AMD’s Zen 2 Architecture Is Coming. What Kind of CPU Improvements Should You Expect?

The upcoming launch of AMD’s third-generation Ryzen microprocessors based on the Zen 2 architecture is a major moment for the company. Not only will this be the first time AMD has ever beaten Intel to a new node, but it’s also a necessary and important demonstration of AMD’s ability to execute an iterative roadmap. This kind of long-term improvement cycle is critical to winning business and server space. Given the importance of the launch to both consumers and corporations, what sort of parts should we expect to see from AMD?

Note: According to AMD, the formal name for the CPU family is third-generation Ryzen with Zen 2 microarchitecture. The Ryzen 3/Zen 2 split we adopt in this article reflects this differentiation.

I decided to revisit this question after data from the European retailer Mindfactory.de caught my eye. We covered the company’s sales figures in a separate story; this is the relevant slide in question:

Since last December, there have been rumors that AMD would use the launch of Ryzen 3/Zen 2 to push core counts higher across its entire product family while slashing prices. The chart below is representative of this rumor, which predates CES:

The product positioning and pricing that this chart (and others similar to it) predicts is incorrect. The net effect of these changes would slash AMD’s current per-core pricing by over 50 percent. While we expect AMD to aggressively introduce new, faster CPUs, the company isn’t going to do it this way. Let’s talk about why.

Margins and Markets

First, AMD doesn’t have financial room to pull this kind of trick, even if it wanted to.

As of Q4 2018, AMD’s gross margin was 37.84 percent. This represents a significant long-term improvement over the pre-Ryzen era. The company’s actual operating and net profit margins, however, are much lower — in the 5-6 percent range. This useful tool from MacroTrends.net lets you compare all three values; a screenshot of AMD’s most recent results and historical data is shown below:

Post-2011, AMD has needed to maintain a gross margin of 33-37 percent in order to remain profitable. The company’s net profits are only 5.2 percent of revenue. Intel’s net profit margin, for comparison, is 29.72 percent. AMD literally can’t afford to gut its own pricing in this fashion without destroying its own profits.

In the past, moving to new nodes could be counted on to reduce long-term silicon pricing and chip costs. This is less true now. Design costs and wafer costs have both risen sharply with every product generation. In his 7nm launch discussion, Mark Papermaster noted that “we’re not getting the frequency lift we used to get,” and called 7nm “a rough lift that added masks, more resistance, and parasitics.”

AMD has made several smart moves that should allow it to maximize profits even at relatively low yields, including pairing 7nm chiplets with a 14nm central I/O die. But these moves should be understood as responses to a tough cost curve and rising engineering difficulty, not a magical unicorn that allows the company to ignore them. Wafer costs and design costs have risen sharply with each new node.

Investors have made it clear on conference calls that they want AMD’s margins to come up. AMD itself has identified critical markets that it wants to compete in, including servers, laptops, and the burgeoning AI/deep learning market, which is expected to define and drive computing in the next decade. While its server market share is rising, AMD’s share of the entire AI/DL space seems to be basically nil. The company has done some work with Google and there have been rumors that it built Radeon Instinct as a semi-custom solution for a customer, but, as this extensive machine learning overview from AI expert Tim Dettmers demonstrates, AMD isn’t viewed as a strong competitor by professionals working in this field, at least not right now. He writes: “AMD invests little into their deep learning software and as such one cannot expect that the software gap between Nvidia and AMD will close.”

If AMD wants to expand into these spaces — and it certainly claims that it does — it’s going to invest heavily in doing so to match the level of software and hardware development Nvidia, Intel, Google, and Amazon are pouring into it. Again, that takes revenue. AMD is already collecting a whole lot less of it per CPU core than Intel is, and it shares TSMC’s 7nm foundry capabilities with multiple high-profile companies. The company cannot depend on slashing costs and making it up on volume the way it might have when it owned its own manufacturing facilities and set its own production targets. Capacity allotments are negotiated in advance.

Historical Trends

Second, AMD has never historically gutted its own product families pricing in the manner these various 7nm rumors anticipate. Prior to the introduction of the Ryzen 7 family, which debuted as high as $500, AMD seized leadership of various markets on two occasions: The introduction of the Athlon 64, and the launch of the Athlon 64 X2. The first re-established AMD’s 64-bit chip as a superior solution in many consumer applications after the Athlon XP had lost that crown. The second gave AMD a serious claim to overall superiority, even in rendering and multi-threaded apps where it had previously lagged the P4 thanks to the latter’s support for Hyper-Threading.

These were not cheap CPUs. In both cases, AMD took advantage of its performance improvements relative to Intel to retrench its pricing and improve its ASPs. In fact, AMD’s dual-cores were significantly more expensive than Intel’s Pentium D equivalents at launch for precisely this reason.

If you care about AMD’s long-term success, a return to Bulldozer pricing models is the last thing you want. Companies with low net margins have difficulty absorbing market downturns or changing conditions without slipping into the red. The enormous gap between AMD and Intel’s net margins is partly evidence of how much AMD still needs to improve its own position to establish itself on truly solid footing.

What to Expect When You’re Expecting Ryzen 3

There are a few things we already know about Ryzen 3/Zen 2 and AMD’s 7nm process expectations. We know that AMD has said it can deliver 25 percent more performance on 7nm within the same power envelope. We know, based on AMD’s on-stage demo, that it can deliver a Ryzen CPU with equivalent performance to a Core i9-9900K in Cinebench 15 within a much lower power envelope. (AMD claimed a CB15 score of 2023 with 130W of system power consumption compared with a score of 2042 for the 9900K with 180W of system power consumption.)

We know that there’s a spot on the CPU package where a second chiplet could be installed and that AMD will likely use this to boost its mainstream desktop platform up to 16 cores. Threadripper will likely evolve upwards as well. The rumored IPC improvement for Ryzen 3 over Ryzen 2 has consistently been in the 10-20 percent range.

If AMD follows its historical pattern, it will hit Intel with a targeted mixture of clock speed gains, IPC improvements, and improved power efficiency while seeking to further improve its margins, not slash them. This does not automatically preclude further core count gains or improvements in per-core pricing. It just means the company isn’t going to gut its own margins and revenue to deliver those benefits.

Having spent so much time critiquing other people’s rumors, it seems only fair that I stick my own neck out and make a few predictions of my own.

First, I expect one of AMD’s major stories around Ryzen 3/Zen 2 to be a low-power play. The company didn’t emphasize Ryzen Mobile as much as I thought it would during the first year of that part’s existence. I think the Zen 2 architecture will be better suited to the segment. TSMC’s 7nm started as a low-power process and the performance gains for x86 chips over the past decade, statistically, have been largest in the lower power bands. I don’t know how much of an improvement we’ll see in the desktop parts, but I expect it to be visible in the form of lower TDPs on at least some chips.

Third, if AMD launches new chips with >8 cores, I expect them to slot in above the current 8-core desktop price point ($300 for a Ryzen 7 2700X). It is possible that AMD will move 8-core chips to modestly lower price points, but I do not expect the company to sell an 8-core processor for under $200.There are rumors of a 10-core high-end Intel desktop CPU coming in 2019, but I expect Intel to keep its CPU pricing as high as possible to protect its own margins, limiting AMD’s need to slash its own.

It seems unlikely that we’ll see a top-to-bottom retrench in the desktop market similar to the one that followed Ryzen 7’s initial debut. Instead, we’ll see top-end core counts expand on both AMD and Intel platforms, with AMD likely stepping up to 12-16-core chips on mainstream Ryzen. It’s possible that mobile CPUs might step up to higher core counts this year as well, but there’s no chance that AMD simply drops quad-cores altogether and standardizes on six-core chips as its minimal unit of computing.

Finally, I expect AMD will push performance as far as it can in terms of clock and IPC without blowing its power curves to hell or leaping for core counts that would wreck its pricing model. This is actually a bit of an evolution for me. In the past, I’ve argued that the company would opt to introduce more measured improvements every year as opposed to leaping for every bit of gain it could scrape out of silicon. (To visualize the difference, imagine a static 10 percent gain each year as opposed to an alternating 5 percent/15 percent improvement rate.) But the ASP gap in the Mindfactory.de data and some other rumors I’ve heard make me think I was wrong on this point. AMD is likely to play 7nm as aggressively as it believes it can — just not to the detriment of every other aspect of its product strategy. To that end, I expect the company to push core counts only in the upper range of its market, while competing more aggressively on clock and IPC at every price point, with lower-power parts aimed at the mobile market and desktop users who want to build small form factor systems.

Continue reading

Intel’s Raja Koduri to Present at Samsung Foundry’s Upcoming Conference

Intel's Raja Koduri will speak at a Samsung foundry event this week — and that's not something that would happen if Intel didn't have something to say.

A 15-Minute Smartphone COVID Test Is Coming to the US

The Department of Defence and Department of Health and Human Services have awarded $230 million to an Australian biotech firm Ellume to produce a 15-minute COVID test that sends results directly to your phone.

Framework Unveils Modular Laptop Coming This Summer 2021

Laptop startup Framework believes it's finally cracked the code on customizable mobile I/O.

Rocket Lab’s Upcoming Reusable Rocket Is Designed for Deploying Mega-Constellations

The company says its upcoming Neutron rocket will be ideal for deploying mega-constellations, and it'll have a reusable first-stage a la the Falcon 9.