Intel Arrow Lake-H CPU Analysis: Core Ultra 200H Makes Lunar Lake Almost Redundant

[Redaktion]

Intel is introducing its new Arrow Lake-H series mobile processors (Core Ultra 200H), covering a power range of 28 to 115 watts. We examine two implementations of the high-end Core Ultra 9 285H. But are the improvements enough to compete with AMD's Zen 5 and Qualcomm's Snapdragon X Elite?

https://www.notebookcheck.net/Intel-Arrow-Lake-H-CPU-Analysis-Core-Ultra-200H-Makes-Lunar-Lake-Almost-Redundant.959328.0.html

[bernstein]

Nobody cares about NPU perf

[Aras]

The "Performance with different power limits" section provides a really nice overview of the CPUs that is missed by many reviewers. It reveals that Ryzen AI 9 HX 370 provides the best performance at all power levels, and it's a pity that it's not easy to find laptops with the Ryzen CPU. It's not necessarily the best CPU in every case, though, because some other CPUs may consume less power when idle or during extremely light tasks, hence, provide a better battery life in that case.

[spicychile]

The "Performance with different power limits" section provides a really nice overview of the CPUs that is missed by many reviewers.

Agreed! This should be a standard for future CPU reviews, even on desktop like for comparing ECO mode vs. stock on AMD CPUs. It's especially helpful for SFF builds where it can be more difficult to keep temps under control, so knowing whether the tradeoffs of more heat for more performance is worth it is valuable.

[Don't Fear the Future]

I'm impressed by the 285H.

At 55 watts, the 285H is within 12% of the Ryzen AI 9 370HX.  That is 14 threads vs 24 Threads. 

(To my knowledge, the 2 LP-E cores on the 285H are NOT on the Compute Tile of the CPU; therefore, would not participate in much in a Cinebench Multicore run.  Hence why I stated 12 threads instead of 16 threads)

[Don't Fear the Future]

ooops, meant to say "14 threads instead of 16 threads"

[dada_dave]

I'm impressed by the 285H.

At 55 watts, the 285H is within 12% of the Ryzen AI 9 370HX.  That is 14 threads vs 24 Threads. 

(To my knowledge, the 2 LP-E cores on the 285H are NOT on the Compute Tile of the CPU; therefore, would not participate in much in a Cinebench Multicore run.  Hence why I stated 12 threads instead of 16 threads)

ooops, meant to say "14 threads instead of 16 threads"

Depends on how you count it. Hyperthreading is really only expected to give an increase of 20-30% in most, even highly parallel, workloads. Thus a 12-core part like the HX 370 will behave as though it were a 14-16 core part with no hyperthreading. I believe Tom's Hardware did do a hyperthreading on/off test for CB 24 and if I remember correctly it was around the 25% mark.

So one *could* argue that it's a "15-thread" equivalent AMD part versus a 14-16 thread Intel.

[dada_dave]

I'm impressed by the 285H.

At 55 watts, the 285H is within 12% of the Ryzen AI 9 370HX.  That is 14 threads vs 24 Threads. 

(To my knowledge, the 2 LP-E cores on the 285H are NOT on the Compute Tile of the CPU; therefore, would not participate in much in a Cinebench Multicore run.  Hence why I stated 12 threads instead of 16 threads)

ooops, meant to say "14 threads instead of 16 threads"

Depends on how you count it. Hyperthreading is really only expected to give an increase of 20-30% in most, even highly parallel, workloads. Thus a 12-core part like the HX 370 will behave as though it were a 14-16 core part with no hyperthreading. I believe Tom's Hardware did do a hyperthreading on/off test for CB 24 and if I remember correctly it was around the 25% mark.

So one *could* argue that it's a "15-thread" equivalent AMD part versus a 14-16 thread Intel.

Of course then there is the additional complication of "c"-cores vs "E"- & "LP"- cores + silicon die area vs fabrication processes. Basically, normalizing MT performance in heterogeneous processors with and without hyperthreading is ... difficult. :)

[Don't Fear the Future]

"Depends on how you count it. Hyperthreading is really only expected to give an increase of 20-30% in most, even highly parallel, workloads. Thus a 12-core part like the HX 370 will behave as though it were a 14-16 core part with no hyperthreading. I believe Tom's Hardware did do a hyperthreading on/off test for CB 24 and if I remember correctly it was around the 25% mark."

Hahaha, yes, this is exactly how I remember reading that too.


But while I too thought of that analogy, it's more of hyperthreading can increase the performance of a Core by 25%; which isn't exactly the same as increasing the amount of cores by 25% (though it does make sense)

So, lets take the HX 370 score of 1167 at 55 watts.  Without hyperthreading the score may very well be 875. (1167 x.75); as all of the cores of the HX 370 have hyperthreading.

So, with hyperthreading turned off, the HX 370 with 12 cores could get a score of 875, versus the score of the 285H with 14 cores of 1042. (like I said above, the 2 lp-e cores are not on the compute tile)

For me, the AMD is going to be much better where hyperthreading is truly utilized; which is basically used in every synthetic benchmark used today.

But in instances where hyperthreading doesn't work, the Intel is going to shine pretty good.

But this just my 2 cents.

[Don't Fear the Future]

"So, lets take the HX 370 score of 1167 at 55 watts.  Without hyperthreading the score may very well be 875. (1167 x.75); as all of the cores of the HX 370 have hyperthreading."

ooops again.  I did the math wrong.  I would be 934 (1167/1.25), but I think I got the idea across.

[dada_dave]

"Depends on how you count it. Hyperthreading is really only expected to give an increase of 20-30% in most, even highly parallel, workloads. Thus a 12-core part like the HX 370 will behave as though it were a 14-16 core part with no hyperthreading. I believe Tom's Hardware did do a hyperthreading on/off test for CB 24 and if I remember correctly it was around the 25% mark."

Hahaha, yes, this is exactly how I remember reading that too.


But while I too thought of that analogy, it's more of hyperthreading can increase the performance of a Core by 25%; which isn't exactly the same as increasing the amount of cores by 25% (though it does make sense)

So, lets take the HX 370 score of 1167 at 55 watts.  Without hyperthreading the score may very well be 875. (1167 x.75); as all of the cores of the HX 370 have hyperthreading.

So, with hyperthreading turned off, the HX 370 with 12 cores could get a score of 875, versus the score of the 285H with 14 cores of 1042. (like I said above, the 2 lp-e cores are not on the compute tile)

For me, the AMD is going to be much better where hyperthreading is truly utilized; which is basically used in every synthetic benchmark used today.

But in instances where hyperthreading doesn't work, the Intel is going to shine pretty good.

But this just my 2 cents.

"So, lets take the HX 370 score of 1167 at 55 watts.  Without hyperthreading the score may very well be 875. (1167 x.75); as all of the cores of the HX 370 have hyperthreading."

ooops again.  I did the math wrong.  I would be 934 (1167/1.25), but I think I got the idea across.

Aye as I said in my second post normalizing MT scores for processor size across processors with very different architectures is pretty difficult. When it is extreme, it is obvious, but it's not immediately clear how 4 Zen 5 + 8 Zen 5c cores with "24" threads should be compared to 6 Lion Cove + 8 Skymont + 2 LP Skymont cores. If available, sometimes I like to compare CPU die areas but then you still have to account for differences in fabrication. So yeah ... non-trivial and agreed, it can be very test specific. Even for this test, it's power level specific.

As power level drops, performance drops faster for the MSI Arrow Lake than for the Zen. Unfortunately, unlike in some previous analyses articles they didn't also measure wall power as they changed power levels. But assuming they AMD and Intel machines are drawing roughly the same power when set to the same TDP, at 20-28W the AMD processor is indeed ahead by 19-28% (in this one test). At higher power, as you noted, Intel closes the gap. 

I have to admit I am a little confused by this power-performance relationship in the Intel chip though. This is especially apparent in the 285H Zenbook MT results compared to the MSI where we have wall power measurements. The Zenbook's MT clocks are obviously down clocked but it didn't get any efficiency increase - i.e. MT performance fell linearly with power which shouldn't happen. In all the other devices with the same processor at different power levels, efficiency goes up as power/performance goes down as power should fall faster than performance when one lowers clocks. And yet it didn't for the Zenbook (which otherwise had the same/better single threaded efficiency than the MSI machine so it isn't like the Zenbook is just bad). Just comparing the TDP/performance chart, we do also see a much bigger performance drop off for MSI than the HX 370 as power is lowered (though again no wall power measurements), so maybe that's just the way Arrow Lake H performs at low power. But, it seems ... odd.

========

Small typo in the article: the Arrow Lake H GPU tile is manufactured on 5NP not 5NB.

[dada_dave]

But assuming they AMD and Intel machines are drawing roughly the same power when set to the same TDP, at 20-28W the AMD processor is indeed ahead by 19-28% (in this one test). At higher power, as you noted, Intel closes the gap. 

...


Just comparing the TDP/performance chart, we do also see a much bigger performance drop off for MSI than the HX 370 as power is lowered (though again no wall power measurements), so maybe that's just the way Arrow Lake H performs at low power. But, it seems ... odd.
 

Pretty sure the AMD chips are drawing more wall power at the same TDP level. The HX 370 results come from their HX 370 analysis article where they indeed measured wall power at each TDP and at 35W getting 1022 points the HX370 actually drew 57W. Meanwhile the MSI drew 52W to get 991 which is more points than it got at its 45W TDP level. Now there is a difference in idle power - probably due to the fact that the AMD device has a dGPU - but even so that means the TDP chart may be doing the Intel chips a disservice. However, the Arrow Lake H Zenbook's performance is still odd and its power was measured at the wall. So it is still possible that Arrow Lake H's performance drops off faster for some reason at lower power than the Strix Point's. However, we can state that draw roughly 50W of wall power accounting for differences in device power Arrow Lake H and Strix Point are almost identical in terms of performance and efficiency where Strix Point's hyperthreading lets it make up the difference in lower core count.

Again, all in Cinebench R24. As you say, other tests may show different results, some benefiting one or the other. But here, they're surprisingly equivalent at this power level for this test. 

[PHVM_BR]

To my knowledge, the 2 LP-E cores on the 285H are NOT on the Compute Tile of the CPU; therefore, would not participate in much in a Cinebench Multicore run.

I confirm this.

On the Ultra 7 155h the 2 LP-E don't work when running Cinebench R23 or 2024 and the CPU usage only reports 91% when running these benchmarks.

[Don't Fear the Future]

[
[/quote]

I confirm this.

On the Ultra 7 155h the 2 LP-E don't work when running Cinebench R23 or 2024 and the CPU usage only reports 91% when running these benchmarks.
[/quote]

Thanks for the verification.