AMD’s Ryzen processors have indubitably reshaped the mainstream PC in the year since their release. Four-core, eight-thread CPUs reigned in those systems for the better part of eight years, but first-generation Ryzen parts brought core and thread counts typical of high-end desktop chips within range of the average builder for the first time.
The Zen microarchitecture has since proven itself worthy in a broad range of gaming and productivity tasks, and enthusiast-friendly perks like capable stock coolers, universally unlocked multipliers, and soldered heat spreaders have won the hearts of many a DIY builder. Some fundamental disadvantages of the Zen core versus Intel's Skylake architecture, like SIMD units that provide half the potential throughput of the blue team's cores, will require major architectural changes if AMD chooses to address them. Massive re-architecting like that will likely need to wait for the move to 7-nm-class process technologies and the bounty of extra transistors they could offer.
AMD has been listening to Ryzen owners over the past year for more pragmatic changes it can implement to make its products more competitive, however, and it came up with a few relatively straightforward fixes. First and foremost, enthusiasts have demanded a better-behaved and lower-latency integrated memory controller for use with high-speed DDR4 RAM, important characteristics for feeding as many as eight cores with dual-channel memory. Overclockers have pined for more potential from Ryzen CPUs, many of which top out at 3.8 GHz to 4 GHz all-core speeds. Folks who don’t overclock want higher stock clock speeds, as well. Finally, AMD felt it could reduce access latencies at the various levels of the processor’s cache and memory hierarchy to improve performance.
At least on paper, AMD has ticked off every box on that wish list with the Zen+ microarchitecture that underpins second-generation Ryzen CPUs. The company says the die size, transistor count, and fundamental logic of the Zen core remains unchanged in the transition from GlobalFoundries’ 14-nm FinFET process to its 12LP process.
Instead, the company is reaping the benefits of the better transistors available from that process to improve performance on critical paths of the chip. In sum, that means higher peak clock speeds, lower cache latencies, a more robust memory controller, and lower voltage requirements for the same performance.
With the improvements of 12LP in mind, some might find it odd to see that the TDP of the top-end Ryzen 7 2700X has actually increased 10 W, to 105 W. Part of this change is because of second-gen Ryzen’s Precision Boost 2 dynamic-voltage-and-frequency-scaling technology. More on that chip's specs in a second.
Instead of first-generation Ryzen CPUs’ simple concept of single-core, two-core, and all-core boost speeds, Precision Boost 2 allows AMD’s SenseMI power- and temperature-monitoring tech to vary second-gen Ryzen parts’ all-core clock speeds more or less linearly, from one to as many as eight loaded cores and one to sixteen threads.
Precision Boost 2 means a second-generation Ryzen chip can take full advantage of the power and thermal headroom available to it, and AMD says it’s intentionally allowing Ryzen second-generation parts to burn more power under load to fully realize Precision Boost 2’s potential in cases where the original Precision Boost would have had to fall back to an all-core boost speed. To be clear, higher power usage alone should not be seen as a regression, at least in theory. This decision makes perfect sense if it allows Ryzen second-gen parts to consume less energy over the course of a task thanks to higher performance-per-watt.
The move to a more linear dynamic voltage and frequency scaling curve means that the behavior of AMD’s Extended Frequency Range (XFR) technology is changing, as well. XFR 2 does away with the idea of a fixed frequency increase across single-core and all-core workloads, as seen on all first-generation Ryzen products to some degree. Instead, XFR 2 works more like the Mobile XFR feature we first saw on Raven Ridge mobile chips.
The peak Precision Boost 2 speed on any Ryzen second-generation part will still rise (by about 50 MHz, in our experience) if better-quality cooling is installed, but the bigger change is that users should see higher sustained frequencies on multi-core workloads with a more capable cooler, and hence better performance in those tasks. That sustained clock-speed improvement is the second key way that AMD is improving performance with its second generation of Ryzen parts.
All second-generation Ryzen CPUs enjoy a more robust memory controller than first-generation Ryzens, too. As it does with Raven Ridge desktop parts, AMD rates second-gen Ryzens for DDR4-2933 support from single-rank, two-DIMM memory configurations (although only from motherboards with six PCB layers, oddly enough). Using more DIMMs or dual-rank memory will cause stock memory speeds to drop off, just as with first-generation Ryzen parts. Even with that in mind, our hands-on testing with AMD's demo X470 systems at an event in New York suggested that overclocked memory speeds in the range of 3600 MT/s with two DIMMs could be achievable with some care. That's a major improvement over first-gen Ryzens, where speeds greater than 3200 MT/s proved difficult to reliably achieve without exacting choices of memory kits and motherboards.