AMD Has Scaled Ryzen

Skype October 10, 2020 Technology Leave a comment 226 Views

Ryzen2-Feature

Note: This article assumes that Zen 3 will deliver the 1.19x IPC improvement it has promised over the Zen 2 / Ryzen 3000 CPU family. AMD’s Zen disclosures have consistently been accurate enough that I am comfortable making that assumption.

When AMD launches Zen 3 on November 5, it isn’t just going to be another iteration of the company’s CPU family. Mathematically, it’s going to break — or at least match — one of its own records that it hasn’t challenged for nearly 15 years. Statistically, the amount of cumulative improvement delivered between March 2, 2017, and November 5, 2020, represents the highest, fastest scaling AMD has delivered at any point since the year 2000 — including during the Athlon 64/Athlon 64 X2 era.

The Athlon 64 3200+ launched at 2GHz (single-core) on September 23, 2003, while the Athlon 64 X2 6000+ launched at 3GHz on February 20, 2007. That’s a 1.5x improvement in core clock and a 2x improvement in core count delivered over 1,246 days. IPC improvements over this period from architectural changes were very small; most gains came from moving to dual-core, using dual-channel memory via Socket 939 as opposed to single-channel Socket 754, and increased CPU clock.

The Ryzen 7 1800X launched on March 2, 2017, while the Ryzen 9 5950X launches on November 5, 2020. The Ryzen 7 1800X tended to clock around 3.7GHz under load. We don’t know how the Ryzen 9 5950X will clock under load, but ~3.8GHz was typical for full-load on the 3950X, so let’s assume a very modest improvement of 100-200MHz.

Over the course of 1,344 days, AMD will have once again doubled its top-end CPU core count and improved effective IPC (which is to say, IPC + sustained higher clocks) by ~1.51x. If AMD has improved the Ryzen 9 5950X’s all-core boost by 100-200MHz, that’s a further 2.6 – 5.2 percent improvement, which would cumulatively match or exceed the theoretical scaling performance AMD delivered during the most competitive time in its history. If anything, I’d expect modern scaling to be better than what we saw on Athlon 64/Athlon 64 X2, simply because far more software is multi-threaded now compared with then. And if AMD can squeeze a little more clock out — say, 4.2 – 4.4GHz — it would cleanly beat the previous time period. The Cinebench R20 benchmark Lisa Su showed implied AMD may have done just that.

The benchmarks AMD showed during its Zen 3 unveil showed a single-threaded CB20 score of 631 and claimed a 640 for the 5950X. Zen 1 scored a 381 in the same benchmark back in 2017.

Cinebench R20 is one test, these are AMD-provided numbers, and feel free to eat a salt shaker if you feel it necessary. But then sit down and think about the fact that AMD has improved single-threaded performance by ~1.68x in 3.5 years. For comparison, in 2017, an Intel Core i7-7700K scored 446 in CB20. The Core i9-10900K scores 518 today. That’s a gain of 1.16x for Intel in single-threaded performance over the same amount of time. AMD may have started from farther behind, but they’ve clearly outpaced Intel.

This achievement is even more impressive when you consider that from 2003-2007, AMD could still take advantage of the waning days of Dennard scaling. From 2017-2020, the company had no such luxury. AMD has matched or exceeded — and I’m betting on “exceeded” — its fastest performance scaling in two decades. It has done so without relying on clock speed as its primary method of improving performance, though clock gains have contributed.

To beat the cumulative rate of scaling AMD claims it will deliver from March 2017 to November 2020, you have to go all the way back to 1999, when Dennard scaling was still firing on all thrusters.

An original Slot A Athlon. Image by Tullius from Wikimedia Commons, provided under the GNU Free Documentation License

On June 23, 1999, AMD launched the original Athlon, at speeds of up to 600MHz. The arrival of the K7 heralded a new chapter in AMD’s competitive war against Intel, dubbed the “Gigahertz Wars.” On June 5, 2000, AMD launched the Athlon “Thunderbird” 1GHz CPU.

This technically wasn’t AMD’s first 1GHz chip, but it was the first with an integrated, full-speed L2 cache that brought AMD up to par with what Intel was delivering in that regard with the Coppermine-based Pentium 3, so I’m counting it as a better marker of capability than the earlier slot-based 1GHz chip, with its L2 clocked at 1/3 CPU speed.

On June 5, 2000, AMD could claim to have delivered a 1.66x boost in clock speed, as well as full-speed L2 cache, in under a year. Original reviews of the CPU show that AMD picked up between 1.3x and 1.5x by moving from the old, slot-based Athlon 600 to the Thunderbird-based Athlon 1000. These types of radical performance leaps are well and truly gone until someone invents a new post-CMOS architecture, so I wouldn’t look to any company to return to the heady days of what we took for granted at the time.

But the fact that AMD has matched or exceeded the performance scaling it offered during its first golden age says volumes about the company AMD is today versus the company it has been at any previous point in history. For years, AMD was notorious for its ability to snatch defeat from the jaws of victory, but if you think that phrase still describes the company today, you haven’t been paying attention. Zen wasn’t luck and Zen 2 wasn’t a fluke. AMD continues to deliver exactly the CPU performance improvements it promises, year after year.

Intel, needless to say, has not kept up with AMD’s onslaught. Coffee Lake was an excellent upgrade over Kaby and chips like the Core i9-9900K temporarily held off Zen+, but given the size of the gains AMD is predicting with Zen 3, Intel’s Rocket Lake will need to be something truly special to put the larger company in a desktop leadership position again.

AMD’s second golden age shows no signs of slowing down.

Now Read: