Last week, the first purported benchmarks of AMD’s upcoming “Zen” microprocessor leaked, giving us an early glimpse of how the chip performs vis-à-vis various Intel products. While the leaks haven’t been confirmed as genuine, it was interesting to see the conversation around Zen and AMD’s expected performance level. Then, at an event on Wednesday, AMD showed some demos of Zen running clock-for-clock against Intel’s 6900K. While these demos were not performed at full speed, they offered some evidence that AMD could match Intel clock-for-clock in certain workloads — and while such demos are always cherry-picked to favor one’s own company, they gave some additional context on what Zen can do.
What I want to talk about here has less to do with AMD’s demo or specifics of the architecture and more to do with what AMD needs to deliver to be considered a success. One thing I’ll say up front: I have no secret sources, no hidden informants, and no leaks to credit for what I’m about to write. Agree or disagree, these are my own thoughts and observations.
AMD claims that Zen will have 40% better IPC than its previous Excavator architecture and it’s shown demos illustrating how Zen can compete clock-for-clock against the 6900K in some workloads. This is both a solid indication of where the chip currently is and a generalized prediction of where it could end up. The details and specifics will depend on a number of factors including final clock speeds and application-level optimization.
What does 40% more IPC mean, anyway?
IPC stands for instructions per clock or instructions per cycle (the two terms are synonymous). It’s a metric of CPU efficiency — the higher a CPU’s IPC, the more work that CPU can perform in a given amount of time.
There’s typically an inverse relationship between a CPU’s clock speed and its IPC. While IPC has always been used to measure CPU performance, the term became common in enthusiast circles when Intel debuted the original Pentium 4. While the P4 was clocked much higher than its Pentium 3 predecessor, the amount of work it performed per clock cycle was significantly lower. The net result of this was that the Pentium 4 1.5GHz often struggled to outperform the slower Pentium 3 1GHz or its Athlon K7 counterpart from AMD.
The most common way to parse AMD’s statement is as follows: If you took an Excavator CPU and a Zen CPU and ran them at the same clock speed, the Zen CPU should be 40% faster, on average, then the AMD CPU it replaces. But even here, there have been questions: Does that 40% IPC improvement include the impact of simultaneous multi-threading (what Intel calls Hyper-Threading) or not? Is that based on two chips clocked at the same frequency or not? Finally and most importantly, which tests were used to derive that figure? IPC isn’t a constant — it fluctuates a great deal based on the behavior of the target application.
The benchmark above is from tests I ran when Bulldozer launched nearly five years ago. Despite losing to AMD’s older six-core X6 1100T in multiple benchmarks, the FX-8150 was actually significantly faster in Maxwell Render — one of the few tests it unilaterally won. This kind of variance is normal, which is why it’s so difficult to draw conclusions about relative chip performance based on any single metric. AMD’s claimed 40% IPC uplift should be treated like a general or average prediction and not the guaranteed result of any single test.
How will Zen compare with Intel CPUs?
The real question, of course, isn’t whether Zen will improve on Carrizo, but whether it’ll give AMD a CPU that can compete with Intel. The realistic answer is “It’ll depend on where you look and what you’re looking for.”
Here’s what I mean by that. Compare AMD’s current top-end Piledriver, the FX-9590, against Intel’s Core i7-6700K using Anandtech’s Bench tool. I’ve snipped a section of tests to include below, but you can view the full comparison here. In the graph below, blue is for the FX-9590, while the orange-ish bar is Intel’s Core i7-6700K.
Assume for a moment that these FX-9590 results actually reflected Excavator performance instead of Piledriver. Now, assume they were 40% faster than they actually are. Would this theoretical top-end chip offer competitive performance with Intel? In some tests, it absolutely would. In other tests, the gap between AMD and Intel is large enough that even a 40% improvement isn’t enough to allow AMD to catch its competitor. Those of you who want to perform this comparison using actual data from Excavator can refer to this AT review of the chip and do the math from there. Either way, the point holds — there are tests where a 40% improvement would absolutely be enough to allow AMD to catch Intel and tests where it wouldn’t be. When we say “It’s going to depend,” it’s not a dodge or an excuse, it’s a fact. AMD’s IDF showcase gave some additional information but not enough to change this basic situation.
How do you eat an elephant?
One of the common arguments raised in the Ashes of the Singularity benchmark thread was that Zen’s comparatively weak performance compared with the Core i7-6900K is proof that AMD’s eight-core chip won’t be able to compete against Intel. The simple fact that these test results were run on an ES chip of unknown stepping and vintage make that an extremely premature conclusion. Worse, it’s a conclusion founded in an incorrect premise — specifically, that AMD has to match Intel’s top-end performance in order to compete at all. While AMD’s recent demos should assuage some of these fears, the early unveil wasn’t meant to be a complete overview of every aspect of the chip’s performance.
Right now, AMD’s FX-8350 is a $160 CPU, while the top-end FX-9590 retails for $229. Intel’s quad-core Core i5 desktop processors start at $185 for older Haswell parts and $190 for Skylake. The cheapest Core i7 you can buy starts at $295 for Haswell and $305 for Skylake (all prices from NewEgg). This means the FX-8350 is priced against a high-end Skylake Core i3 (the Core i3-6320) while the FX-9590 faces off against the Core i5-6600 — a comparatively tiny chip with an integrated GPU and one-third the TDP
In the 15 years I’ve been reviewing CPUs, the gap between AMD and Intel has never been larger than it is today. That’s part of why Zen is so important, but it’s also why it’s important to keep perspective on what kind of performance improvement AMD can reasonably deliver in a single product cycle.
Based on AMD’s single 40% IPC figure, it’s extremely unlikely that the company will deliver a CPU that can match Intel’s performance at every particular and at every price point — and it doesn’t need to. A 40% IPC boost would allow AMD to challenge Intel’s Core i3/i5/i7 line-up by leveraging larger core counts and its own SMT implementation far more effectively than Piledriver ever did. The drastically reduced TDP (Zen has reportedly targeted 95W at the high end) will allow it to compete much more aggressively on power consumption. Higher CPU efficiency means that AMD won’t need to rely on high clocks to hit performance targets, making it easier to push into laptops when Zen-based APUs come to market. Trade-offs between core count and clock speed also mean AMD can probably offer lower core counts at higher clocks, the same way that Intel does.
It’s easy to forget, but AMD didn’t launch K7 in 1999 and seize 20% of the server market 12 months later. K7 was competitive with the Pentium 3, but not necessarily faster — particularly since Intel’s Coppermine P3’s ran a full-speed L2 cache while the slot-based K7 and K7.5 used a half-speed or 1/3 speed L2. Socket A and Thunderbird closed this gap in June 2000, but AMD didn’t enter the server market until 2001. It didn’t see serious success in the server market until 2003, when K8 gave it the legs it needed to go head-to-head with Intel’s high-end Xeons.
Right now, Intel’s least expensive eight-core processor is the Core i7-5960X, at $1,015. AMD doesn’t need to hit Intel in the $1,000 CPU market to drastically improve its own fortunes or competitive standing.
Zen will be judged a failure or a success based on how well it performs today and on how well it positions AMD to deliver future improvements. It may not going to be strong enough to go toe-to-toe with Skylake across every SKU and price point, but that’s not what it needs to do. AMD needs Zen to perform well at low power so it can be paired with an APU and slipped into notebooks. It needs Zen to be more efficient than Excavator so Sunnyvale isn’t left pricing eight-core chips against Intel’s Core i3. And it needs Zen to be strong enough that investors and enthusiasts see the core as a future to be built upon, rather than an anchor around the company’s neck.
I genuinely don’t know how well AMD will deliver on these goals. But those are the criteria I expect the chip to need to fulfill and the metrics by which I’ll judge its overall position. I’ll have more to say about Zen, including a deep dive into its architecture, when I return next week.
August 19, 2016