News Intel's next-gen desktop CPUs have leaked — Arrow Lake Core Ultra 200 series share similar core counts with Raptor Lake Refresh

[Admin]

Chinese news outlet Benchlife shares the alleged specifications for Intel's forthcoming Core Ultra 200 series (Arrow Lake) processors.

Intel's next-gen desktop CPUs have leaked — Arrow Lake Core Ultra 200 series share similar core counts with Raptor Lake Refresh : Read more

 

[Metal Messiah.]

It appears Intel did change the nomenclature, as the site claims. Because the numbers ending in "90", "70" and "60" series, would be replaced by "85", "65" and "45" instead. And non-K SKUs have slightly different Model numbers as well.

So the full "rumored" lineup:

• Core Ultra 9 285K (Intel Core i9-14900K Successor)
• Core Ultra 7 265K (Intel Core i7-14700K Successor)
• Core Ultra 5 245K (Intel Core i5-14600K Successor)

Add these non-K parts in the table if you wish:

• Core Ultra 9 275
• Core Ultra 7 255
• Core Ultra 5 240

Most likely the last SKU is the Core Ultra 5 10c/10t (6+4) processor, 240/240F. This processor was recently spotted in a coreboot patch.

Total 13 processors as the site claims.

3 K
3 non-K
2 F
5 T, 35 Watts series.

Speculation:

Either Intel might opt for a bigger "8+16" die configuration/H0 for its top Core Ultra 9 & Ultra 7 SKUs, and use the 6+8 die/C0 for its Core Ultra 5 SKUs.

Or, another possibility is that all three "K" SKUs would be featuring the 8+16 die config, while the non-K SKUs could end up using the 6+8 die instead.

 

[peachpuff]

Is ht gone due to those mitm attacks?

 

[TerryLaze]

I very much doubt that it won't have HTT it would make them be slower, or in the best case same speed if the new rentable units thing can make up for the loss, than the previous gen which would be very bad for sales.

But then again intel had good sales in the past with minimal to no change in performance so who knows.

 

[Metal Messiah.]

Let me point this again. Copy/paste of my previous comment.

Arrow Lake might not come with HyperThreading

Speaking of HT, I'm still wondering of Intel's patent which focuses on "Rentable Units" which they filed back last year. The patent calls it as “Instruction Processing Circuit“.

News went silent on this one though, so I'm not sure what to say.

https://www.freepatentsonline.com/y2023/0168898.html

IDK, if you go through the patent carefully, some details are mentioned. Here in the figure below, Intel has highlighted the difference between hyper-threading and Rentable Units on hybrid-core processors, assuming this is still in Intel's consideration for future client CPUs.

The densely shaded areas are P-cores while the lightly shaded ones represent the E-cores.

The Rentable Unit splits the first thread of incoming instructions into two partitions, assigning two different cores to each based on the complexity. There are two threads (1 and 2). The scheduler divides each into three partitions (A, B, C).

The first two partitions of 1 (1A, 1B) are executed on the P-core, while the third (1C) is handed to the E-core. Likewise, the first partition of thread 2 (2A) is processed by the E-core, while the other two (2B, 2C) are executed by the P-core.

So it would appear that the Renting Unit is passing part of the work of the E-Core to the P-Core so that the second one does not find a part of the time stopped.

The patent also mentions that "Rentable Units" will use timers and counters to measure P/E core utilization and send parts of the thread to each core for processing. This inherently requires larger cache sizes, and Arrow Lake is already rumored to have 3 MB of L2 cache per core.

So basically, this appears to be a pseudo-multi-threaded solution that splits the first thread of incoming instructions into two partitions, assigning them to different cores based on complexity more like.

 

[TerryLaze]

So it would appear that the Renting Unit is passing part of the work of the E-Core to the P-Core so that the second one does not find a part of the time stopped.

The patent also mentions that "Rentable Units" will use timers and counters to measure P/E core utilization and send parts of the thread to each core for processing. This inherently requires larger cache sizes, and Arrow Lake is already rumored to have 3 MB of L2 cache per core.

So basically, this appears to be a pseudo-multi-threaded solution that splits the first thread of incoming instructions into two partitions, assigning them to different cores based on complexity more like.

They will split every thread into complicated and not-so-complicated parts, this first thread thing in the description is just to make it easier to follow the explanation, and then cram them into as many p and e cores as they can to get the fastest result.
It's a great concept for general computing and should provide a good boost to the user experience.

The problem for me is that reviews just run benchmarking programs and those are only complicated, from beginning to end, so I don't think that it will make enough of a difference since all cores e and p are already fully, or almost fully, used by those, I very much doubt that it will be able to make up the 20-30% that HTT provides to cinebench or similar.

 

[usertests]

The problem for me is that reviews just run benchmarking programs and those are only complicated, from beginning to end, so I don't think that it will make enough of a difference since all cores e and p are already fully, or almost fully, used by those, I very much doubt that it will be able to make up the 20-30% that HTT provides to cinebench or similar.

Background: MLID has been saying HT is dropped in Arrow Lake for a long time. He also leaked "rentable units" but has said they're not ready. Tom's Hardware has been running with the no HT narrative since at least January.

A decent IPC increase could counteract the loss of HT in the great benchmarking game. We've heard as low as 10%, but maybe it will be better. A multi-threaded performance regression is bad but not the end of the world. Most users wouldn't notice. Though we have to see how games will handle 8 P-core threads instead of 16.

If Arrow Lake increases power efficiency significantly (avoiding the current instability fiasco), has a better than 10% IPC increase and maybe a small clock increase, and 50% larger L2 cache as rumored, then it could end up looking decent. Also, the iGPU may be significantly improved from UHD 770, good for when office PCs hit the used market in 5+ years.

An Arrow Lake Refresh could add another 16 E-cores at the top, taking it to 8+32 and ensuring it has better multi-threading chart-topping potential than the 14900KS.

 

[Amdlova]

There's no reason at all to upgrade from a raptor lake to a new cpu from any brand.
You can skip the dd5 generation and goes directly to ddr6 and pci 6 gen

 

[Eximo]

There's no reason at all to upgrade from a raptor lake to a new cpu from any brand.
You can skip the dd5 generation and goes directly to ddr6 and pci 6 gen

I'm more curious if they are going to stick to DIMMs or consider CAMM2 or future versions for desktop. Getting to that point where the trace length and signal integrity are going to be very tricky to reach higher speeds.

Or will CPUs just keep piling on the cache to alleviate memory bandwidth problems.

 

[TerryLaze]

Background: MLID has been saying HT is dropped in Arrow Lake for a long time. He also leaked "rentable units" but has said they're not ready. Tom's Hardware has been running with the no HT narrative since at least January.

A decent IPC increase could counteract the loss of HT in the great benchmarking game. We've heard as low as 10%, but maybe it will be better. A multi-threaded performance regression is bad but not the end of the world. Most users wouldn't notice. Though we have to see how games will handle 8 P-core threads instead of 16.

If Arrow Lake increases power efficiency significantly (avoiding the current instability fiasco), has a better than 10% IPC increase and maybe a small clock increase, and 50% larger L2 cache as rumored, then it could end up looking decent. Also, the iGPU may be significantly improved from UHD 770, good for when office PCs hit the used market in 5+ years.

An Arrow Lake Refresh could add another 16 E-cores at the top, taking it to 8+32 and ensuring it has better multi-threading chart-topping potential than the 14900KS.

IPC might have counteracted the loss if the benchmarks would run the same amount of threads but now it has to compensate for the missing HTT and fewer threads so less work send to the CPU in the first place and less for the scheduler and rentable units to choose from.

Games are one of the things were this might show a benefit, at least for games that still have a main thread, in those cases simpler parts of the code could be send to the e-cores allowing the p-core(s) to run more of the complex parts in the same time increasing overall performance. But those games only become fewer and fewer.

 

[alceryes]

It appears Intel did change the nomenclature, as the site claims. Because the numbers ending in "90", "70" and "60" series, would be replaced by "85", "65" and "45" instead. And non-K SKUs have slightly different Model numbers as well.

So the full "rumored" lineup:

• Core Ultra 9 285K (Intel Core i9-14900K Successor)
• Core Ultra 7 265K (Intel Core i7-14700K Successor)
• Core Ultra 5 245K (Intel Core i5-14600K Successor)

Add these non-K parts in the table if you wish:

• Core Ultra 9 275
• Core Ultra 7 255
• Core Ultra 5 240

Most likely the last SKU is the Core Ultra 5 10c/10t (6+4) processor, 240/240F. This processor was recently spotted in a coreboot patch.

Total 13 processors as the site claims.

3 K
3 non-K
2 F
5 T, 35 Watts series.

Speculation:

Either Intel might opt for a bigger "8+16" die configuration/H0 for its top Core Ultra 9 & Ultra 7 SKUs, and use the 6+8 die/C0 for its Core Ultra 5 SKUs.

Or, another possibility is that all three "K" SKUs would be featuring the 8+16 die config, while the non-K SKUs could end up using the 6+8 die instead.

Good info. Thanks.
They're getting rid of HT and supposedly using the rentable units. I'm not surprised.
I wonder if anyone has done performance testing on HT cores vs E-cores? I still think E-cores would win. HT cores are just pre-computation cache fillers. HT adds heat and power, and security vulnerabilities. Intel now thinks they can use that same heat and power elsewhere and gain a net positive in performance.

Personally, I have SMT off on my 7950X. I have no need for more than 16 cores and would rather keep my CPU cooler and able to clock higher in games (giving better performance vs having SMT enabled).

The one potential issue I see coming up is that, the more Intel fiddles with the execution order and core assignments for work, the more resources will be used for just this ordering of work. It's like they're making a super-HT that might allow for performance gains now but may become more of a mess down the road. Time will tell.

 

[TerryLaze]

I wonder if anyone has done performance testing on HT cores vs E-cores? I still think E-cores would win.

Doesn't matter if the e-cores win because HTT is still additional performance on top of that, you still get 20-30% (in bench apps, more in simpler apps) more performance out of every p-core you have compared to not having HTT on them.

 

[alceryes]

Doesn't matter if the e-cores win because HTT is still additional performance on top of that, you still get 20-30% (in bench apps, more in simpler apps) more performance out of every p-core you have compared to not having HTT on them.

This is untrue. The performance gain/loss is workload specific. The lower clocks when using HT (due to both heat and power profile limits) sometimes negate the performance gain that hyper threading provides. Hyperthreading has never been a free X% performance boost.

A simple test for you and others to do (if you have at least 8 P-cores) is to run 3DMark's Time Spy both with HT enabled and disabled. Results will vary depending on arcitecture, cooling, etc. but many people (myself included) see a performance increase with HT disabled due to the regular cores being able to clock higher, due to the aforementioned more power and less heat. It is actually a common performance tweaker move to turn off SMT/HT for gaming as no games are optimized for more than 8 cores, yet.

Another way to look at it - Intel would not be getting rid of hyperthreading if it just gave a free X% performance boost. They've no doubt thrown tens (maybe hundreds) of millions of dollars into researching the best ways to increase CPU performance. Axing hyperthreading is the first step.

 

[Eximo]

Yeah, it does seem to go in favor of P cores at the moment with Alder/Raptor when it comes to power consumption. Even though one P core is roughly the same die area as 4 E cores.

From the images of Arrow Lake out there, very similar size ratio between P and E cores again, so they didn't save themselves much by ditching hyperthreading. Though the actual look of the silicon seems a bit different, that could be the area recovered by not having HT, which could also be an improvement.

https://i.pcmag.com/imagery/articles/06CxDSYvSdzmSJIrjCuUdiy-2.fit_lim.size_768x.png

https://en.wikichip.org/wiki/File:intel_raptor_lake_die_(8+16).jpg

 

[TerryLaze]

This is untrue. The performance gain/loss is workload specific.

That's what I said, benches are going to have less improvement than other apps

The lower clocks when using HT (due to both heat and power profile limits) sometimes negate the performance gain that hyper threading provides. Hyperthreading has never been a free X% performance boost.

You only get lower clocks with HTT if you are heavily overclocking your CPU...
At default clocks HTT has no effect on clocks at all, it will create more heat and use more power but it will also increase performance by the same amount.

Another way to look at it - Intel would not be getting rid of hyperthreading if it just gave a free X% performance boost.

Agreed, that's why said that I don't believe that they will in the first place.

 

[usertests]

I'm more curious if they are going to stick to DIMMs or consider CAMM2 or future versions for desktop. Getting to that point where the trace length and signal integrity are going to be very tricky to reach higher speeds.

Or will CPUs just keep piling on the cache to alleviate memory bandwidth problems.

There are desktop motherboards with SO-DIMMs on them. I think DIMMs and CAMM will co-exist for at least the confirmed DDR6 generation. After DDR6, who knows?

Micron's new 64 GB LPCAMM2 is pretty small. Maybe it will be adopted in Mini-ITX and mini PCs. There is a possibility of stacking two of them but I don't know if that's finalized yet.

From the images of Arrow Lake out there, very similar size ratio between P and E cores again, so they didn't save themselves much by ditching hyperthreading. Though the actual look of the silicon seems a bit different, that could be the area recovered by not having HT, which could also be an improvement.

Rather than for the purpose of saving space, it may be that the groundwork for rentable units necessitated killing hyperthreading in the new microarchitecture. But then rentable units isn't ready supposedly.

We may never see thread counts in the marketing again. Keeps it simple. "24 cores". "40 cores". Do not inquire within.

 

[Eximo]

Well, the logic chip is pretty tiny anyway. Silicon space should directly equate to power required. Because the P cores clock higher and are like 90% the size of 4 E Cores and perform a bit better with hyperthreading, there is potential performance loss on the P cores without hyperthreading. Actual space savings only matter to the bean counters at Intel.

Marketing was already doing that for them. Constantly seeing things like 14 core CPU with no explanation of why that is. Mostly because a lot of the retailers didn't update their information blocks to contain any subtlety.

 

[alceryes]

That's what I said, benches are going to have less improvement than other apps

You only get lower clocks with HTT if you are heavily overclocking your CPU...
At default clocks HTT has no effect on clocks at all, it will create more heat and use more power but it will also increase performance by the same amount.

Agreed, that's why said that I don't believe that they will in the first place.

Hyperthreading is, in fact, a detriment to performance is some scenarios.

In addition to heat, lower clocks also comes from maxing out your power profile. If you hit your PPT/TDP max, your CPU's frequency will react (downclock) accordingly. Hyperthreading takes power. Power that Intel believes can be better utilized by clocking a lower number of cores to a higher frequency, and using more advanced scheduling to send jobs to E-cores.

Also, important to note that I do NOT think hyperthreading is going away in the commercial/industrial segment, at least not anytime soon. The types of workloads those Xeon-level CPUs perform are practically made for symmetric multithreading. HT still has a benefit there.

I'm curious to see what your results are in the Time Spy test I mentioned. If you run it, please list your results and full system specs.

 

[TerryLaze]

Hyperthreading is, in fact, a detriment to performance is some scenarios.

In addition to heat, lower clocks also comes from maxing out your power profile. If you hit your PPT/TDP max, your CPU's frequency will react (downclock) accordingly. Hyperthreading takes power. Power that Intel believes can be better utilized by clocking a lower number of cores to a higher frequency, and using more advanced scheduling to send jobs to E-cores.

HTT lowers the maximum possible overclock, that's different from what you are saying.
The whole fiasco with intel baseline is about the maximum overclock NOT being the normal clock but in fact being overclock.

 

[TerryLaze]

I'm curious to see what your results are in the Time Spy test I mentioned. If you run it, please list your results and full system specs.

That's useless, you (anybody) can make the results be anything you want to if you want to just by tweaking affinity priority or whatnot.

 

[alceryes]

HTT lowers the maximum possible overclock, that's different from what you are saying.
The whole fiasco with intel baseline is about the maximum overclock NOT being the normal clock but in fact being overclock.

I am not making any reference to the current Intel baseline settings fiasco.

FACTS - Hyperthreading (any symmetric multithreading, in fact) increases power usage and heat output. CPUs will throttle once they hit their TDP/PPT limit. CPUs will hit their TDP/PPT limit sooner and more frequently due to having hyperthreading enabled.

As far as YOU running Time Spy being useless, that is up to you to decide.

Since we're not getting anywhere with this discussion we'll just have to agree to disagree.

 

[PCWarrior]

On the discussion about no Hyperthreading . Sure, in applications like Cinebench hyperthreading (HT) offers 25-30% improvement in the multithreaded (MT) performance of the P-cores. But hyperthreading doesn’t offer a 25-30% improvement in the MT performance of a hybrid 14th gen chip as the e-cores never had HT and the 14900K has 16 of them. In particular for the 14900K the Cinebench score with HT enabled is around 40K and with HT disabled the score falls to around 35K. So enabling HT improves the performance of the chip by around 14%. This is something that can be compensated by improving the per core performance (during multithreaded workloads) of both the P-cores and E-cores by 14% through a combination of better IPC and higher all-core clockspeeds.

The E-cores used in Meteor Lake use the Crestmont architecture, the successor of the Gracemont architecture used in Raptor lake. Gracemont to Crestmont sees a 3% IPC uplift. For Arrow lake we will have Skymont the successor to Crestmont. If it has another 5% IPC gain, that is 8% (=1.03*1.05=1.0815) better IPC for the e-cores (Arrow Vs Raptor). If they can also get the e-cores to boost to 4.7GHz (from 4.4GHz they are now in the 14900K) that is another 7%. So overall you get roughly 15% improvement.

The P-cores with the Lion Cove architecture might as well offer a 15% IPC improvement over Raptor Cove, while the all-core clock-speed stays at least the same. It should be noted that normally for a 14900K when only the P-cores are loaded (or active because e.g. you disabled the E-cores) the all-P-core clockspeed is 5.7GHz. But once you have all the e-cores active and loaded the all- P-cores' clockspeed drops to 5.1GHz. So it is possible for Intel to get even more MT performance out of Arrow lake by running at a higher all-P-core speed especially as the P-cores will have no hyperthreading and they are manufactured on a more efficient node.

 

[Metal Messiah.]

So, I went to linked Chinese site (https://benchlife.info/arrow-lake-s-intel-core-ultra-2-naming-will-change), copy/pasted article into Perplexity, and this was in the first translated sentence,

"According to @OneRaichu, cited and reported by VideocardZ..."

Yeah. I too noticed that at first. It would appear to me that "Benchlife" wanted to credit the Twitter user OneRaichu instead of Videocardz here though. Or maybe they first saw videocardz reporting on one of his previous leaks pertaining to Arrow Lake, and thus quoted them as a reference ?

IDK. But in either case it's a total mess ! Videocardz should never have been the "original" source in this case though.

 

[Metal Messiah.]

Seeing the non-K "i5" being nerfed was surprising to me. I can't see Intel killing off the midrange in this manner. I expect more SKUs to complete this line-up.

I don't think Intel is nerfing or killing the Core i5 or the mid-range parts in general.

The Core Ultra 5 SKUs should come in both K and non-K variants, and some models will obviously be F variants lacking integrated graphics. The same as the current Raptor Lake-S/H lineup.

But do make a note that, coming Arrow Lake we won't be seeing any Core i3 or Core Ultra 3 SKUs for sure. This has been confirmed by numerous sources. So Intel might cut-down some of the Ultra 5 chips to make room for F "igpu-less" variants if need be, sporting lesser cores as well.

 

[ingtar33]

They're killing hyperthreading because it's a free 15% ipc improvement thanks to all the specter and other exploit fixes over the years. killing off hyperthreading allows them to drop the performance killing exploit protections.