Something that has been interesting about the last handful of years in PC technology has been watching AMD claw back a steady and growing footing in the PC market for CPUs while Intel has gradually been slipping on a series of banana peels they threw in their own path.
With the first gen Ryzen launch in 2017, it wasn’t that bad yet – the CPUs were good but not enough to win in gaming compared to the 7th gen Core series from Intel, but in productivity workloads, AMD had pulled out a lead while maintaining a close-enough margin on gaming for all but the most single-threaded of games and those games where clock speed mattered more than core count – situations like high-refresh rate for FPSes. Second-gen Ryzen continued this trend, but offered a modest performance improvement and a lower price at the top-end. Last year’s third-generation Ryzen made waves for being a totally new to consumer-space chiplet design, allowing AMD to pack up to 16 cores into the same socket packaging they’d been using for Ryzen since launch, while offering better memory speed and compatability, much better single-threaded performance, which, clock-for-clock, has a slight lead on Intel which is only negated by the Intel parts clocking higher, and was the point at which AMD could finally claim to have silicon process superiority on Intel.
That last point has been clearly stuck in Intel’s craw the most, however. Intel’s oldest roadmaps for their manufacturing division had them on-track to have a 10 nanometer silicon process in mass production by 2016, used in desktop products, datacenter, and all over – a top-to-bottom product stack. Intel’s process development was and in many ways still is tightly integrated with their product development. Their CPU cores are developed explicitly for and optimized around the use of the manufacturing process in question. When 10nm was delayed, this caused Intel to keep the same designs in use for, at this point, 5 years. Skylake was their last truly new architecture, released in first iteration in 2015 as their second 14nm product, and all subsequent releases to date in the CPU market they’ve made have been different codenames spun on the same core design. For both datacenter and mainstream desktop CPUs, they’ve basically bolted on additional cores to keep pace with AMD and their customer demands, with the current Comet Lake 10th gen Core CPUs leading with a 10-core design, still built around the same old Skylake architecture.
To Intel’s credit, Skylake is a relatively good design. It is very potent and has actively benefited from the improvements and tweaks they’ve made to 14nm, as it now reliably clocks over 5 GHz such that Intel sells CPUs that offer boost clocks up to 5.3 GHz on a single core. However, the problem of not being able to sell a new core design makes a potent problem. Intel has a multi-generational problem with security, with multiple hardware-level security flaws that need hardened fixes at the silicon level. Skylake still has these flaws, with a mix of firmware fixes and software solutions that have cut processor performance. For the average gamer, a lot of these issues aren’t inherently concerning or performance affecting, but in datacenters, these issues have led Intel to crippling patches and has caused many dedicated datacenter customers to buy more CPUs from them to feed a loss of performance while in maintenance contracts or in an ROI period for existing datacenter purchases – filling in unused sockets in 2S motherboards, for example.
The problem has been that Intel’s Sunny Cove and Golden Cove core designs – their next-gen CPU cores – were designed for 10nm, a process that was delayed 2 years for basic laptops and still has yet to yield a meaningful number of chips, none of which are in the desktop space and none will be in the datacenter space for a while. For Intel’s 11th-gen Core CPUs, codenamed Rocket Lake, they’re supposedly finally backporting the Sunny Cove core design to 14nm, which is going to be a real issue. Will it port? Sure, and after some work, it will probably do relatively well. However, porting a new core design to the existing process means losing some of the optimization work that has been done. Skylake has clocked really highly and done very well because it was built for 14nm and hasn’t needed a ton of extra tweaking. A new core design not built from the ground-up for the process will lose some of those clock speed gains.
This alone is a concern if you’re Intel, but especially since AMD has maintained an aggressive once every 15-18 months release cadence for new Ryzen CPU generations. In the datacenter, Epyc based on the Zen 2 cores is cheaper, higher core count, and overall higher performance in most workloads. On the desktop, Zen 2 in the Ryzen 3xxx CPUs has closed most of the gaming performance gap while remaining the best overall productivity chips around, which makes them great for single-system Twitch streaming, YouTube editing, etc (it’s why I have a 12-core Ryzen 9 3900X, afterall). By the end of 2020, AMD will have Zen 3 based CPUs on shelves and available in the datacenter, keeping core counts flat but offering an improved 7nm process courtesy of their manufacturing partner TSMC which should allow clock speeds to push a little bit higher, and with rumored architectural changes to the way Zen works that will eliminate the gaming performance gap by reducing core-to-core latency by a substantial amount across each 8-core chiplet. Sunny Cove from Intel promises their first IPC increases in the last 5 years, with around 15% improvement or so over Skylake from most rumors, but Zen 3 is set to offer a rumored 10%+ over their already slightly higher IPC compared to Skylake, and some of the rumors have Zen 3 reaching 20% improvement in some workloads. Couple that with even a modest 100 Mhz clock boost over the current parts, and Intel might be in trouble in gaming.
Gaming is, after all, Intel’s refuge in the consumer market. Since Ryzen launched, Intel has had to aggressively market as the best gaming CPU, promoting the lead they do rightfully have in gaming workloads in a mix of popular titles. And, to be honest, if literally all you want to do is play games on a computer at the highest possible framerates to feed a 144 or 240Hz monitor, then sure – the current Intel Core i9-10900k is the part to get, security issues and all. However, AMD has been getting more aggressive in marketing – sometimes stupidly, like comparing desktop parts to laptop ones – but in a lot of cases, the value proposition for AMD is stronger. While a lot of PC gamers don’t stream or make videos or use their PCs for other things, a lot do. Even when I’m not trying to get back into video editing or streaming, I do a lot of 3D rendering and my investment in my Ryzen pays dividends there.
But Intel has a lot of trouble ahead, which they announced is getting worse today. For years, 10nm has been delayed to the point that Intel started promoting that they’d effectively “start over” with lineups on their 7nm process, and that those would be really impressive. However, today’s news is that Intel has had to delay their 7nm CPU lineup by 6-12 months, which, on current timelines, means they won’t have 7nm parts until 2023! So they have a slow-rolling 10nm process that won’t see meaningful datacenter rollout until late this year and consumer rollout outside of laptops until next year, and the 7nm process that was supposed to be the real exciting move to bring them back into prominence is now 3 years out? Yikes, Intel.
Of course, it is worth saying that right now, Intel is fine from a business perspective still. They have strong FPGA, AI, and accelerator businesses largely built on acquired companies, and their datacenter business has strong roots with locked-in contracts and sweetheart deals designed to keep people chained to Intel. Even in the consumer space, while AMD has made headway with system integrators, Intel has a large chunk of the low-end prebuilt PC market locked up in a similar fashion to datacenter. However, as delays mount for Intel, AMD stands a chance to win more market share and the longer it takes Intel to bust out of their funk, the worse it gets for them. AMD is set over the next year or so to have 5nm from TSMC in production for them, which will only make things worse if Intel is stuck on 10nm, much less 14nm!
Today, Epyc is increasing AMD’s marketshare in datacenter slowly, but it is the winning CPU for most workloads and comes in at a lower TCO compared to Xeon, unless you’re a precious enterprise customer for whom Intel will throw CPUs away at low cost. More software is being built and optimized around the Zen design, and it is becoming more common to see major system manufacturers building on AMD. AMD has also rolled out a Threadripper Pro CPU lineup for workstation use with 8-channel memory controllers, and a Ryzen Pro 4xxx lineup built on the strong Renoir laptop CPU design with up to 8 cores and integrated graphics, including low-power variants to use in rolling out low cost desktops for enterprise deployments like you might see in an average cubicle farm – a type of business Intel has had on lockdown for ages, even when AMD was riding high in the aughts.
Overall, it is fascinating to see just how much the market has changed because it was inconceivable that the mighty Intel would make this many tactical errors a few years ago. In 2017, it felt like a solid back and forth was about to happen, but while Intel has brought progressively less-impressive products to market to fight back, AMD has made large strides with each new Zen generation, and by the time Intel is ready to fight back, it might already be too late.