The processor technology war is heating up in a way it hasn’t since the late nineties through early 2000’s, and that is really exciting.
My last post on this topic, back in March, focused on rumors and some claims of performance about the new Ryzen 3000 series CPUs from AMD. Things have largely remained the same since then, with the CES 2019 demo and the New Horizon event last November being the only peeks at official performance figures, with New Horizon showing the server performance of the Zen 2-based Epyc CPU with 64 cores, and the CES 2019 demo showing performance of an 8-core Ryzen 3000 part. Both have a lot of unknowns – base frequency, boost frequency, actual model numbers, or anything that would allow you to draw a direct comparison across the competitive landscape, but yet, they served their purpose – both events got a lot of people very excited for the upcoming launch of these processors.
Rumors have since leaked out detailing possibly more information, as the products get closer to launching – the desktop Ryzen series seems ever more likely to be available from 6 cores up to 16, with the 6 core model being the base level (huge if true). Rumors have added details about higher levels of DDR4 speed support, with the possibility of supporting as high as 5,000 MHz memory (this would be highly overclocked, but the motherboards that seem to support this rumor have memory dividers that make it play a bit nicer). Several X570 motherboard designs have leaked, and while these are not necessary for most existing Ryzen users, they help to confirm little details about the new processors. Of course, in just under two weeks, the official Computex keynote from AMD is expected to answer these questions and finally set an available date for these new CPUs.
What is fascinating here, and is what I want to touch on today, is the baffling lack of response from Intel. Intel, in case you’ve never touched a computer, is the (mostly) undisputed king of computing, whose processors have powered a majority of the world’s desktops, laptops, and servers for decades. Intel has, until lately, been on a decade-long winning streak of having the best processors on the market in terms of overall performance. Sometimes they’ve been beat by AMD in certain segments of the market (usually budget) or in certain terms (in price-to-performance, AMD often won), but in terms of the overall top performance, Intel has held that crown for a long time.
This makes it particularly confusing that, when, in 2016, the full extent of the Zen design from AMD became known, Intel’s response was to do…nothing, pretty much. Their plan of rehashing the Skylake architecture continued on, and still continues on, as the most likely release for the desktop from Intel this year is yet another refresh of Skylake, rumored to have 10 cores. Intel has had some troubles that are worth recapping, however, before we analyze this.
The Loss of Leadership – Intel Loses Foundry Leadership
Intel has nearly always been the leader in silicon manufacturing, being one of the only remaining companies that owns their own foundries to manufacture their processors. They’ve pioneered several of the major advances in chip engineering, including new materials, insulator technologies, and are also ramping up 3D stacked packaging with their FOVEROS technology. They also, for most of their history, maintained a process node advantage, being able to build smaller, more space-efficient transistors than their competition. When TSMC (the leading 3rd party foundry) was at 16 nanometer transistor sizes, Intel was at 14 nanometer. Intel’s big jump after 14nm was scheduled to be to 10nm, a shrink that would yield some decent performance gains. 14nm was introduced by Intel in 2015 with a shrink to their existing processor lineup, and then was expanded in 2016 with the Skylake flagship lineup. The plan was that 14nm would be relatively short lived, as Intel had big plans in mind for 10nm.
However, 10nm hit a lot of walls. For a variety of different reasons, the technology just wasn’t working as intended, and Intel has been stuck since 2015 releasing parts on 14nm exclusively while trying desperately to tool up their 10nm lines and get yield high enough to release. Intel released two 10nm parts in late 2017, which yield so poorly that the most commonly available was a dual-core that has to disable the integrated graphics part – which is on the chip, but is so often defective that it is laser-cut in order to salvage a working dual-core CPU from the wreckage. Rumors have it that these parts on 10nm have a single-digit yield percentage, which is disastrous for silicon manufacturing (for comparison, most new process nodes have around a 70% yield at a point they begin full-time production, and tend to mature towards 90+% yield at full development).
In the meantime, Intel’s competition has not slept, and has in-fact lapped Intel for the first time in history, with both Samsung and TSMC both having mature 10nm process nodes and having moved volume production to a 7nm process (to be fair, because nothing can be simple anymore, while the nanometer size in a process node once stood for the actual physical width of the transistors manufactured, now it is not the case, so Intel’s 10nm is closer in size and power savings to Samsung and TSMC’s 7nm process, and those foundries’ 10nm processes are larger than Intel’s own 10nm). The 7nm node from TSMC is what the new lineup from AMD will be using, with the Radeon VII graphics card released earlier this year being built on the process and the full Zen 2 lineup also scheduled to be built on it.
This means that AMD has access to better process nodes than Intel, and will be shipping a smaller-node product for the first time in history. Intel, in the meantime, is expected to release their 4th (or fifth, depending on how you define it) refresh of the Skylake architecture on 14nm yet again, with the defining addition being a likely 10-core flagship intended to help compete against what is expected to be a lineup from AMD with 12-core Ryzen 3000 CPUs at the enthusiast mid-range and a possible 16-core Ryzen as the flagship part. These new CPUs from Intel will have no new features other than the added cores, and the intention seems to be that the first 10nm consumer mass-market parts will be laptop processors in late 2019, followed by desktop CPUs in 2020. What’s worse is that when Intel finally does release those 10nm desktop CPUs, most indications are that these processors will actually be lower-clocked due to process limitations, and while in theory Intel could refine the process over a few years, as they have with 14nm (which released with low-to-mid 4 GHz turbo clockspeed CPUs and now releases a decent lineup with 5GHz turbo speeds), the rumor mill suggests that Intel has already removed much of its 10nm capacity and will instead focus on their 7nm process after the launch of the 10nm lineup.
Why have they removed the 10nm capacity? Well, after figuring they had time to tool up those production lines, the competitive response from AMD (particularly with Epyc in the datacenter and the expected migration to the Zen 2-based Epyc parts over the next 6 months) forced Intel to refocus on their current products, where they still have a performance lead, which resulted in 3 out of 4 fabrication plants being switched back to 14nm from 10nm to produce the currently-leading Xeon lineup and the desktop-leading Core i9 9900k, which have been hit with supply shortages. These shortages further detracted from Intel’s plans, as while they could have sold a boatload of these CPUs (and they still did, make no mistake), they had to limit inventory into the enthusiast DIY market right as rumors began to circulate about what to expect from AMD’s Zen 2 and the CES 2019 demo that showed the 8-core Ryzen 3000 part beating the Core i9 9900k. The CPUs that were available were marked up to nearly double the price of the top end Ryzen 7 2700x, which is only around 15% slower in only certain use cases, and would offer the possibility of a drop-in upgrade to Ryzen 3000 once that launched. So you could have either spent nearly $600 for the Core i9, and had leading performance for over half a year, or you could instead have spent $300 on the Ryzen 7 2700x, been a bit behind, and then leapfrogged the i9 9900k half a year later and end up spending around the same cost for it in the end, but now with a CPU you can pass down or resell.
Sure, it is a string of dumb luck, but AMD has aggressively played every inch they can get from Intel. Intel messes up supply? Well now, buying a Ryzen gets you a bundle of free games. Intel wants to change socket type every two years, forcing a motherboard upgrade and a new copy of Windows? AMD will end up supporting socket AM4 for around 4 years, so a good motherboard from the Ryzen 1000 launch will work for around 3 generations of upgrades. Intel’s stock coolers are loud and lousy at their job? Well, AMD bundles better stock coolers with all of their CPUs in their mainstream lineup, which itself can be a money saver if you aren’t looking to overclock much past a few hundred MHz. Now, right as Intel predicts they will be able to end their supply shortage of 14nm CPUs, AMD will be launching their new lineup, whose performance is predicted to beat Intel. While earlier, I did say it is dumb luck for AMD in some of these cases, the fact that Intel has made so many unforced errors in this time of increased competition boils down to Intel being largely unprepared for an AMD that is competitive.
Because time is a flat circle, this isn’t even the first time this has happened between these two, either. In the late nineties and early 2000s, Intel made a choice to deepen their execution pipeline on the Pentium 4, pushing for higher raw clockspeeds and theoretical better performance per clock. The problem is that the deeper pipeline required that the CPU be constantly fed with data, and that it was able to accurate get ahead of the current operations to predict the next step through smart branch prediction. Intel’s branch prediction in the Pentium 4 lineup was awful, though, and so the large pipeline went mostly unused and the performance of the processors suffered as a result.
AMD, meanwhile, used a smaller execution pipeline in the Athlon XP lineup, and the end result of that was that in most tasks, the AMD CPUs were better, and in some cases, much better. Intel refined the Pentium 4 design slightly over the years, but was stuck with variants of the flawed design of that CPU from 2001 until 2006, when the Core lineup began on desktop. The Pentium 4 design was so awful that when the Core 2 parts launched, they used a completely new architecture derived from the power-sipping laptop designs Intel was pushing at the time, rather than trying to refine and sharpen the Pentium 4’s “NetBurst” architecture.
For nearly 5 years, AMD had the best-performing parts, and it is only due to the variety of lucky breaks – the power of Intel’s Pentium marketing in the 90’s, the prevalence of their Xeon CPUs in the less-massive datacenters of that era – but also, the iffy business practices they used at that time, like providing CPUs at vastly reduced costs to OEMs like Dell and HP – that Intel managed to maintain a marketshare lead. During this time, they had other failures too – Intel’s attempt at moving to 64 bit was the short-lived Itanium and the related IA-64 instructions, which lost to AMD-64 to the extent that Intel licenses AMD-64 and it is the prevailing x86-64 instruction set to this day. That is without even getting into the misread of the market that led to Intel making the early Pentium 4 chipsets only compatible with RDRAM, a short-lived technology that ended up in most homes thanks to its use in the Nintendo 64, but was too expensive and esoteric for home desktop usage (RDRAM was a dual-channel memory, but with 4 RIMMs (yes that was the acronym) capable of being slotted into most Pentium 4 motherboards, the slots not populated with memory required terminating RIMMs be inserted for signal continuity. This blunder led to Intel having to quickly develop chipsets for other memory types, which led to an additional blunder, as the first chipset out of the gate was an SDRAM one, which was anemically slow, before they finally got a DDR version up and running and into consumer’s hands.
This history seemingly mirrors what we expect out of the CPU space over the next several years. Intel will be stuck with their legacy design as AMD overtakes them on performance, while their next generation lineup will need to deliver impressive IPC gains to beat what we expect from Zen 2, only to have Zen 3 scheduled for next year with possible further improvements to performance (the current rumors on Zen 3 are that it will have SMT4 technology, meaning each core can run 4 threads at once, which is expected to be pared down for non-datacenter CPUs but will likely still result in SMT3 on desktop, meaning a 12-core on this architecture will have 36 threads to run!). It is also likely that AMD will see a similar clockspeed regression to the Intel 10nm parts in this move, but if SMT4 and other possible improvements like cache on the IO die see the light of day, it is possible that performance will remain intact or see some increases under the right types of workload. The tweaks made to Zen 2 to optimize for gaming performance will carry over in all likelihood, which will result in Ryzen being much better at gaming than it is today (which, to be fair, the modern Ryzen parts aren’t bad, but they are certainly lacking the single-threaded performance most games still lean on).
Intel is, to their credit, beginning to diversify their lineup more than they have over their history. In the next several years, their XE discrete graphics are expected to launch, and based on what was shown of the Generation 11 integrated GPUs at recent press events, these parts will be worth keeping an eye on. Their SSD lineup and X-Point flash memory technology has been great and an Intel SSD is one of the best performance buys you can make in the storage market.
However, it is absolutely worth noting that Intel’s core competencies have always been CPU design and their intimidating lead in semiconductor manufacturing, and to see these fall as they have has been fascinating. Ultimately, I expect that Intel, while it will face rough road over the next 5 years, will manage to recover and will likely hold a market lead as well. Too many consumers are bought in to the Intel brand, and while AMD does look to be making massive advances in the datacenter market, Intel benefits from the upgrade lifecycle in that space being 3-5 years to see a return on investment in the hardware. For customers with upgrade cycles coming up in the near-term (1-2 years), Intel is liable to lose those customers to AMD Epyc, particularly the Zen 2-based Rome architecture parts and the likely 2020 rollout of “Milan” based on Zen 3. However, some of those customers may remain Intel loyal, and for those that waver, Intel is likely to make discount plays to keep these customers through the current storm while they regroup from their own lack of preparation.
To their credit, Intel has also published their roadmap for the next several products and, assuming these remain accurate, Intel will be making some effort to regain any lost ground. Their first 10nm core design, Sunny Cove, has IPC gains which should help offset any clock regression, and seemingly every other design is targeting some attempts at IPC gains, which should keep them gaining performance at a steady clip over the early 2020’s. The unknown in this phase of things is AMD – while their roadmap goes through Zen 4 in 2021 and confirms Zen 3 is already designed and undergoing validation, what changes these new architectures will bring is unknown, where Intel’s attempts to add some detail through brief descriptions of the design improvements each future architecture will bring.
However, it remains a curious case why, when in 2015-2016, the rumors about what Zen would bring growing louder, Intel decided on the course of action that would leave them least able to counter-attack once AMD came to market. Lucky on AMD’s part? Perhaps, but there are a lot of lucky breaks they’ve received that have led to this moment.