Today is AMD’s Financial Analyst Day, a semi-annual analyst-centric gathering for the company. While the main purpose of this event is to contact AMD with investors, analysts and others to show the company’s performance and why it continues to invest, FAD has also become AMD’s de facto product roadmap event. I am. After all, if you don’t know what’s coming next, how can you invest smartly in AMD?
As a result, the half-day series of presentations is filled with small chunks of information about company-wide products and plans. Everything here is high level and don’t expect AMD to distribute a floor plan for Zen 4 transistors. But it’s easy to see AMD’s product plans for the next few years.
It’s always the Zen architecture roadmap that launches FAD2022 with AMD’s most interesting updates. The Zen architecture, which is the basis of AMD’s recovery and resurgence into a competitive and competent player in x86 processor space, is the foundation of everything from AMD’s smallest embedded CPUs to the largest enterprise chips. Therefore, going through the pipe in the next few years is a huge problem for AMD, and for the industry as a whole.
Zen 4: Improved performance and performance per watt, shipped later this year
AMD is currently in the process of enhancing its Zen4 architecture-based product.This includes the Ryzen 7000 (Raphael) client CPU and its fourth Generation EPYC (Genoa) server CPU. Both of these will be available later this year.
So far, we’ve seen a piece of information about Zen 4, but recently we’ve announced the Ryzen 7000 at Computex. Zen 4 brings new CPU core chipsets and new I / O dies, adding support for features such as PCI-Express 5.0 and DDR5 memory. In terms of performance, AMD also aims to significantly improve performance per watt and clock speed compared to current Zen 3 based products.
AMD, meanwhile, is following Computex’s announcement by clarifying a few things. In particular, the company is addressing questions about the expectations of the Instruction per Clock (IPC), stating that Zen 4 expects an 8-10% improvement in IPC over Zen 3. The first Computex announcements and demos seem to suggest most of AMD’s performance. Gain was due to improved clock speeds, and AMD is working to accommodate it without spending months after the product launches.
Coupled with that, AMD also discloses that it expects an overall single-threaded performance improvement of over 15% with a focus on “greater than”. Since ST performance is a combination of IPC and clock speed, AMD has not locked down the final clock speed at this time, so it cannot be overly specific. However, as we saw in the Computex demo, for low thread workloads, 5.5GHz (or higher) is now in the Zen4 table.
Finally, AMD has also confirmed that Zen 4 will have ISA extensions for AI and AVX-512. At this point, is either (or both) of these extensions included in all Zen 4 products, or a subset – AVX-512, for example, consumes a bit of space and power – but at least these are It makes sense to expect it to appear in Zen4 server parts. CPU AI performance is already on the battlefield for chip makers, so adding AI instructions will allow AMD to catch up with Intel and other competitors in the short term. How this affects AMD’s competitiveness depends in most cases on the instructions (and data types) added.
AMD produces three flavors of Zen4 products. This includes the Vanilla Zen 4 core and the previously announced Zen 4c core. It is a compact core for high density servers and will be installed in a 128 core EPYC Bergamo processor. AMD has also confirmed for the first time that there are also Zen 4 parts with V-Cache. This is new information, but not surprising given the success of AMD’s V-Cache consumer and server parts.
Interestingly, AMD plans to use both 5nm and 4nm processes for the Zen4 family. We already know that Ryzen 7000 and Genoa will use one of TSMC’s 5nm processes and the Zen4c chiplet is configured to be built with the HPC version of N5. So it’s not immediately clear where 4nm fits into AMD’s roadmap, but AMD’s terminology here is because TSMC’s 4nm process is a derivative of 5nm (rather than a completely new node). It cannot be ruled out that you are playing a little faster and looser. It is usually initially classified as a 5nm variant.
At this point, AMD expects Zen 4 to perform more than 25% better per watt than Zen 3 (based on a desktop 16C chip running CineBench). On the other hand, the overall performance improvement is over 35%, both with the per-thread architecture performance improvement and the higher TDP previously disclosed by AMD (especially useful for improving MT workload performance). There is no doubt that you are using it. And yes, these are terrible graphs.
Zen 5 Architecture: A brand new microarchitecture for 2024
Meanwhile, AMD’s Zen architecture roadmap will be introduced in 2024 with the Zen5 architecture announced today. Given that AMD hasn’t shipped Zen 4 yet, the details about Zen 5 are, of course, at a very high level. Nonetheless, they also show that AMD isn’t content with their glory and that some aggressive updates are planned.
The big news here is that AMD calls the Zen5 architecture a “whole new microarchitecture.” In other words, it’s not just a gradual improvement over Zen4.
In reality, no major vendor designs a CPU architecture from scratch. There’s always enough to reuse, but the message from AMD is clear. In other words, we plan to significantly rewrite the core CPU architecture in order. To further improve performance and energy efficiency.
As for AMD’s disclosure for now, Zen 5 will re-pipeline the front end and expand its issuance again. The devil has more details here, but since it comes from Zen 3 and its 4 instruction / cycle decode rate, it’s easy to understand why AMD wants to focus next. Especially on the backend, the integer unit already has 10. Wide issue width.
Meanwhile, in addition to Zen 4’s new AI instructions, Zen 5 also integrates AI and machine learning optimization. AMD hasn’t said much else here, but it has an important library of tools that cover everything from AI-focused instructions to the addition of support for more data types.
AMD expects the Zen5 chip stack to be similar to Zen4. In other words, we expect the three designs of Vanilla Zen 5 core, compact core (Zen 5c), and V-Cache to be the same. You have enabled the core. This kind of continuity is very important for AMD customers. This ensures that AMD’s more bespoke configurations (Zen 4c and V-Cache) will be the successors for the period after 2024. From a technical point of view, this isn’t all that surprising, but from a business point of view, we want to make sure our customers aren’t adopting dead-end hardware.
Finally, AMD is planning an interesting manufacturing mix for Zen5. The Zen5 CPU core is manufactured with a combination of 4nm and 3nm processes, but unlike Zen4’s 5nm / 4nm mix, TSMC’s 4nm and 3nm nodes are very different. 4nm is an optimized version of 5nm, while 3nm is a brand new node. Therefore, if AMD’s manufacturing plans are as advanced as they are today, Zen5 will span major node jumps. That said, it’s not unreasonable to suspect that AMD is hedging their bets here and leaving 4nm on the table in case 3nm isn’t as far as they want.
In summary, the Zen 5 architecture is planned for 2024. Looking at Zen 3 and Zen 4, AMD has released / will be released later, but hasn’t provided any further information about when it will be in the year. 2020 and 2022, respectively. Therefore, the H2 / EOY 2024 is as good as a guess.
