First, the rumors that AMD would move Zen 3 to 5nm have been disproven by a slide from AMD:
As expected, Zen 3 is a 7nm chip, while Zen 4 is 5nm. With a typical one-year cadence between launches and with Zen 3 expected to launch around Q4, we should, therefore, expect a late 2021 / early 2022 debut for Zen 4. The roadmap document implies a launch before 2022, but an early 2022 debut at or around CES would still satisfy the document well enough.
AMD is still doubling down on the idea that RDNA2 will be a 1.5x uplift over RDNA. AMD’s promises of GPU performance improvements haven’t always materialized as well as their CPU improvements have, but RDNA delivered the improvement it promised relative to the RX500 family, and a 1.5x improvement in performance-per-watt would help Navi 2 compete with expected GPUs from Nvidia’s Ampere family. A 2.25x improvement in performance per watt over GCN (1.5x * 1.5x) would be a substantial overall gain for a single node transition, even if it took several years to deliver the improvement.
We’ve seen some hints of what RDNA2 is capable of from the console demos and discussions that have already surfaced, like the Unreal-powered next-gen gaming engine demos from several weeks ago. The specs on the Xbox Series X and PlayStation 5 indicate that Microsoft and Sony were both aggressive on GPU frequency and capability, with Microsoft potentially pushing the envelope even farther than its Japanese rival. We expect RDNA2 to move both PC and console gaming forward — and with the entire PC and console industry supporting ray tracing after these launches, we’ll likely see more ray tracing-enabled games as well.
Like RDNA, RDNA2 is a 7nm architecture. RDNA3 is described as shipping by 2022 on an “Advanced node,” implying AMD may lead with GPUs at 5nm as it did on 7nm. AMD’s CDNA roadmap will follow the RDNA playbook at a delay, with CDNA2 moving to compute graphics only in 2021 at 7nm. No 5nm “Advanced node” transition is shown on the document. CDNA2 debuts with third-generation Infinity Fabric:
The major features AMD is talking about right now for IF3 include full cache coherency across multiple GPUs with up to eight cards affected, but we’d be surprised if Gen 3 didn’t introduce some new improvements to power efficiency and bandwidth allocation across the CPU as well. There’s no sign AMD will move away from its new chiplet-based methods of moving data around the CPU, but further tweaks to improve power consumption and latency are likely.
The document is light on any expected changes to Epyc, but Genoa (Zen 4) is expected by the end of 2021 as well, implying a quick turn-around time between desktop and server. As for the impact of these changes, they’ve paid major dividends in two of AMD’s three markets:
AMD’s claims on this slide don’t entirely track the data we’ve reported from other companies that also track x86 market share. The claim of 16.9 percent market share in client for 2019 aligns with Mercury Research reporting that found AMD with a 17.1 percent market share in Q1 and Q2 of that year, but the server claims don’t match at all, even with the “Excludes IoT” caveat:
With that said, there are reasons to believe AMD’s server division is performing well. Epyc’s high core counts and overall performance have won AMD fresh business from old partners and a lot of customer curiosity. During AMD’s Q1 conference call, it reported $348M for the Enterprise, Embedded, and Semi-custom business unit, while saying console revenues made up a negligible part of the segment this time around. This implies that the enterprise GPU and CPU business combined are currently worth about $350M — small, to be sure, but better than the approximately $0 in server CPU sales AMD was recording in the late Piledriver era.
It’s not clear where AMD’s market share data is coming from on this slide, but I do agree that the overall trajectory is positive.
The Big Picture
Scuttlebutt says that for AMD, 2020 is more about GPU refreshes and console launches than major leaps forward in CPU perf, though we’ll see what Ryzen refreshes and the Zen 3 launch offer later this year. We don’t expect to see core count jumps in 2020 — AMD has driven the number of cores in a mainstream PC from 4C/8T to 16C/32T in just three years, and software developers need time to adapt to those changes.
The Ryzen Mobile launch earlier this spring put the company on the right foot for the rest of the year, and Zen 3 may launch against Intel’s Rocket Lake late this fall, but AMD’s 2020 looks to be more focused on console launches and GPU share as opposed to aggressively pushing CPU core counts or clock speeds. Just because the company isn’t making noise about new products doesn’t mean it isn’t quietly winning market share.
One of the advantages of competing against companies like Intel and Nvidia, which dominate their respective server markets for CPU and GPU, is that any amount of market share AMD can seize represents a significant gain. Moving from 5 to 10 percent market share is headline news for AMD while capturing 80 percent versus 85 percent of the CPU/GPU server market isn’t as big of a social win for Intel or Nvidia.