" (XFR), which AMD claims will boost the clock speed of Ryzen outside of the (as yet unspecified) typical range if a user has suitably robust cooling. Those running big air coolers, all-in-one liquid coolers, or full watercooling loops will apparently see a big improvement compared to a stock cooler. The process is entirely automated, so even those without any overclocking skills can gain a performance boost. That said, fully manual overclocking is still supported for power users.
The final two tweaks—"Neural Net Prediction" and "Smart Prefetch"—help to shuffle data through the CPU more effectively. AMD claims there's a "true artificial network" and "learning algorithms" inside Ryzen that are able to predict future decisions, pre-load instructions, and learn application data access patterns in order to improve performance. That's aided by a beefy 4MB of L2 cache and 16MB of L3 cache.
Ryzen will be released alongside a new chipset, dubbed AM4, which will support many things that Intel users take for granted, including DDR4 memory, PCIe Gen3, USB 3.1 Gen2, and NVMe storage. AMD says that AM4 is designed to "run multiple processors over multiple generations," so here's hoping they stick a few forward-looking technologies in there to keep it competitive over the years.
There's still no word on a price for Ryzen or an exact release date, but if it's anything close to the price of Intel's enthusiast desktop parts such as the Core i7-6700K, AMD could have a hit on its hands. Intel has been happily gouging power users with its E-series processors for years now (the 8-core i7-6900K costs a whopping £1000/$1050).
If AMD can bring 8-core chips to the mainstream, and can close the gap on Intel's single-threaded performance superiority, the CPU market might get the kick up the arse it sorely needs.
Read on for the original high-level unveil of AMD Zen...
This is not a Zen chip. I just included it because, well, the original AMD Athlon holds a special place in my heart. It's also the last time that AMD challenged Intel for the performance crown. Click through to the rest of the Zen slide deck. InF!
Updated, August 24: AMD has provided a few more details of the new Zen CPU architecture at the Hot Chips symposium, most notably that Zen's power envelope has been pushed below 100W TDP at the top end. That's significantly less than the 140W of the Intel Broadwell-E CPU that AMD benchmarked Zen against, although—given how little we know about the overall Zen platform at this point—it'd be wise to take AMD's claims with a pinch of salt.
Unlike previous AMD architectures, which favoured pure performance, Zen aggressively targets efficiency. "Power had been intersecting more at the backend of the design, not at the beginning," explained AMD's Mike Clark, "We needed power analysis in the architecture from day one...a derivative was not going to work to get where we needed."
AMD says the much-touted "40 percent higher IPC" over the Excavator core came from three design goals: core, cache, and power. For the core, AMD made everything bigger and wider, introducing a micro-op cache (something Intel has been using for some time), as well as a larger dispatch, larger retire, larger schedulers, and better branch prediction. On the cache side there's a faster prefetch, while L1 and L2 bandwidth has been doubled, and L3 more then quadrupled. Full details on the improvements are in the slides below.
AMD also shared more details on Zen's design, confirming that it features truly independent cores—versus the module-based design of Bulldozer that shared several resources between cores—as well as support for SMT (simultaneous multithreading) similar to current Intel processors. Unfortunately, the juicier Zen details—clock speeds, platform, and products—won't be disclosed until later in the year.
Original story
AMD's new Zen CPU architecture has been officially delayed until "early 2017." The first Zen chips, which will be produced on a 14nm FinFET process, had originally been expected sometime in Q3 or Q4 2016.
At an event in San Francisco AMD also revealed a few more low-level details of Zen's architecture—and in a multithreaded Blender rendering demo showed that an 8-core/16-thread "Summit Ridge" Zen CPU outperformed an 8C/16T Broadwell-E CPU (presumably the Core i7-6900K) at the same clockspeed.
AMD showed off a dual-CPU Windows server setup using the 32-core/64-thread "Naples" enterprise-oriented Zen CPU at the same event, but didn't provide any kind of performance figures.
Architecture-wise, we already know that Zen is a brand new design that corrects most of the mistakes made with AMD's line of Bulldozer-based CPU cores. Clustered multithreading (CMT) is out; simultaneous multithreading (SMT), which has been used to great effect by Intel, is in. We also knew that AMD planned to seriously beef up Zen's caching system (slow caches were one of the big reasons behind Bulldozer's poor performance).
Today, without getting into too much detail, AMD confirmed that desktop Zen parts will have a "new cache memory hierarchy with 8MB of L3 cache" with an enhanced pre-fetcher, a large unified L2 cache, and separate low-latency L1 instruction and data caches.
AMD had previously said it was targeting a 40 percent uplift in instructions-per-clock (IPC) performance over Excavator, the last core in the Bulldozer family; now the company has provided some harder figures on how it plans to get there: Zen has "enhanced branch prediction," plus "1.75x instruction scheduler window" width and "1.5x issue width and execution resources" over Excavator.
AMD also briefly discussed some design tweaks aimed at improving power consumption—"aggressive clock gating with multi-level regions"—but I imagine they will pale in comparison to the gains rendered by manufacturing Zen on a 14nm FinFET process (Excavator was stranded at planar 28nm bulk silicon).
AMD isn't yet discussing Zen's target clockspeed, but some leaked benchmarks last week purported to show an 8C/16T Zen engineering sample clocked at 2.8GHz base/3.2GHz turbo—which is pretty impressive when you consider that engineering samples are often clocked significantly lower than retail parts. I don't want to get my hopes up just yet, but the benchmarks did appear to show pretty decent performance for the Zen chip.
It's also important to note that Zen will be AMD's only x86 CPU architecture moving forward: the company is doing away with its current bifurcated lineup of mobile (Bobcat, Jaguar, Puma) and desktop (Bulldozer, Piledriver, Excavator) CPU architectures. Much like with Intel's recent chips, it's safe to assume that AMD will have made some mobile-first design decisions with Zen.
We expect AMD will offer journalists a deeper dive into Zen in the next few weeks; we'll bring you more info when we have it.
Sebastian is the editor of Ars Technica UK. He usually writes about low-level hardware, software, and transport, but it is emerging science and the future of technology that really get him excited.
