Zen 4 on the Floor: AMD Promises 35 Percent Performance Jump For Next-Gen CPUs

This site may earn affiliate commissions from the links on this page. Terms of use.

At its Financial Analyst Day today, AMD shared its upcoming CPU roadmap and associated node transition plans for the next few years. While the company has a lot to say about various aspects of its business, it’s the Zen 4 CPU announcements we want to touch on today. Zen 4 is apparently a much larger improvement over Zen 3 than AMD had previously disclosed.

The Zen “4C” core referenced here is Bergamo, AMD’s tweaked CPU design with support up to 128 cores. The standard Zen 4 core, Genoa, tops out at 96 cores. AMD is planning to manufacture chips with TSMC’s 4nm node in the future, though it hasn’t specified which CPUs will use it. According to TSMC, 4nm is an enhanced 5nm node with some small area savings (~6 percent) but lower process complexity “via mask reduction” and better performance/power “through device and BEOL enhancement.” BEOL stands for Back End of Line — it’s a reference to a specific step in the IC fabrication process, where interconnects are laid down and metal layers are deposited. AMD hasn’t made a very big deal over its use of 6nm and it might not make much noise over 4nm, either.

One thing AMD confirmed today is that V-Cache will appear on future Zen 4 client desktop CPUs, though the company refrained from saying much about how many SKUs might offer the extra L3. V-Cache makes the Ryzen 7 5800X3D a particularly interesting upgrade option for older Ryzen owners with slower DRAM, as we recently discussed. AMD didn’t give a timeline for V-Cache introduction with Zen 4, however, so it may not launch when the rest of the chips debut.

Zen 4 Hits the Gas on Frequency

When it comes to Zen 4, it appears AMD might have been sandbagging a bit at Computex. Back then it said its upcoming architecture will offer ~>15 percent single-threaded performance uplift over Zen 3, but that was the only number it threw out. Now AMD says it’s looking at greater than 35 percent increase for Zen 4 over Zen 3, though it’s not clear if that figure refers to single-threaded performance, multi-threaded performance, or a mix between the two. The company also predicts that Zen 4 will offer a 25 percent performance per watt improvement over Zen 3.

Right now it looks as if most of these gains are likely to come from clock speed. AMD recently showed a 16-core CPU running at 5.5GHz and claimed nothing beyond an off-the-shelf AIO cooler was required to hit these clocks. If we assume a all-core maximum frequency increase from 4.5GHz to 5.5GHz — which is honestly enormous — AMD would need to offer approximately 1.1x additional IPC to hit a 1.35x performance improvement. AMD has raised its CPU TDP to 170W and maximum socket power to 230W, so the company apparently plans to give Zen 4 a little more room to breathe in 2022.  Mark Papermaster claimed an 8-10 percent gain in IPC from stage today, so the math checks out. If anything, a 1.15x performance improvement for single-thread seems low.

Hitting frequencies like this while simultaneously improving IPC and preventing performance per watt from falling implies TSMC’s 5nm is hitting clock frequencies Intel might well envy. We’ve asked AMD for more details on how it improved its frequencies so dramatically but the company isn’t ready to talk about the architecture at that level of detail yet.

Hitting these performance targets would more than close the gap with Alder Lake — it would also put the company on a solid footing against Intel’s upcoming 13th Gen platform, codenamed Raptor Lake. The implication of AMD’s Financial Analyst Day is that Intel can’t count on an easy win against its smaller rival. All manufacturer data should be taken with a grain of salt, and AMD is no exception, but AMD’s CPU business has nailed its projections for years now. A 1.35x performance increase is larger than expected, but AMD’s publicly demonstrated clock speeds suggest a path for getting there.

Managing Editor Joel Hruska contributed to this article.

Now Read: