Overclocking Results Show We’re Hitting the Fundamental Limits of Silicon
Update (4/30/20): The formal unveiling of Intel’s 10th Generation Core i9 family is an excellent opportunity to revisit the points made in this November 2019 article. As of this writing — some five months later — Silicon Lottery is out of 9th Gen chips and waiting on Comet Lake CPUs to arrive. The handful of 7nm AMD CPUs show very similar patterns to what we identified back in November. A 3950X @ 4GHz is just $750, but an all-core 4.1GHz is $850 and a 4.15GHz chip is $999.
Now, with its 10th Gen Comet Lake, Intel has adopted strategies like die-lapping and adding copper to its IHS to improve thermal transfer off the core, while allowing much higher levels of power consumption. It’s not that there’s something wrong with the parts from either company — manufacturers increasingly have no additional firepower to leave on the table for enthusiasts to enjoy.
Original story below:
Silicon Lottery, a website that specializes in selling overclocked Intel and AMD parts, has some 9900KS chips available for sale. The company is offering a 9900KS verified at 5.1GHz for $749 and a 9900KS verified at 5.2GHz for $1199. What’s more interesting to us is the number of chips that qualify at each frequency. Thirty-one percent of Intel 9900KS chips can hit 5.1GHz, while just 3 percent can hit 5.2GHz. The 5.2GHz option was available earlier on 11/4 but is listed as sold-out as of this writing.
The 9900KS is an optimized variant of Intel’s 9900K. The 9900K is Intel’s current top-end CPU. Given the difficulties Intel has had moving to 10nm and the company’s need to maintain competitive standing against a newly-resurgent AMD, it’s safe to assume that Intel has optimized its 14nm++ process to within an inch of its life. The fact that Intel can ship a chip within ~4 percent of its apparent maximum clock in sufficient volume to launch it at all says good things about the company’s quality control and the state of its 14nm process line.
What I find interesting about the Silicon Lottery results is what they say (or said, as of November 2019) about the overall state of clock rates in high-performance desktop microprocessors. AMD is scarcely having an easier time of it. While new AGESA releases have improved overall clocking on 7nm chips, AMD’s engineers told us they were surprised to see clock improvements on the Ryzen 7 3000 family at all, because of the expected characteristics of the 7nm node.
AMD and Intel have continued to refine the clocking and thermal management systems they use and to squeeze more headroom out of silicon that they weren’t previously monetizing, but one of the results of this has been the gradual loss of high-end overclocking. Intel’s 10nm process is now in full production, giving us some idea of the trajectory of the node. Clocks on mobile parts have come down sharply compared with 14nm++. IPC improvements helped compensate for the loss in performance, but Intel still pushed TDPs up to 25W in some of the mobile CPU comparisons it did.
I think we can generally expect Intel to improve 10nm clocks with 10nm+ and 10nm++ when those nodes are ready. Similarly, AMD may be able to leverage TSMC’s 7nm node improvements for some small frequency gains itself. It’s even possible that both Intel and TSMC will clear away problems currently limiting them from hitting slightly higher CPU clocks. Intel’s 10nm has had severe growing pains and TSMC has never built big-core x86 processors like the Ryzen and Epyc chips it’s now shipping. I’m not trying to imply that CPU clocks have literally peaked at 5GHz and will never, ever improve. But the scope for gains past 5GHz looks limited indeed, and the 5.3GHz top frequency on Comet Lake doesn’t really change that.
The advent of machine learning, AI, and the IoT have collectively ensured that the broader computer industry will feel no pain from these shifts, but those of us who prized clock speed and single-threaded performance may have to find other aspects of computing to focus on long-term. The one architecture I’ve seen proposed as a replacement for CMOS is a spintronics approach Intel is researching. MESO — that’s the name of the new architecture — could open up new options as far as compute power density and efficiency. Both of those are critical goals in their own right, but so far, what we know about MESO suggests it would be more useful for low-power computing as opposed to pushing the high-power envelope, though it may have some utility in this respect in time. One of the frustrating things about being a high-performance computing fan these days is how few options for improving single-thread seem to exist.
This might seem a bit churlish to write in 2020. After all, we’ve seen more movement in the CPU market in the past 3 years, since AMD launched Ryzen, than in the previous six. Both AMD and Intel have made major changes to their product families and introduced new CPUs with higher performance and faster clocks. Density improvements at future nodes ensure both companies will be able to introduce CPUs with more cores than previous models, should they choose to do so. Will they be able to keep cranking the clocks up? That’s a very different question. The evidence thus far is not encouraging.