Today is AMD’s Financial Analyst Day, the company’s biannual gathering for analysts. While the primary purpose of the event is for AMD to engage investors, analysts and others to demonstrate the company’s performance and why they should continue investing in the company, the FAD has also become AMD’s de facto product roadmap event . After all, how can you invest wisely in AMD if you don’t know what’s next?
As a result, the half-day presentation series is packed with small bits of information about products and plans across the company. Everything here is of a high standard – don’t expect AMD to release the Zen 4 transistor floorplan – but it’s easily our best look at AMD’s product plans for the next few years.
Kicking off FAD 2022, AMD’s always most interesting update is the Zen architecture roadmap. As the cornerstone of AMD’s recovery and resurgence as a competitive and capable player in the x86 processor space, the Zen architecture powers everything from AMD’s smallest embedded CPUs to the largest enterprise chips. So what’s coming off the line over the next few years is a very big deal for AMD and the industry as a whole.
Zen 4: Improved performance and performance per watt, shipping later this year
AMD is currently in the process of ramping up its Zen 4 architecture based products. These include the Ryzen 7000 (Raphael) client CPUs and their 4thth Generation EPYC (Genoa) server CPUs. Both are expected to be on the market later this year.
We’ve seen information about Zen 4 so far, most recently with the announcement of Ryzen 7000 at Computex. Zen 4 brings new CPU core chiplets and a new I/O die, and adds support for features like PCI-Express 5.0 and DDR5 memory. And in terms of performance, AMD is aiming for significant improvements in performance per watt and clock speeds over their current Zen 3-based products.
In the meantime, AMD follows up on this Computex announcement by clarifying a few things. Specifically, the company addresses questions about Instruction per Clock (IPC) expectations, stating that it expects Zen 4 to offer an 8-10% IPC increase over Zen 3. The initial Computex announcement and demo seemed to indicate that most of AMD’s performance gains came from clock speed improvements, so AMD is working to respond without showing too much of their hand months after product launch.
In connection with this, AMD also announces that they expect an overall single-threaded performance gain of more than 15% – with an emphasis on “greater than”. ST performance is a mix of IPC and clock speeds, so AMD can’t get overly specific at this point as they haven’t finalized the final clock speeds. But as we’ve seen with their Computex demos, the Zen 4 currently has 5.5GHz (or higher) on the table for lightly threaded workloads.
Finally, AMD also confirms that there will be ISA extensions for AI and AVX-512 for Zen 4. At this point, the company isn’t clear on whether either (or both) of these additions will be included in all Zen 4 products, or just a subset – AVX-512, for example, is a bit of a space and power hog – but at least it is it’s reasonable to expect these to show up in Zen 4 server parts. The addition of AI instructions will help AMD keep up with Intel and other rivals in the short term as CPU AI performance has already become a battleground for chipmakers. What this means for AMD’s competitiveness, however, depends in large part on what instructions (and data types) are added.
AMD will produce three flavors of Zen 4 products. These include the vanilla Zen 4 core as well as the previously announced Zen 4c core – a compact core for high-density servers that will be integrated into the 128-core EPYC Bergamo processor. AMD is also confirming for the first time that there will also be V-cache equipped Zen 4 parts – which is new information but unsurprising given the success of AMD’s V-cache consumer and server parts.
Interestingly, AMD plans to use both 5nm and 4nm processes for the Zen 4 family. We already know that Ryzen 7000 and Genoa are planned to use one of TSMC’s 5nm processes, and that Zen 4c chiplets are planned to be built on the HPC version of N5. So it’s not immediately clear where 4nm fits into AMD’s roadmap, although we can’t rule out AMD playing a little fast and loose with the terminology here, given that TSMC’s 4nm processes are an offshoot of 5nm (rather than an entirely new one nodes) and are typically initially classified as 5 nm variants.
As of this writing, AMD expects the Zen 4 to deliver >25% performance per watt over the Zen 3 (based on desktop 16C chips running CineBench). Meanwhile, the overall performance improvement is >35%, undoubtedly taking advantage of both the architecture’s higher per-thread performance and the higher TDPs previously announced by AMD (which are especially handy for uncorking more performance on MT workloads). And yes, those are terrible graphics.
Zen 5 architecture: all-new microarchitecture for 2024
Meanwhile, AMD’s Zen architecture roadmap to 2024 carries the Zen 5 architecture, which will be announced today. Since AMD hasn’t yet delivered the Zen 4, the specifications for the Zen 5 are understandably very high. Nonetheless, they also point out that AMD won’t be resting on its laurels and has some aggressive updates planned.
The big news here is that AMD calls the Zen 5 architecture an “all new microarchitecture”. That means it won’t just be an incremental improvement over Zen 4.
In practice, no major vendor designs a CPU architecture from scratch – there will always be something good for reuse – but AMD’s message is clear: they’ll be making some significant revisions to their core CPU architecture in turn to further improve their performance and energy efficiency.
What AMD will reveal now, Zen 5 will reroute the front end and increase the output width again. The devil is in the details here, but given Zen 3 and its 4-instruction/cycle decode rate, it’s easy to see why AMD might want to focus on that next – especially when in the backend the integer units already have a 10- . large booklet width.
Meanwhile, Zen 5 integrates more AI and machine learning optimizations in addition to Zen 4’s new AI instructions. AMD doesn’t say much more here, but they do have an extensive library of tools to choose from, covering everything from AI-focused instructions to adding support for even more data types.
AMD anticipates that the Zen 5 chip stack will be similar to Zen 4 – meaning they will have the same trio of designs: a vanilla Zen 5 core, a compact core (Zen 5c) and a V-cache enabled core. For AMD’s customers, this type of continuity is very important as it gives customers a guarantee that AMD’s more bespoke configurations (Zen 4c & V-Cache) will have successors in the 2024+ time frame. From a technical perspective, this isn’t too surprising, but from a business perspective, customers want to make sure they’re not adopting dead-end hardware.
Finally, AMD has planned an interesting manufacturing mix for Zen 5. Zen 5 CPU cores are manufactured in a mix of 4nm and 3nm processes, which unlike the 5nm/4nm mix for Zen 4, which uses 4nm and 3nm Nodes from TSMC are very different. 4nm is an optimized version of 5nm while 3nm is a whole new node. So if AMD’s manufacturing plans move forward as currently planned, Zen 5 will span a major node jump. However, it’s not unreasonable to assume that AMD is hedging its bets here, leaving 4nm on the table in case 3nm isn’t as far as they’d like.
Finally, it should be noted that the Zen 5 architecture is planned for 2024. AMD doesn’t provide any further information on when that might be in the year, although Zen 3 and Zen 4 were both released later in 2020 and 2022, respectively. So H2/EOY 2024 is pretty much any guess.