News Apple's A18 delivers solid performance gains, challenges Ryzen 9 9950X in single-core Geekbench

[Admin]

Apple's vanilla A18 shows solid performance in Geekbench, challenges M4 and AMD's Ryzen 9 9950X.

Apple's A18 delivers solid performance gains, challenges Ryzen 9 9950X in single-core Geekbench : Read more

 

[Cooe]

🤦... For the love of freaking god, can we PLEEEEEASE stop pretending that Geekbench is an even REMOTELY accurate representation of modern AMD CPU performance??? That hasn't been the case for basically the entire time the Zen macroarchitecture has existed, but got DRAMATICALLY WORSE after Zen 2! 🤷

(Let ALONE when trying to compare CPU's across different ISA's like x86 to ARM! Which is risky, questionable business even WITH properly optimized/scoring CPU architectures like something from the modern branches of the Intel Core macroarchitecture or ARM's modern core designs.)

Just look at Threadripper multi-core performance scaling in Geekbench vs literally any actual real-world heavily multi-threaded CPU workload! It's absolute freaking NONSENSE GARBAGE!!!

In closely related news, CPU-Z is another popular synthetic benchmark that's become essentially WORTHLESS for modern AMD CPU's, but at least in that case it didn't start until 2022's Zen 4.

(That said, it's still totally INEXCUSABLE that it remains utterly broken/unoptimized on Zen 4/5 to this very day! And even more so that Geekbench remains borked for practically every single Zen core ever! [Although it gets notably worse the newer the Zen architecture is.])

 

[Pierce2623]

🤦... For the love of freaking god, can we PLEEEEEASE stop pretending that Geekbench is an even REMOTELY accurate representation of modern AMD CPU performance??? That hasn't been the case for basically the entire time the Zen macroarchitecture has existed, but got DRAMATICALLY WORSE after Zen 2! 🤷

(Let ALONE when trying to compare CPU's across different ISA's like x86 to ARM! Which is risky, questionable business even WITH properly optimized/scoring CPU architectures like something from the modern branches of the Intel Core macroarchitecture or ARM's modern core designs.)

Just look at Threadripper multi-core performance scaling in Geekbench vs literally any actual real-world heavily multi-threaded CPU workload! It's absolute freaking NONSENSE GARBAGE!!!

In closely related news, CPU-Z is another popular synthetic benchmark that's become essentially WORTHLESS for modern AMD CPU's, but at least in that case it didn't start until 2022's Zen 4.

(That said, it's still totally INEXCUSABLE that it remains utterly broken/unoptimized on Zen 4/5 to this very day! And even more so that Geekbench remains borked for practically every single Zen core ever! [Although it gets notably worse the newer the Zen architecture is.])

I’m a fan of AMD but I’m also a fan of transparency. AMD vs Intel in single threaded Geekbench is actually a fairly good representation of how they perform against each other in an average real work load. Of course it doesn’t scale well with Threadrippers, it’s a consumer biased benchmark for chips up to about 32 threads. Creating a synthetic workload that properly illustrates single threaded performance while scaling well to an infinite number of cores is basically impossible. Regardless you can’t show me that AMD does significantly worse in Geekbench vs Intel than they do in Cinebench vs Intel. That being said, Arrow Lake isn’t shaping up any better than Zen 5.

 

[parkerthon]

🤦... For the love of freaking god, can we PLEEEEEASE stop pretending that Geekbench is an even REMOTELY accurate representation of modern AMD CPU performance??? That hasn't been the case for basically the entire time the Zen macroarchitecture has existed, but got DRAMATICALLY WORSE after Zen 2! 🤷

(Let ALONE when trying to compare CPU's across different ISA's like x86 to ARM! Which is risky, questionable business even WITH properly optimized/scoring CPU architectures like something from the modern branches of the Intel Core macroarchitecture or ARM's modern core designs.)

Just look at Threadripper multi-core performance scaling in Geekbench vs literally any actual real-world heavily multi-threaded CPU workload! It's absolute freaking NONSENSE GARBAGE!!!

In closely related news, CPU-Z is another popular synthetic benchmark that's become essentially WORTHLESS for modern AMD CPU's, but at least in that case it didn't start until 2022's Zen 4.

(That said, it's still totally INEXCUSABLE that it remains utterly broken/unoptimized on Zen 4/5 to this very day! And even more so that Geekbench remains borked for practically every single Zen core ever! [Although it gets notably worse the newer the Zen architecture is.])

You never said what is a good test to compare the cpus that you mention have an unfair shake vs apple chips. Or are you saying it’s not possible?

 

[rjranay]

at the end of the day it still boils down which architecture is supported by your day to day applications

 

[CelicaGT]

All that performance on the table, utilized by most to doomscroll through TikTok...

 

[TheHerald]

🤦... For the love of freaking god, can we PLEEEEEASE stop pretending that Geekbench is an even REMOTELY accurate representation of modern AMD CPU performance??? That hasn't been the case for basically the entire time the Zen macroarchitecture has existed, but got DRAMATICALLY WORSE after Zen 2! 🤷

(Let ALONE when trying to compare CPU's across different ISA's like x86 to ARM! Which is risky, questionable business even WITH properly optimized/scoring CPU architectures like something from the modern branches of the Intel Core macroarchitecture or ARM's modern core designs.)

Just look at Threadripper multi-core performance scaling in Geekbench vs literally any actual real-world heavily multi-threaded CPU workload! It's absolute freaking NONSENSE GARBAGE!!!

In closely related news, CPU-Z is another popular synthetic benchmark that's become essentially WORTHLESS for modern AMD CPU's, but at least in that case it didn't start until 2022's Zen 4.

(That said, it's still totally INEXCUSABLE that it remains utterly broken/unoptimized on Zen 4/5 to this very day! And even more so that Geekbench remains borked for practically every single Zen core ever! [Although it gets notably worse the newer the Zen architecture is.])

The single thread portion of Geekbench is actually fairly decent. In other words, nah, you are wrong.

 

[dada_dave]

"Hence, the intrigue is whether the eight-core A18 Pro can challenge Apple's M3."

The A18 Pro CPU is still 6 cores, it just has access to a larger cache (presumably SLC rather than L1/L2). The A18 Pro might be ever so slightly faster because of that extra cache in certain workloads and I think there's a difference in thermal designs of the chassis the chips are put in which may affect tests with greater emphasis on endurance than GB, but otherwise the A18 and A18 Pro should be pretty similar for CPU speed - which is what we see in the current pre-release benchmarks:

iPhone 16's A18 Pro chip outperforms the M1 chip in new benchmarks [Updated] - 9to5Mac

We got our first look at a Geekbench result from the iPhone 16 earlier this week, with somewhat disappointing results....

9to5mac.com

 

[parkerthon]

All that performance on the table, utilized by most to doomscroll through TikTok...

Not wrong since most people are oblivious that the more you watch some topic, the more you are suggested to consume more… but I laugh my ass off on TikTok when I choose to dive into it(no alerts, ever). The ones after the debate were fantastic.

 

[federal]

🤦... For the love of freaking god, can we PLEEEEEASE stop pretending that Geekbench is an even REMOTELY accurate representation of modern AMD CPU performance??? That hasn't been the case for basically the entire time the Zen macroarchitecture has existed, but got DRAMATICALLY WORSE after Zen 2! 🤷

(Let ALONE when trying to compare CPU's across different ISA's like x86 to ARM! Which is risky, questionable business even WITH properly optimized/scoring CPU architectures like something from the modern branches of the Intel Core macroarchitecture or ARM's modern core designs.)

Just look at Threadripper multi-core performance scaling in Geekbench vs literally any actual real-world heavily multi-threaded CPU workload! It's absolute freaking NONSENSE GARBAGE!!!

In closely related news, CPU-Z is another popular synthetic benchmark that's become essentially WORTHLESS for modern AMD CPU's, but at least in that case it didn't start until 2022's Zen 4.

(That said, it's still totally INEXCUSABLE that it remains utterly broken/unoptimized on Zen 4/5 to this very day! And even more so that Geekbench remains borked for practically every single Zen core ever! [Although it gets notably worse the newer the Zen architecture is.])

Geekbench was originally developed for Apple products (but back when they used Intel CPUs). and its core code is still primarily developed on and for Apple products and their instruction sets, both hardware and OS calls. It's ported to other products but not as optimized or tuned for those I/O calls and chip instruction sets, so the results for Qualcomm, AMD, or Intel chips will be somewhat random - sometimes over, more often under performing versus most anecdotal user experiences.
It's not intentionally trying to favor anything, but you have to develop and initially test on something, and that something is going to be the most consistent (and usually best performing) in terms of benchmark comparability across CPU families. It seems it's still mostly Macs, for Geekbench (but they're closed source at this point, making it hard to say for sure).
All benchmarks are tuned by weighting a mix of instructions and data to arrive at a score. That weighting has to be tuned on something. For an example of how much weighting affects scores, watch frame rates change on a game as it's played. Same app, same hardware, same everything except for the particular mix of calls being made at any given time. Different OS's and hardware do better or worse depending on the instruction and where and how much data is being loaded/stored. You have to tune for something...

 

[dada_dave]

Geekbench was originally developed for Apple products (but back when they used Intel CPUs). and its core code is still primarily developed on and for Apple products and their instruction sets, both hardware and OS calls. It's ported to other products but not as optimized or tuned for those I/O calls and chip instruction sets, so the results for Qualcomm, AMD, or Intel chips will be somewhat random - sometimes over, more often under performing versus most anecdotal user experiences.
It's not intentionally trying to favor anything, but you have to develop and initially test on something, and that something is going to be the most consistent (and usually best performing) in terms of benchmark comparability across CPU families. It seems it's still mostly Macs, for Geekbench (but they're closed source at this point, making it hard to say for sure).

That's not really accurate. Jon Poole was inspired to write benchmarking software because of the lack of good cross platforming benchmarking software back in the G4/G5 days and he wanted to compare how his Macs were actually running, but Geekbench itself has always been cross platform from the beginning with an emphasis on cross platform. A good history of it here:

Geekbench’s creator on version 6 and why benchmarks matter in the real world

“How hard can it be to write a benchmark? Maybe I should write my own.”…

arstechnica.com

Every test has strengths and weaknesses and GB is no exception but it correlates well with SPEC and other such multi-test CPU benchmark suites. Different versions have had different optimization issues and outright bugs, including some over the years for the Mac, but in general Linux (added in GB 2) has actually tended to be the fastest OS when normalized on the same hardware - which again is mirrored by results from other cross platform benchmarks.

As for the first response here about AMD CPUs, I'm not certain but I'm assuming he's upset by the changes to MT benchmarking in GB 6 which does indeed tend to limit the effectiveness of high core count systems like threadripper on many of its subtests. This is covered in the article but GB 6 is probably more accurate as a result for the target audience as many real world applications run by everyday users are not in fact embarrassingly parallel. GB 6 still has some embarrassingly parallel workloads, it didn't get rid of them entirely, but it is trying to reflect what the average user rather than a workstation user will experience. Thus it isn't that AMD CPUs in particular are being treated unfairly but rather all high core/thread count systems will see diminishing returns in MT GB 6 compared to 5. Again, this was a deliberate choice to focus on certain kinds of workloads and user audience. Poole felt that GB 5 in particular was being misused to over sell the average computer user on systems that they didn't need, couldn't take advantage of, or even worse, were potentially slower than other, cheaper systems on the tasks users actually were doing daily.

 

[_Shatta_AD_]

How did this ‘tech’ article pass editorial AND proof-reader when they make such silly mistakes as “Hence, the intrigue is whether the eight-core A18 Pro can challenge Apple's M3” when Apple in the very same event announced the A18 Pro inside the iPhone 16 Pro(& Pro Max) featuring the same 6 cores(2+4) with a larger cache, 1 more GPU core, extra media engine, ISP, etc.??? Did the writer left the chat right after the iPhone 16 reveal?? 😂

 

[dada_dave]

How did this ‘tech’ article pass editorial AND proof-reader when they make such silly mistakes as “Hence, the intrigue is whether the eight-core A18 Pro can challenge Apple's M3” when Apple in the very same event announced the A18 Pro inside the iPhone 16 Pro(& Pro Max) featuring the same 6 cores(2+4) with a larger cache, 1 more GPU core, extra media engine, ISP, etc.??? Did the writer left the chat right after the iPhone 16 reveal?? 😂

Well before Anandtech's passing, Ryan talked about this and confirmed that sadly dedicated copy editors who proofread stories before publication at tech news sites are an extinct species (and are an endangered one at general news sites from what I can gather). And proofreading itself is often just not done given how tight deadlines are - especially surrounding events and embargoes. Altogether, it's just considered too expensive. And while corporate greed certainly plays a role here, it's not like we're paying subscriptions to most sites and what little ad revenue gets soaked up by Google/Facebook/Tabula (so back to corporate greed) which is why sites like Tom's look like something that would make a Nascar driver blush without Ad blockers or Reader mode (which decreases revenue further) to the point where the site can become literally non-functional with ads causing web page crashes, blocking text and buttons, etc ... Kind of a vicious cycle and expenses like copy editing are often the first thing to get axed.

As for why Anton made this mistake ... well there were a lot of rumors about Apple splitting the A18 and A18 Pro up and some of the most popular of those hypothesized that the A18 Pro might get an extra couple of E-cores in addition to a sixth GPU core. He probably had that stuck in his head even though Apple clearly said otherwise during their presentation. It happens, but things like that is why copy editors (used to) exist.

 

[stellarashes]

Like iPad, all this performance is mostly useless on iOS, and most people wouldn't be taking advantage of this performance even if available

 

[Dave Haynie]

You never said what is a good test to compare the cpus that you mention have an unfair shake vs apple chips. Or are you saying it’s not possible?

One of the big discrepancies between Geekbench multiprocessing and real world application multiprocessing is that Geekbench only runs one thread pre real core. So the benefit -- if there is one -- of multi threaded processors like AMDs and Intel's is not measured in the benchmark. When you get a ,ore real world benchmark like CineBench, multithreading is exploited.

There are two more interesting ways to run such a benchmark. One would allocate 2-3x as many threads as real processors. The other would be to allocated some large number, like 1000 threads with fixed work per thread, the lower number the better. The first would match typical CAD/media processing workloads, the latter would match typical internet server workloads.