How fast will Intel’s next-gen processors be? The first sight of Arrow Lake turns expectations upside down

AMD processors with the Zen 5 architecture are a bit late (that is, when it comes to the desktop version), but we already got some indications about the performance. A much bigger mystery are competing Intel Arrow Lake processors, which will be released in October and their clocks are not yet known, leaving massive question marks over their performance. Fortunately, the first preliminary benchmarks are appearing in this uncertainty, which finally indicate what we can look forward to.

Remember that this is still preliminary data, which should be taken with a grain of salt. But their originator is that leaker Jakihn, who is most likely a legitimate insider and has already made a lot of money from Intel’s plans in the last month. This week, Jaykihn published some performance benchmarks to show the current capabilities of the Core Ultra 9 285K, which should be Arrow Lake’s most powerful desktop processor with 8 Lion Cove cores and 16 efficient (E-Core) Skymont cores.

First tests of the Core Ultra 9 285K sample

These tests are allegedly performed on a QS (qualification sample) processor, which should already have the same parameters as the finished Core Ultra 9 285K. However, the QS samples are said to be available only internally at Intel so far, so it is possible that Jaykihn got the measured scores second hand. But the important thing is that the performance is probably not definitive, according to the leaker, it will probably improve even more. While the parameters (frequencies) of the CPU should probably already match, the firmware is probably not yet adjusted three months before the release, so in the final the performance could be better in some tests. Especially probably in programs that work a lot with memory, the settings of which probably don’t need to be adjusted yet.

In these benchmarks, the qualification sample (QS) is compared with the older and slower engineering sample (ES2) and for comparison with the Core i9-14900K from the current generation of Intel processors. All processors are set to 250W PL2, i.e. the same maximum consumption (take into account that until recently this was not the standard for the i9-14900K, which in reviews usually had a higher consumption and thus a higher multi-threaded score).

The first benchmark that Jaykihn shows is CrossMark by BAPCo. This test should represent overall computer performance and responsiveness, based on existing application models. In this test, Arrow Lake’s qualifying sample has 2587 points, surpassing o +6,4 % Core i9-14900K score. This is a composite test score, individual subscores are not shown.

The following are two web browser tests where, unfortunately, things turn out badly for Arrow Lake, both ES and QS being slower than the old Raptor Lake processor. In WebXPRT4 3.73 is Arrow Lake slower by -4,2 %ve Speedometer 2.1 an even greater difference emerges, -9,4 % for Arrow Lake. Both tests use Chrome v114 on all processors.

The single thread test is also infamous GeekBench 5.4.5 (ST)where the qualifying sample Arrow Lake obtained 2455 pointswhich is just o +0,9 % better than the Core i9-14900K score. The new generation should therefore have a marginal improvement in single-threaded performance, according to this. This could be because the big cores are clocked at just 5.7GHz during max boost according to the leaks (for the top-of-the-line Core Ultra 9 285K model), which is a regression against the up to 6.0GHz Raptor Lake.

But beware, Arrow Lake surprisingly snatched victory in the multi-threaded score. IN GeekBench 5.4.5 (MT) has 27,381 points which is o +14,6 % better than Core i9-14900K performance.

And similarly, Arrow Lake also boasts in Cinebench R23 (multi-threaded) where it has 43,118 points and that’s even o +17,5 % better score than Raptor Lake. It must be said that you often find a higher score of over 40,000 points in reviews for the Core i9-14900K, but this is only when it has (against Intel’s “recommendations”) unlocked consumption, while respecting the 250W limit will reduce performance quite a bit. This shows that Arrow Lake is more power efficient (which it definitely should be when it uses a 3nm process versus Raptor Lake’s 7nm technology).

Strong in multi-threaded software, weak in single-threaded?

Even so, these are surprising results. It was generally expected that Arrow Lake would not be very strong in multi-threaded performance, since it has the same number of P-Core and E-Core cores as Raptor Lake, while its large cores will lose HT and thus some of the performance in multi-threaded tasks like Cinebench. But it seems that the new efficient Skymont core overcomes this deficit playfully. It is supposed to have tens of percent better performance at 1 MHz compared to the previous E-Core and possibly even higher clocks.

Surprisingly, Arrow Lake could be strong precisely in multithreaded performance, where probably the maximum configuration with 8+16 cores will successfully face (and in programs like Cinebench R23, therefore possibly even beat) even the fastest Ryzen 9 9950X with 16 Zen 5 cores and 32 threads . On the contrary, in single-threaded performance, where Intel’s strength used to be historically and we expected it to make more progress with Arrow Lake and maybe lead over AMD again, it almost looks like a loss now. So completely against expectations.

But we will see how the performance will possibly change with the further tuning of the platform. In theory, this could probably improve those performance tests in the browser, so then even single-threaded performance could look better.

Unfortunately, it will probably take quite a while before we get a complete picture of the performance, as the Arrow Lake desktop processors will not be released until October of this year. But in two weeks or so, reviews of the Ryzen 9 9950X will be published, at least it will be possible to look at its score in Cinebench R23, which will certainly be measured in some review, and compare it with these alleged QS results of the new generation Intel sample.

Source: Jaykihn