Intel "Arrow Lake-S" Engineering Sample Posts Over 25% 1T Perf Gain Over i9-13900K, Falls Behind in nT

[R0H1T]

This is pure Intel vs Intel test. Don't bother comparing this to Zen 5 or any AMD architecture as CPU-Z has always strongly favored Intel.

Ok but even then what's the baseline? Intel's new predefined "defaults" or day one from reviews 2 years back? Also this is just one score & probably one of the better ones!

 

[Darmok N Jalad]

This is pure Intel vs Intel test. Don't bother comparing this to Zen 5 or any AMD architecture as CPU-Z has always strongly favored Intel.
For example Zen 5 vs Zen 4 is supposed to be 19% faster but still at ~910 points: https://www.techpowerup.com/322906/...formance-at-5-80-ghz-found-19-over-zen-4?cp=2

Also other benchmarks show Zen 5 being 55% faster compared to 13900K:

AMD's new Zen 5 chip up to 55% faster than Intel's Core i9-13900K in leaked benchmark — AMD's Ryzen 9 9950X purportedly shines in AVX workloads

Ryzen 9000 'Granite Ridge' processors are scheduled to launch in July.

www.tomshardware.com

The truth is that while it's fun to speculate we wont know the full performance breakdown until end of the year. Perhaps not even until Q2 2025 if Zen 5 X3D launches that late.
There are too many variables in these leaks like ES, non-final clocks, unknown cooling, unknown memory configuration that can all affect the accuracy of the results.
If one source leaked both Zen 5 and Arrow Lake-S benches with same cooling and same RAM/Speed then it would be more representative but this rarely happens that someone gets a hold of both ES near final or even qualification chips and knows how to test properly.

Yes, and we can't forget the history of CPUz, where once upon a time, an alleged anomaly cause the original Ryzen to outperform its Core counterparts. Perhaps this is the case again, where the architecture is enough of a departure to outpace the benchmark.

 

[N/A]

If it is a 6c12t CPU that is pretty good.

This is a 14c14t in a 6P+8E tile configuration on Intel 20A node. The 8P+16E compute tile is on TSMC 30A so it will have different power characteristics and less overclock. 6P is the way to go for games.

 

[ratirt]

honestly im not counting the Ecores for single thread but rather multithread. 6p for games is the way to go for games I'm not sure. I guess it depends on games.
If your argument would be, ecores are great for gaming, I wonder why Intel did not use only ecores. I guess it is obvious why but i digress.

 

[N/A]

6p will be on a more advanced node and reach 10-15% higher clock. Now that Hyperthreading is gone, these aren't just e-Cores, they serve as rental units which somehow better. I still can't figure it out.

 

[Tomorrow]

honestly im not counting the Ecores for single thread but rather multithread. 6p for games is the way to go for games I'm not sure. I guess it depends on games.
If your argument would be, ecores are great for gaming, I wonder why Intel did not use only ecores. I guess it is obvious why but i digress.

I wondered the same. If E-C cores are as performant as current gen P-Cores then why not double down and make a pure E-Core (client) CPU or reduce the number of P-Cores?

Im guessing they're not as performant as Intel is claiming.

 

[pressing on]

Hmm, trying to figure how many P core/ E core cpu this is. Going by the CPU-z chart it looks like a 6 P core/ 8 E core CPU which if it is, it's pretty impressive!

Yes, wccftech is saying that it is being reported as "...a 6+8 SKU which should offer 6 P-Cores and 8 E-Cores.". But Lunar Lake will be out in September so that will be the first solid evidence of what Lion Cove and Skymont performance actually is.

 

[Jaskara]

I wondered the same. If E-C cores are as performant as current gen P-Cores then why not double down and make a pure E-Core (client) CPU or reduce the number of P-Cores?

Im guessing they're not as performant as Intel is claiming.

Intel’s claim is that they match Raptor Cove’s IPC, not it’s 1T performance. It would make sense that Skymont could still match the IPC, but be at a substantial 1T deficit if it only clocks to, say, 4.2-4.5ghz, whereas raptor cove went up to 6.2.

It’s still true that Intel’s claim could be rosy, but there’s nothing to prove it right or wrong just yet. That said, if Intel is telling the truth, it would be pretty interesting to see what the E-Cores could do if Intel really tried to open up their design, considering how much more power/area efficient they are over P core today.

 

[Tomorrow]

Got any TPU source link?

TSMC N3P actually. Not sure what the hell is 30A as it's the first time im hearing this. TSMC wont start using A monikers until below 2nm with 16A that should be 1,6nm or something like that.

 

[P4-630]

TSMC N3P actually. Not sure what the hell is 30A as it's the first time im hearing this. TSMC wont start using A monikers until below 2nm with 16A that should be 1,6nm or something like that.

Arrow Lake will be on Intel 20A (TSMC 2nm)

 

[N/A]

Some arrows, below ultra 5. Ultra 7 and 9 are rumored to be outsourced to tsmc and it makes sense. 8 cores plus 16 fits exactly in the same area as the lunar lake compute tile. Provided the soc and gpu are separate tiles. And ultra 5 could be entirely monolithic or two tiles max.

 

[Tomorrow]

Arrow Lake will be on Intel 20A (TSMC 2nm)

20A is more like TSMC 3nm at best. Certainly not 2nm. Also since this will be Intel's first desktop chiplet CPU (Meteor Lake-S was canned) some tiles will be made by TSMC. Yes the CPU tile housing cores will be made by Intel but the GPU tile will be made by TSMC. Not sure about SoC and IO tiles.

 

[Redwoodz]

liquid helium

 

[TF-GrayWizard]

Sadly its a CPU-Z bug

Reproducible by running the version 19 AVX beta test then switch it back to the version 17 test but don't run it then select a CPU to compare to and the ver 19 results show up.

 

[Steevo]

What a time to be alive, we finally have enough cores to swap out the extra die space for more cache, a fatter prediction unit and reap the rewards. I hope AMD is paying attention and drops a 12 core monolithic X3D without HT for gaming so Intel doesn't rape us all to death.

 

[InVasMani]

If this is the i5 it's actually about 30% faster in MT than the old i5.

It is likely clocked to it's eyeballs but still -- shows great potential.

Yeah, but everything is clocked to it's eyeballs now. It's not exactly just Intel it's close to same with AMD and Nvidia. Overclocking is mostly dead now.

 

[phanbuey]

Yeah, but everything is clocked to it's eyeballs now. It's not exactly just Intel it's close to same with AMD and Nvidia. Overclocking is mostly dead now.

first gen i5's are usually the exception for product segmentation. Usually just a cut down i9 with cores disabled and artificially supressed to low clocks.

12600k oc'd to 20-30% perfomance increase. Im still kickin myself for upgrading to 13700k -- I had a 12600k sitting at 5.3 Ghz with tuned ram and ring. Went to 13700k for nice 5 C increase to my ambient room temp.

13600k/14600K can all hit clocks of 13900/14900k... I'm thinking intel will follow the pattern here. i9 Juiced to the eyeballs, i5 on 20A experimental, but clocked lower to sell i7/i9's -- and with loads of headroom.

 

[Crackong]

Maybe not relevant, but I remember there is a thing called "SMT1"
Could this be the reason why 2way HT isn't available in the new CPUs?

 

[Minus Infinity]

Yes, wccftech is saying that it is being reported as "...a 6+8 SKU which should offer 6 P-Cores and 8 E-Cores.". But Lunar Lake will be out in September so that will be the first solid evidence of what Lion Cove and Skymont performance actually is.

One leaked benchmark of LL with only 4+4 was beating Mediocre lake 6+8 in single thread using only 17W and only 11% behind in multithread. Pretty impressive if actually true.

I'm eagerly awaiting Arrow Lake vs Zen 5/5X3D for productivity and gaming.

20A is more like TSMC 3nm at best. Certainly not 2nm. Also since this will be Intel's first desktop chiplet CPU (Meteor Lake-S was canned) some tiles will be made by TSMC. Yes the CPU tile housing cores will be made by Intel but the GPU tile will be made by TSMC. Not sure about SoC and IO tiles.

20A uses GaaFET, BSPD and is way ahead of TSMC in that regard. We might not see TSMC move to that tech until N2 or A18. Intel is moving to 18A rather quickly for Clearwater Forest and Panther Lake.

 

[watzupken]

Better wait for official and independent review results. I have no doubts that next gen processors should see a good bump in performance. In the case of Intel, the bump in performance and efficiency seems to come at the cost of multi-threaded performance, and not an overall improvement. Hence, depending on your workflow, the lost in multi-threaded performance may not make it attractive. And I still don't like the idea of having P and E cores with significant difference in performance since it is very dependent on a software layer to assign the right cores to the tasks.

20A uses GaaFET, BSPD and is way ahead of TSMC in that regard. We might not see TSMC move to that tech until N2 or A18. Intel is moving to 18A rather quickly for Clearwater Forest and Panther Lake.

In my opinion, I won't even bother with these numbers and technology. What does 20A or 3nm even mean when we know that the underlying is not really 3nm or 20A? Gate all around sounds great one paper, but I am not sure if we will see any significant improvement since it is in a nascent stage. All these numbers just gives us some indication on the "newness" of the fab, but ultimately as consumers, we only want to see the below that will drive our purchase decision,
1. How performant is the chip
2. How power efficient.

Just as you, I am keen to see how next gen processors stack up against one another.

 

[Minus Infinity]

Better wait for official and independent review results. I have no doubts that next gen processors should see a good bump in performance. In the case of Intel, the bump in performance and efficiency seems to come at the cost of multi-threaded performance, and not an overall improvement. Hence, depending on your workflow, the lost in multi-threaded performance may not make it attractive. And I still don't like the idea of having P and E cores with significant difference in performance since it is very dependent on a software layer to assign the right cores to the tasks.


In my opinion, I won't even bother with these numbers and technology. What does 20A or 3nm even mean when we know that the underlying is not really 3nm or 20A? Gate all around sounds great one paper, but I am not sure if we will see any significant improvement since it is in a nascent stage. All these numbers just gives us some indication on the "newness" of the fab, but ultimately as consumers, we only want to see the below that will drive our purchase decision,
1. How performant is the chip
2. How power efficient.

Just as you, I am keen to see how next gen processors stack up against one another.

Just consider them labels not feature size. Got to call them something. GaaFET and BSPD are a big deal, we may see some impressive improvements in power efficiency.

 

[N/A]

Better wait for official and independent review results. I have no doubts that next gen processors should see a good bump in performance. In the case of Intel, the bump in performance and efficiency seems to come at the cost of multi-threaded performance, and not an overall improvement.

6 integer ALUs compared to 5 in alder lake and many more improvements. But that to me is more telling as it translates to more instructions per cycle.

 

[Tomorrow]

20A uses GaaFET, BSPD and is way ahead of TSMC in that regard. We might not see TSMC move to that tech until N2 or A18. Intel is moving to 18A rather quickly for Clearwater Forest and Panther Lake.

GaaFET is not a silver bullet. Samsung has it too before TSMC but with abysmal yields (reported). We dont know Intel's yields.
BSPD is a good step but until we see it in action we dont know if the promises hold true.

Also im scared to think how expensive these chips will be compared to TSMC N4P.
Also if Intel is moving to 18A so quickly then that does not instill a lot of confidence in 20A.

 

[MrMilli]

Arrow Lake-S will have HT. It's only Arrow Lake-U that's confirmed not to have HT.

 

[stimpy88]

Now we know why AMD is panicking. Time AMD took a small risk and made the extra L3 cache part of the die, instead of a cash grab band-aid to compete with Intel.