AMD Zen 5 Core CPU Arch Alleged to Be over 40% Faster Than Zen 4 | Hacker News Reader

AMD Zen 5 Core CPU Arch Alleged to Be over 40% Faster Than Zen 4

hothardware.com

43 pointssigna1122 comments

This is essentially rumor on the basis of an anonymous forum comment saying 40% on SPECint. Cool if it's true but we cannot evaluate that and there are zero interesting details.

A lot of people thinks this is impossible or BS. But assuming this is on a clock per clock basis, Zen 5 would still be ~10% slower than A17. And we are expecting Apple to get a little bit faster with A18.

The only difference would be we can expect Zen 5 to scale up to much higher clock speed.

As noted in the article, it is known that Zen 5 is a new microarchitecture, not just one with small incremental changes, like Zen 4 over Zen 3.

Therefore it is expected that the performance increase at the same clock frequency will be substantial, at least 20% to 25% on average.

It is also expected that some applications will be accelerated much more than the average. Usually SPEC has a performance increase not too different from the average, so 40% seems high, but not impossible, because the IPC would remain lower than for the Apple CPUs.

By the end of the year, both AMD with Zen 5 and Intel with Arrow Lake and Lunar Lake are expected to introduce new microarchitectures, after several years with only minor microarchitecture improvements in Zen 3/Zen 4 and Alder Lake/Raptor Lake/Raptor Lake Refresh/Meteor Lake.

Hopefully both will demonstrate good advancements over the current CPUs.

>Therefore it is expected that the performance increase at the same clock frequency will be substantial, at least 20% to 25% on average.

This is absolutely not given. So I disagree with the word "expected". I cant even count how many new uArch we have had in history that simply fail to perform. Both AMD and Intel or even outside x86.

So having a new uArch that has this huge increase in performance is exciting. Although arguably most of what Intel and AMD are doing is simply looking and Apple and how it could be applied to x86.

> Zen 5 would still be ~10% slower than A17

Stange that they would lose ground in a new architecture. Zen 4 is currently ~10% faster than the A17. In addition, that's for single-core synthetic performance, which is a completely useless metric. Ignoring AMD's core count, or Intel's awesome e-cores, is like ignoring Apple's unbeatable perf/watt.

Apple remains a great mobile processor. This may change with the A18 but the A17 certainly isn't competitive with the reckless horsepower of desktop CPUs.

>Apple's unbeatable perf/watt

It is worth noting that, by default, Ryzens run at waaay higher power than where their best perf/watt point is. Using the power limit option, it is easy to get 80% of the performance for 35% of the power or 90% of the performance for 50% of the power. Apple ARM still wins (especially in idle), but it's not such a blowout anymore.

https://www.anandtech.com/show/17641/lighter-touch-cpu-power...

> Stange that they would lose ground in a new architecture.

Is it strange though? AMD needs to ensure backwards compatibility with x86 and AMD64 architectures, and can't simply tell it's customers to just run their old software on an emulator and hope it does not crash.

Apple can and does do just that, and as it controls all of its ecosystem then it can invest what it can to micro-optimize and fine-tune any layer of their tech stack to meet specific needs.

The way I see it:

Intel if you want awesome common man performance (IdleM@ster[1] and best single thread performance).

AMD if you want to rip all the threads (best multithread performance).

Apple M if you want to save power (goes without saying).

[1]: https://news.ycombinator.com/item?id=38829531

The X3D models are also ridiculously efficient desktop CPUs. Perfect for ITX builds.

AFAIU, you can get very similar efficiency by setting power limits on non-X3D models. The X3D are just more power limited at stock.

Most desktop chips are run far beyond their perf/watt sweet spot, so this is true - but you would loose the stellar gaming performance.

Most AMD desktop chips. Intel chips have a fairly linear power vs performance curve, i.e. you will lose significant performance when you reduce power.

Zen 4 is faster due to a greater IPC * clock frequency product.

In IPC, the Apple CPUs continue to be much better than the Intel and AMD CPUs, which is the reason that while the latter have greater absolute performance, the former have greater energy efficiency.

Mostly true, but IPC is not the only thing influencing power consumption - process node also plays a significant part, so Apple will be ahead as long as they use the newest nodes at TSMC.

Apple's IPC mostly comes from their 8-wide instruction decoder that isn't possible in x86_64 because of variable instruction lengths.

> the A17 certainly isn't competitive with the reckless horsepower of desktop CPUs

Except when they put it into a laptop/desktop and crank the power up to typical levels ...

On a Clock per Clock basis. Not absolute performance.

Absolute performance and absolute performance per watt - given a specific set of tasks - are the only metrics that matter in practice (for desktop and mobile/data center respectively). nm (which actually has had no meaningful correlation to size for at least a decade), IPC, or whatever other metrics are only actually useful to two audiences: the people designing and fabbing the chip.

"Apple would be faster than AMD/Intel if it ran at 5.5GHz," well it doesn't, does it? Even on the desktop SKUs. The high IPC is a specific trade-off that the designers made in order to hit targets that they care about. We aren't yet at the point where one chip is good at everything (well, not entirely true, Epyc does hit an unreasonable amount of targets).

In this case it looks like they weren’t talking about what matters “in practice,” but whether the improvement would indicate that some aspect of the design must have improved to an implausible degree. So, which type of benchmark you think is best seems slightly tangential to the issue.

"per clock" is IPC, which was the entire point of my comment.

If AMD were running at 4GHz they might have had comparable IPC, or they might not because they could use the headroom to target something else. But they aren't at 4GHz. Nor are they ARM. It's all Apple marketing speak (others aren't innocent of this nonsense either).

Comparing IPC across the same chip line does make sense, to get a rough idea if how far the envelope is being pushed. Even comparing M3 to A17 would be extremely questionable.

What does "absolute" performance mean in this context? All performance measurements are relative to the use cases the measurements were designed around. The iPhone's CPU would be awful in a gaming computer and the Ryzen would be a terrible CPU for a smart watch.

In common CPU benchmarks, the M3 gets blown out of the water by the 7950X3D, even in single core performance. Terrible CPU to have in a laptop, but in a workstation or desktop computer there's no competition which CPU to pick.

Clock per clock is a bad metric in comparing how fast CPUs are, it's like normalizing for engine CC or RPM. CPU designers make a chip do more work per clock cycle while trading off how fast it can clock, and vice versa.

(It can be a useful number in analyzing CPU design tradeoffs or other details of course)

The metric work done per clock is not a useful one. Efficiency, i.e. energy spent per work done is a lot better. The latter, however, is heavy influenced by the process node - so comparing different processes when it comes to efficiency is rather futile.

Higher IPC may mean inability to clock higher. In applications where work done per time matters (say games, HFT) it also doesn't provide much.

Reply on news.ycombinator.com