Hey, AMD: It’s time for Ryzen’s rebel moment

The PC industry is currently obsessed with power efficiency, long battery life, and AI. But who’s going to stand up for the people who want raw performance and don’t care about anything else?

That performance advocate could be you, AMD. You’ve got what it takes. So dig out that bomber jacket, put on those aviator shades, and lean in. Be the rebel chip vendor we want you to be.

This isn’t a story about AMD executives smashing a server with a sledgehammer, or spray-painting graffiti all over Intel’s headquarters. In fact, at its recent Tech Day, AMD spent two days with reporters talking over PowerPoint presentations about IPC uplift, data paths, and the time to first token. But the subtext was how well positioned both the Ryzen 9000 and the Ryzen AI 300 are to take over the PC market.

There were two clear messages from AMD: First, AMD’s Ryzen 9000 (Granite Ridge) and its Ryzen AI 300 (Strix Point) are nearly here — the latter will ship July 28, according to retailers and manufacturers. Intel’s Lunar Lake is decidedly not so imminent, giving AMD an advantage.

Second, AMD’s chips will deliver performance that is significantly above what Intel will offer, and without the compatibility issues that have hindered Windows on Arm laptops from Qualcomm’s Snapdragon X Elite. Unfortunately, we still don’t know the prices of the Ryzen 9000 processors, even though those chips will ship on July 31, AMD said.

Building Ryzen for speed

AMD executives began with the abstract — the Zen processor architecture itself — then drilled down to the two chips themselves, and from there on to the tools that users can take advantage of to tweak performance further.

Both the Ryzen 9000 and AI 300 are based upon AMD’s latest Zen 5 architecture, the fifth iteration since AMD originally launched Zen in 2017. Mark Papermaster, AMD’s chief technical officer, said that performance is the overarching goal, with the general Zen 5 architecture achieving on average 16 percent more IPC (instructions per clock) performance than its predecessor, the short-lived “Hawk Point” CPU inside December’s Ryzen 8000 chip. He also confirmed that AMD does plan an eventual Zen 5c core, as expected but not previously confirmed.

“What we said at the time [in 2017] was that…we’re going to come out, consistently, with bringing more performance to x86 CPUs,” Papermaster said. “We’re going to simply provide outright leadership — that was the goal that we set and that’s what we’ve done. And what I can tell you is that there will be no letups, and Zen 5 will not disappoint you with the kind of performance improvements that we’ve brought and we won’t let up going forward. We’re doubling down on Zen 5 and what you’re going to see is that it really represents a huge leap forward. And in fact, it’s going to be a pedestal that we’re going to build upon the next several generations.”

Papermaster supplied some internal details of the Zen 5 architecture, such as a 512-bit datapath for processing AVX-512 instructions. The integrated RDNA 3.5 graphics engine — which AMD improved after a collaboration with Samsung in the mobile space — delivers 32 percent faster performance in 3DMark’s Time Spy benchmark than the Ryzen 8000. The chip also includes what Papermaster called a “breakthrough” improvement of its math acceleration unit, which delivers a 32 percent improvement in machine learning for AI applications.

Zen 5 will use both 4nm (Strix Point) and 3nm technology, Papermaster said, leaving open the question of when Zen 6 will ship and on what process technology.

AMD’s Ryzen AI 300, meanwhile includes two CCD core clusters and “enhanced boost sensitivity” that apparently will allow them to transition quickly into a boost state, increasing performance. Papermaster also reminded the audience that the Ryzen 9700X is on average 12 percent faster in gaming than the cache-laden 105W gaming monster Ryzen 5800X3D, which debuted just two years ago in 2022.

Joe Macri, AMD’s chief technical officer of the client division, said that the AM5 socket that the 9700X fits into was designed to live “six, seven years — four generations of CPU cores” after its launch in September 2022. “So many teams come together to build the same platform and ensure that longevity so that you guys get the ultimate experiment,” he said. “You don’t have to go change the motherboard every other generation, like some other folks.”

AMD even pulled off what it claimed to be a world record in overclocking, achieving a Cinebench R20 multithreaded score of 55,046 on a 5.7GHz (stock) Ryzen 9 9950X overclocked to 6.494GHz using liquid nitrogen. (The chip hit 6.698GHz on a later attempt, but the PC crashed.)

AMD achieves what the company claimed was a world record on Cinebench R23, with a 6.5GHz overclock on the Ryzen 9 9950X.

New overclocking tools for Ryzen chips

While you may not have a vat of liquid nitrogen in your garage, you will have access to new AMD software to push your Ryzen even more aggressively than before.

Part of the overall performance boost you can expect is an addition to AMD’s overclocking tools. AMD already has a feature called Precision Boost Overdrive, which is the simplest way to get more out of your system. PBO’s “one-click overclocking” allows you to get 15 percent more performance out of the Ryzen 7 9700X, executives said, with 6 percent boosts on the 9600X and 9900X.

Beginning with the Ryzen 7000, AMD implemented a feature called Curve Optimizer that allowed users to tweak the settings within the Ryzen Master utility software, specifically with regard to Precision Boost Overdrive and the power management framework. While some users prefer to overclock their systems, consuming more power, others prefer to undervolt their PC, trying to maintain their performance while using less power and lowering the system temperature. Curve Optimizer was a way to “shift” the curve by varying the voltage as the CPU went up and down in frequency.

Curve Shaper, the latest tool, offers users more flexibility. Essentially, a user will be able to take the tool and set their own steps depending upon temperature and frequency. (Users will have three temperature and five frequency bands to work with.) The process will be manual, and users will have to come up with their own adjustments. However, AMD executives said they imagine that the community will develop and share their own configurations.

“This allows users to further reduce voltage from bands that are stable, and add voltage for bands where instabilities are observed,” AMD said. (That’s important, given that a CPU and a PC can crash if it doesn’t receive enough power.)

On the memory side, AMD’s AGESA base firmware supports up to overclocked DDR5-8000, JEDEC support for DDR5-5600, and with memory overclocking enabled for all of AMD’s Ryzen chipsets.

For now, AMD’s NPUs lead in performance, too

In AI, AMD too is a leader. (For now, anyway — things change quickly in the AI space.) But the problem isn’t really increasing performance — it’s managing it.

AMD’s Strix Point NPU supplies 50 TOPS (trillions of operations per second) of AI power versus the 48 (Intel’s Lunar Lake) and 45 (Qualcomm Snapdragon X Elite) its rivals provide. Vamsi Boppana, AMD’s senior vice president of its AI group, claimed that AMD was up to 5X faster than the current Intel Core Ultra 7 155H in “time to first token,” or how responsive an LLM (large language model) is when running on an AMD NPU.

Typically, the more parameters an AI chatbot or LLM offers, the more accurate it is. But the more parameters an LLM offers, the slower it runs and the more memory and memory bandwidth it requires.

An inaccurate LLM is useless; a glacially slow one is too. AMD is working to implement what’s called a “block FP16” format for ISPs, to allow the accuracy of large language models with the processing speed of small language models. AMD also showed off a number of apps, including those used by the PC makers themselves, to demonstrate the power of the NPU.

The bottom line is this: Do you want a productivity PC that can last all day? Or an AI machine? A gaming powerhouse? If you look at the latest Ryzen 9000 series, you’ll see that the low-end Ryzen 5 and Ryzen 7 chips have cut their power quite a bit. There’s nothing wrong with that.

But consider what you’ve heard from AMD over the years: Threadripper. Ryzen. The X3D family. These aren’t designed to power the wishy-washy PCs that you’ll find at your local library. These are for true PCs, the kind that you grew up on, and I was surprised that AMD didn’t seize the chance to fully embrace them. Low-power PCs are fine. But the world still needs a muscle car.

Editor’s Note: AMD paid for hotel, airfare, and some meals to fly PCWorld and other journalists to its Tech Day event in Los Angeles. AMD did not in any way control the editorial process, including what topics to write about.

CPUs and Processors