AMD Ryzen 7nm CPUs May Not Hit Maximum Boost Frequency on All Cores
AMD’s third-generation Ryzen processors have been a massive hit for the company by all reports, with excellent performance relative to Intel’s Core CPUs. There have, however, been a few questions around yield, overclocking, and boost frequencies. CPU overclocks on Ryzen are notably low, and some enthusiasts have noticed a limited number of cores on their CPUs hit the targeted boost frequencies.
Tom’s Hardware has done a significant deep-dive into this issue and came away with a number of key findings. In the past, AMD CPUs were capable of hitting their top-rated boost frequencies on any CPU cores. Intel chips are designed similarly. With Ryzen 3000, apparently only up to one core needs to be capable of hitting its maximum or near-maximum boost frequency. The scheduler updates baked into Windows 10 were said to speed power state transitions (which they do), but they also assign workloads specifically to the fastest cores capable of hitting a given clock.
These findings may explain why all-core overclocking headroom on these new Ryzen 7 processors is so low. On the Ryzen 7 3600X, only one CPU core proved capable of hitting 4.35GHz, for example, with other cores on the same chip boosting to 75-100MHz lower. AMD has not released exact specs for what frequencies its cores need to be able to hit to satisfy its own internal metrics for launch, which means we don’t really “know” which frequencies these CPU cores will operate at. This is definitely a change from previous parts, where all cores could be more-or-less assumed to be capable of hitting the same boost frequencies, and it may have implications for overclockers but it doesn’t really change my opinion on AMD’s 7nm Ryzen CPUs. If anything, I suspect it’s a harbinger of where the industry is headed in the future.
Building Faster Silicon Today Means Working Smarter, Not Harder
One of the topics I’ve covered repeatedly at ExtremeTech is the difficulty of scaling either IPC (instructions-per-clock, a measure of CPU efficiency) or clock speed as process technology continues to shrink. From December 2018 – June 2019, I wrote a number of articles pushing back against various AMD fans who insisted the company would use 7nm to make huge clock leaps above Intel. When we met AMD at E3 2019, company engineers told us point-blank that they expected no clock improvements at 7nm whatsoever initially, and were very pleased to be able to improve clocks modestly in the final design.
One of the major difficulties semiconductor foundries are dealing with on 7nm and lower nodes is increased variability. Increased variation in parts means the chance of getting a wider “spread” on which cores are capable of running at specific frequency and voltage settings. AMD adapted Adaptive Voltage and Frequency Scaling back with Carrizo in part because AVFS can be used to control for process variation by more precisely matching CPU internal voltages with the specific requirements of the processor. Working with Microsoft to ensure Windows runs workloads on the highest-clocked CPU core isn’t just a good idea; it’s going to be a necessary method of extracting maximum performance in the future.
Intel’s decision to introduce Turbo Boost with Sandy Bridge back in 2011 was one of the smartest moves the company ever made. Intel’s engineers accurately forecast it was going to become increasingly difficult to guarantee maximum clocks under all circumstances. There’s no arguing what AMD is doing here represents a fundamental shift from the company’s approach in years past, but it’s one I strongly believe we’re going to see more companies embracing in the future. Higher silicon variability is going to demand a response from software. The entire reason the industry has shifted towards chiplets is that building entire dies on 7nm is seen as a fool’s errand, given the way cost scales with large die sizes, as the slide below shows.
Why move to AVFS? To decrease variability. Why move to chiplets? To cut manufacturing costs and improve yields overall. Why change Windows scheduling to be aware of per-core boost frequencies? To ensure end-users receive the full measure of performance they pay for. While it’s true Intel CPUs may be able to hit boost frequencies on any core, that doesn’t mean this state of affairs was objectively better for the end-user. Windows’ typical core-shuffling is not some unalloyed good, a fact Paul Alcorn notes in his article. “Typically we would see more interspersed frequency jumps among cores,” Alcorn writes, “largely due to the Windows scheduler’s irritating and seemingly irrational tendency to allocate threads into different cores on a whim.” Meanwhile, we know the boost frequency Intel CPUs will practically hold still depends directly on how many CPU cores are being loaded. The fact that all CPU cores can reach higher clocks does not necessarily benefit the end-user in any way unless said user is overclocking and statistically, most computer users don’t.
But because it’s getting harder to eke out frequency boosts and performance improvements, manufacturers are investing in technologies that tap the reservoir of performance in any given CPU solely for their own use. This is why high-end overclocking is slowly dying and has been for at least the past seven years. AMD and Intel are getting better and better at making limited frequency headroom in their products available to end-users without overclocking because overclocking these CPUs in the conventional fashion blows their power curve out so severely. It wouldn’t surprise me to discover AMD went with this method of clocking because it improved performance more at lower power compared with launching lower-clocked chips with a more conventional all-core boost arrangement.
The old rules of process node transitions and silicon designs have changed. That’s the bottom line. I’m confident we’ll see Intel deploying advanced tactics of its own to deal with these concerns in the future because there is zero evidence to suggest these issues are unique to AMD or TSMC. AMD’s adoption of AVFS, the rising use of chiplets across the industry, the lower expected clocks at 7nm that were turned into a small gain thanks to clever engineering all of these issues point in the same direction. Companies will undoubtedly develop their own particular solutions, but everyone is grappling with the same fundamental set of problems.
Good Communication is Key
AMD, to its credit, did tell users they needed to be running the latest chipset driver and the Windows 1903 update to take advantage of the new scheduler. Implied in that rhetoric was not doing so would prevent you from seeing the full impact of third-generation Ryzen’s improved performance. I do agree the company should have disclosed this new binning strategy to the technical press at E3, so we could detail it during the actual review.
But does this change my overall evaluation of third-generation Ryzen? No. Not in any fashion. The work THG has done to explore this issue is quite thorough, but based on the reading I’ve done on the evolution of process technology in modern manufacturing, I come down firmly on the side of this being a good thing. It’s the extension of the same trend that led ARM to invent big.Little namely, the idea that the OS needs to be more tightly coupled to the underlying hardware, with a greater awareness of what CPU cores should be used for which workloads in order to maximize performance and minimize power consumption.
According to AMD, roughly 25 percent of the performance improvements of the past decade have come from better compilers and improved power management. That percentage will likely be even larger 10 years from now. Power consumption at both idle and load is now the largest enemy of improved silicon performance, and variability in silicon process is a major cause of power consumption. Improving performance in the future is going to rely on different tools than the ones we’ve used for the past few decades, and one of the likely consequences of that push is the end of overclocking. Manufacturers can’t afford to leave 10, 20, 30 percent performance margins on the table any longer. Those margins represent a significant percentage of the total improvements they can offer.
Do these findings have implications for the currently limited availability on the Ryzen 9 3900X? We don’t know. Certainly, it’s possible the two are connected and that AMD is having trouble getting yield on the chip. Ultimately, I stand by what I said in our article on AMD CPU availability earlier today we’ll give the company a little more time to get product into market and revisit the topic in a few more weeks. But the CPU’s performance is excellent. Its power consumption, particularly if paired with an X470 motherboard, is excellent. We’re still working on future Ryzen articles and have been working with these CPUs for several weeks. The performance and overall power characteristics are fundamentally strong, and while the THG findings are quite interesting for what they say about AMD’s overall strategy going forward and what I believe is the general increase in variability in semiconductors as a whole, I view them as broadly confirming the direction the industry is moving in. Dealing with intrinsically higher silicon variability will be one of the major challenges of the 2020s.
I hesitate to bring Intel into this conversation at all, because we haven’t even seen the company’s latest iteration of its 10nm process yet, but it’s surely no accident the company’s upcoming mobile processors have sharply reduced maximum Turbo Boosts (4.1GHz for Ice Lake, compared to 4.8GHz for 14nm Whiskey Lake). Some of that may be explained by the wider graphics core that’s built into Gen 11, but Intel forecast from the beginning that 14nm++ would be a better node for high-frequency chips than its initial 10nm process. That doesn’t mean Intel has adopted AMD’s new clocking method, but it does show the company is grappling with some of these same issues around frequency, variation, and power consumption, and working to find its own ideal balance.