All language subtitles for 酷睿Ultra全面评测：变化很大，但疗效如何？-en

The 12th generation Core two years ago left a deep impression on us.

The large and small cores introduced for the first time on the x86 platform

, the greatly improved IPC and the excellent video codec

all opened up a considerable generation gap between it and its predecessors.

Today’s 14th generation Desktop Core is still a direct relative of the 12th generation.

For desktops and game notebooks,

the performance of the 12th generation is still good,

but for thin and light notebooks and all-round notebooks, it is miserable. The

battery life performance of the 12th and 13th generations has become the best among Intel notebooks. A shortcoming:

We have talked about the battery life issue of x86 platform notebooks before.

The current mobile CPUs have basically been SoC-based,

so in addition to the energy consumption ratio of the CPU core,

the energy consumption performance of the parts outside the core will also affect the battery life

, especially the average CPU load. Reducing the continuous power consumption outside the core is very important for battery life

in low-level daily applications.

The Intel Core Ultra processor we are going to test today

is designed to solve this problem.

Can it avenge its shame and

improve the battery life of Intel laptops ? What is the level of response?

Before the official test begins,

let’s briefly talk about the architecture of Core Ultra.

Because it is very new, this part will be longer.

If you want to see the performance test directly, you

can jump here.

Core Ultra is Intel’s complete restart of the product after the Core i series. From the perspective of

the overall structure of the SoC, it should be the biggest structural change since the 2nd generation Core i series

. In fact, modern processors have long been more than just a simple CPU.

SoC-based processors include CPUs and GPUs. , memory controller

, IO controller, media engine, AI acceleration unit, etc. There are a lot of things

, and the area of these units is no smaller than the CPU itself.

For example, this is the Dieshot of Apple M3

. Look at how much

chip area the CPU takes up. What?

Area is money!

A larger area means that one wafer can cut fewer processors and the yield rate is lower.

Not every area in an SoC requires the most advanced technology

, so in order to save costs,

each company has used a secret weapon.

For example, AMD made the CPU into a CCD

and other things separately to make an IOD.

The CCD was produced using advanced technology.

The IOD was produced using mature technology

and then placed on the same substrate

. This not only reduces the cost of advanced technology,

but also because of the chiplet. The small area brought about improves the yield rate.

It can be said that this design laid the foundation for the success of Zen2

, but it also brings some problems

. On the one hand, the chiplet design will cause multiple heat sources in the chip

, and the area of each heat source is large. Relatively small

, which may lead to "heat accumulation" problems.

On the other hand, longer distances mean higher delays.

The communication delay between CCDs or between CCDs and IODs will be relatively high.

In order to solve this problem,

Intel will also start to improve the "packaging process". "Get started

, and Core Ultra is the first mature application of this technology. If you

look closely at Meteor Lake's DIE,

you will find that it is actually composed of small chips.

These small chips are closely combined to form a large DIE

. Intel calls each small piece of chip

a Tile.

You would never guess that

these Tiles use different processes.

I made a schematic diagram here.

The CPU part uses Intel 4

and the GPU part uses TSMC N5.

Obviously these two areas The IO and SOC parts that require the most advanced processes

use TSMC N6, which

can reduce some costs while ensuring performance.

The most interesting thing is that these small chips are not connected through a substrate like other companies.

They are directly under these Tiles. A Base Die was made using the 22nm process

to connect them, and

one chip was used as the substrate.

These fabs really do whatever they want!

Connecting through such a Base Die

should have higher bandwidth and lower latency than connecting through the substrate.

Its technical principle is similar to that of 3D cache

. Except that it is expensive, it is indeed an extremely advanced idea.

These modular Tiles are put together like a jigsaw puzzle. into a complete SoC

and then completely polished it through a unified packaging process.

Let’s expand it to see what these Tiles are.

First of all, the CPU part

Compute Tile contains the Redwood Cove large core + Crestmont small core.

The overall architecture of the core has changed very little.

Basically, the design from the 12th generation is still used.

The IPC of the small core will be slightly improved.

It will still have 4 small cores in a cluster.

The full blood specifications are still 6P+8E

. In addition, this time the large core is actually divided into applications like the mobile phone SoC.

There are 2 cores in the performance library and 4 cores in the application density library

, which means that two of them are "super large cores".

On Ultra 7 and Ultra 9, when all cores are fully loaded,

they will run to a higher frequency.

In addition to these

, the CPU part also adds two extra small cores

, but they are not in the Compute Tile

but integrated in the SoC Tile

. The architecture of an ultra-small core

is basically the same as that of a small core. It mainly has a smaller cache and lower frequency.

The most important thing is that it is in the SoC Tile,

which means that it does not need to call the ring bus in the Compute Tile when it starts.

In other words, Meteor Lake can directly turn off Compute Tile in some scenarios

, which can greatly reduce the power consumption in these scenarios.

We will also run SPEC2017 of these cores separately later to see

whether the core architecture has been improved.

Let’s see how the ultra-small core performs. Next

, let’s take a look at the GPU.

The GPU architecture this time is called Xe-LPG.

You can understand it as a streamlined

version of the A770’s Xe-HPG. It mainly simplifies the XMX unit

. After all, Meteor Lake has an NPU.

The XMX unit seems a bit unnecessary.

As for other functions such as ray tracing, they are still

in the specification

. Full HP is 128EU.

I am very much looking forward to the performance of the core display performance this time.

This should be the highlight of Meteor Lake.

There is another interesting thing

in the GPU. The display part and media engine part have been moved from GT Tiles to SOC Tile.

The core display media engine has always been a very important part

, especially for video workers,

the core display media engine can effectively accelerate the efficiency of video editing.

This time the media engine It is also stronger

and has added support for the AV1 format.

We will also take a look at its performance later.

In addition to the CPU and GPU,

there is an important new member this time - NPU.

The NPU is located in the SoC Tile.

You can already I saw it in the Windows Task Manager

. Of course, it can run Stable Diffusion

and various AIGC applications that are compatible with OpenVINO

. But its focus is more on energy consumption ratio rather than absolute performance

. Therefore, currently these tasks are mainly focused on The main task of the NPU on the GPU

is to do

applications such as AI keying, video super-resolution, or motion capture.

In addition, from the entry-level to the flagship

NPU, they are exactly the same, without emasculating

the CPU, GPU, NPU, and adding new ones. SoC Tile and IO Tile

constitute the new Meteor Lake processor.

This modular design allows chips of different specifications to

be flexibly matched with different CPUs, GPUs and IO Tiles

, but the delay and bus bandwidth are controllable

, and CPU, GPU and NPU can be used. Completing AI computing

is also the design focus of this generation

. From a conceptual point of view, it is quite advanced.

Welcome back. After talking about the architecture,

I believe everyone is curious about

the performance of Meteor Lake. What

we are testing this time is a 2024 Lenovo. Xiaoxin Pro16

is a 16-inch all-round notebook.

It is equipped with Ultra 5 125H processor,

32G LPDDR5x 7467 memory.

Compared with the previous generation 2024 Xiaoxin Pro16,

the battery has been increased from 75Wh to 84Wh.

We will also test it with the new one. What is the battery life of the processor?

For comparison,

I bought two more

Xiaoxin Pro 16 equipped with i5 13500H and Ryzen 7 7840HS processors

respectively . Although Ultra 7 and Ultra 9 were not tested,

Ultra 5 is also the next main shipping model.

Let’s briefly review the specifications of

4 There are 8 small cores plus 2 ultra-small cores. The

single-core turbo frequency is 0.2GHz lower than the previous generation’s 4.7GHz.

The GPU’s 112EU Intel Arc core display is

slightly lower than the full-blooded 128EU

. So first, please take our SPEC2017 test

as the industry standard. Testing

it can tell us whether its single-core performance has really improved

and what exactly its ultra-small core is.

In SPEC2017,

the large-core performance of the Core Ultra 5 125H is obviously not as strong

as the previous generation i5 13500H. After all, the previous generation 4.7GHz The core frequency is a bit higher than the 4.4-4.5 of this generation,

and the IPC has basically not improved.

This performance is not surprising.

At the same time

, it cannot beat the Zen 4 7840HS, which has a lower IPC.

After all, although the same frequency performance is slightly It is weak

, but a single core can run up to 5.0 under PBO.

This frequency gap is really difficult to make up for

the performance of this large core. Compared with the Apple M3, which has a much higher IPC,

the gap is even greater. The performance of 125H

on small cores

is still significantly higher than 13500H.

The leading margin is also larger than the 0.1GHz difference in frequency,

which shows that the IPC of the small core should be improved.

In order to further compare whether the same-frequency performance has been improved,

I used 14900K to run the same large core 4.4 small core 3.6 memory

as the 125H. Running 6400G4

to roughly compare the performance of Meteor Lake and Raptor Lake at the same frequency,

it is not difficult to see that

the large core score of 125H is lower than that of the 14900K at the same frequency

. Could it be that the IPC of Core Ultra sucks?

This is mainly because the bus performance, cache performance, etc. of desktop CPUs will be higher. The

core IPC should not have changed.

After all, the architecture has not changed. However

, there has been an obvious change in

the small core. The performance of the small core is

even worse when the peripheral components are worse. It is higher than Raptor Lake at the same frequency.

It seems that the official IPC improvement

of the small core is not deceiving. The performance of this small core is also much better than the small core of M3.

Finally, let’s take a look at the newly introduced ultra-small core.

Judging from the SPEC2017 results,

this 2.5GHz small core is not very weak.

Its performance is close to the M3 small core.

If you want to compare it with ARM’s true small core A520

, this one can beat it. Three cores

are not a competitor of the same magnitude

, but in the subsequent tests,

you may be confused about the scheduling strategy of Intel's ultra-small cores like me.

Because there are almost no places to use them,

I also ran an inter-core delay test.

After all, this The physical location of the sub-ultra-small core is completely separate from other parts of the CPU.

As a result,

the delay of this ultra-small core is indeed relatively high

and close to the delay between Apple M series clusters.

Surprisingly, it is

between other cores of the CPU. The latency is also higher than that of Raptor Lake.

I don’t know if the performance of Ring has been reduced to save energy.

Let’s briefly summarize the performance of each CPU core.

The large-core co-frequency performance of Meteor Lake is almost the same as Raptor Lake.

The small-core co-frequency performance has been slightly improved.

As for ultra-small cores, because IPC and small cores are basically the same,

the main reason is the performance gap caused by the frequency difference.

Next

, we asked Cinebench R23, which can run at full CPU

, to see their multi-core performance and performance under full-core full load. Let’s compare the performance.

The first is the performance under the default power wall.

Under the default power wall,

it is a pity that

the Ultra 5 125H with a multi-core score of 70W cannot beat the previous generation i5 13500H,

let alone the 8 large-core Ryzen 7. In terms of single-core results of 7840HS

, similar to the results of SPEC2017,

the 7840HS with PBO is still the strongest

Ultra 5 125H frequency.

Next, let’s control the power consumption wall

to see what the energy consumption performance of Core Ultra’s CPU is. No improvement,

how can I say this result...

There is indeed an improvement

, but the improvement is really small.

Because of the disadvantage in the number of large cores,

it cannot get any advantage in front of the 7840HS

. If you look at this chart carefully,

you will find that A strange thing is

that when the power consumption wall is lower than 28W,

why is the 13500H even more powerful than the Ultra 5 125H?

So I set the power consumption wall at 15W

to observe the frequency strategy under 15W

. Something outrageous happened.

The small core was actually downclocked to 1.2GHz

, while the large core ran at 1.5-1.6GHz?

What kind of scheduling strategy is this?

Generally speaking, if you want to improve the energy consumption ratio,

you must try your best to let the small cores output more performance when the power consumption is insufficient,

so that the big cores can avoid the edge and lower the point. The frequency is used to run

Apple's low-power mode, which is exactly this idea.

However, I can't understand Intel's idea.

Four big cores are desperately sucking blood, and the small core frequency is extremely low.

In this case, it is strange that the performance can be good!

Moreover, the energy consumption performance of the first generation Intel 4 product did not meet my expectations.

Generally speaking,

I think the performance of the CPU part is only unsatisfactory

and has not reached the improvement I expected.

After reading the CPU energy efficiency, it is somewhat disappointing. Let’s take a look at

the performance of the GPU

compared to the CPU that has not been significantly changed.

The GPU can be said to be completely reborn this time.

In the 3Dmark Time Spy GPU test

, even the Ultra 5 125H Arc GPU with residual blood

completely surpassed it under the default power consumption wall. The performance of 780M compared to Ryzen 7 7840HS

can be said to be the shame of Yixue Intel Core Display!

Compared with the previous generation of Xe core display,

the performance this time has even more than doubled.

Of course,

the power consumption limit this time is also higher than before.

Basically, you will find that

its performance has stabilized above 45W

. The power consumption performance is actually relatively close to the 780M performance of the 7840HS,

while the core display specifications of the previous generation are much smaller.

The 25W power consumption wall is almost full.

This GPU

is indeed larger than the previous generation. However, there is one thing that surprised me.

It is within 15W.

The GPU scores of some Ultra 5 125H are lower than the 780M of 7840HS

. It seems that

this GPU is still too big for handheld use

. However, the next U series is the model

H series developed for this low-power platform. It is also mostly installed in models with 28W or above.

This performance is not surprising.

Let’s briefly summarize the results of the theoretical test.

The performance of Meteor Lake’s CPU part is not surprising

. The large core IPC has not improved, and the frequency is lower than the previous generation.

The small core IPC has improved, and performance has improved

. In terms of energy efficiency, there has been a slight improvement in mid- and high-frequency energy efficiency.

Low-frequency performance is not normal due to the scheduling strategy.

The GPU part of

Meteor Lake has greatly improved compared to the previous generation.

The performance in Time Spy has slightly surpassed the 7840HS, and

the performance is good

. After finishing the theoretical performance,

let’s take a look at the performance of Meteor Lake in practical applications.

The first is Cinebench 2024.

In Cinebench 2024,

the multi-core score of Ultra 5 125H is slightly better than 13500H.

This may be because the memory frequency of this generation is as high as 7467MHz.

Cinebench 2024 requires memory performance. It is relatively high

, but there is still some gap between it and the 8-core 7840HS.

In terms of single-core performance

, because the maximum turbo frequency of the large core is lower this time,

the result of 125H did not surprise me.

Then look at the performance of another 3D software - Blender. In

the CPU test,

the Ultra 5 125H is between 13500H and 7840HS,

slightly stronger than the 13500H.

What surprises me is that the GPU

does not support ray tracing because the previous generation Xe core display does not

support it. Although Blender's GPU renderer

7840HS supports light tracing

, Blender's HIP renderer keeps reporting errors

and cannot run the results.

As a result, only Ultra 5 125H successfully ran the GPU results.

It was a surprise.

Then I tried Adobe's software

PugetBench and updated them. The new version test.

This test process is extremely painful

, but fortunately we still finished the test.

Currently, this test includes PS and PR projects.

In Photoshop 2024,

Ultra 5 125H is slightly stronger than 13500H

, mainly because it is stronger in the Filter sub-item.

The core display should play its role. It is not a small effect.

The total score of 7840HS is higher than that of 125H. This

is mainly due to the 8-core CPU,

so the general sub-item score will be higher.

Premiere Pro

relies on the greatly enhanced core display performance

of Ultra 5 125H. The total score of 16 Pro is significantly higher than the other two models.

In the sub-items

involving GPU performance, Intraframe and GPU Effects

have significantly improved compared to the other two

. It seems that the new core display has indeed brought it Considerable improvement.

In the Media Encoder transcoding test,

with its larger core display scale and enhanced codec,

Ultra 5 125H also came out on top.

If paired with a discrete graphics card, the lead would be even greater.

Intel's media engine should be paired with a stronger GPU. It can also have greater potential.

In the 7Zip compression and decompression test, the 125H is slightly ahead of the 13500H,

but in the decompression test it still cannot beat the 8-core 7840HS.

The DaVinci test results are similar to the PR

of the Ultra 5 125H, relying on the substantial improvements in core display and media engine. The evolution

has greatly surpassed i5 13500H and 7840HS .

It seems that the new platform is very suitable for cutting videos. After

completing the application test, it is not difficult to see that

if it is a pure CPU load application,

the improvement of Core Ultra compared to the previous generation is very limited

, but as long as it is related to core display For applications

such as video editing, the performance of

Core Ultra has been greatly improved.

This is also in line with the results of the previous theoretical test.

We also briefly tested three games

, namely the lightweight Genshin Impact, Iron Break and the high-pressure Cyberpunk 2077

Original. The Arc core display of the Shenli

125H is nearly doubled compared to the previous generation

and exceeds the 780M core display on the 7840HS.

The current generation of core display can already meet the performance requirements of 1080P high-definition 60fps.

This performance is suitable for light gaming on business trips. It should be enough

, and the power consumption during the test was not

higher than the previous generation

. This performance is still continued in Honkai Impact: Star Dome Railway

. In Honkai Impact's 1080 low image quality, it can almost It reached 120fps.

Even in 1080P high-definition,

the average frame exceeded 60fps

and reached an average of 74fps.

Compared with the 27fps game experience of the previous generation, it was a qualitative leap.

It also defeated the 780M core display of 7840HS

and the Xe core display of the previous generation of

2077. Playing 2077 sounds a bit funny

, but the 125H core display is obviously much stronger than

the 1080P low image quality

without oversampling. It tied AMD's 780M core display. After turning on

the XeSS balance file,

it can run to an average of 60fps. After using XeSS upsampling

based on deep learning,

the picture quality is also very playable on the notebook screen. The

frame rate is slightly lower than the 780M using FSR.

I am also curious about one thing.

Since there are currently two companies, Intel and AMD. All core displays already support ray tracing

, and their performance has been greatly improved compared to previous core displays.

So what if I use them to run 2077's path tracing?

Hey, it really ran.

Although it only got "e-sports performance" with an average frame rate of 3fps and 2fps,

they didn't crash or explode the video memory

. After all, can 32G of memory give you such an easy burst?

In short, this test is just like Tu Yile.

Think about it. Now the core display can run path tracking

. Well, it’s good for Windows handhelds!

I simply ran a few games

and the performance of Intel's new core display was quite good,

both in 3Dmark and in actual games.

Even the 112EU residual version of Ultra 5 125H

has more wins than AMD's current 780M. Not

considering that AMD's 8040 series in 2024 will focus on one technology and rename it as

Core Ultra. This core display should perform very strongly in 2024. In

addition, the advantages of the media engine

can help accelerate

even when paired with a standalone display and Intel's core display.

This is good news for light gamers and creators.

So our performance test will start here.

I want to say where the weakest point of current Windows laptops is.

It must be the battery life performance.

This time Core Ultra brings so much energy consumption. So

how does it perform in the actual endurance test?

It’s time to bring out our magic mirror - the Geek Bay endurance test model.

Finally, in the balanced performance mode,

Xiaoxin Pro 16 equipped with Core Ultra 5 125H ran a time of 6 hours and 49 minutes.

This result is better than the previous generation equipped with 13500H. Xiaoxin Pro 16 has a longer battery life of 2 hours and 14 minutes,

which means the battery life is 49% longer. It

is also 1 hour and 24 minutes longer than AMD's 7840HS version.

Of course,

the battery capacity of the 2024 Xiaoxin Pro 16 has also been increased from the previous generation's 75Wh to 84Wh

. Calculated in terms of battery capacity, if the battery capacity is equal,

a notebook equipped with Core Ultra will have about 30% longer battery life than the 13th generation Core model,

and about 12% longer than the 7840HS.

Although I think this is not disruptive, it is

still far from the MacBook. There is a considerable gap in battery life

, but at least this should be the platform with the longest battery life

among Windows laptops you can buy currently.

However, higher expectations will be reserved for next year’s Lunar Lake. There

is one more thing I need to mention here

. The two ultra-low-power small cores introduced

for the first time will not be used in daily applications.

They can only be used in standby idle state, when watching videos locally,

or in the software specially programmed for them.

Our battery life test The model has almost no idle time

and no scene for watching videos locally,

so this ultra-small core can hardly bring any gain in battery life in our battery life test.

Maybe it will help in the PCMark10 battery life test

, but I think the battery life of PCMark10 The test model is seriously out of touch with actual usage.

I think Intel should really allow more applications for ultra-small cores

. After all, what kind of performance do I need to type in Word?

If you don’t let me use the ultra-small core at this time, when should I use it?

After all, it is not a small piece of garbage with performance like A520.

To be fair, its performance is actually enough for typing.

I really can’t understand Intel’s scheduling strategy this time.

In addition to these

, Meteor Lake also introduces the AI acceleration unit NPU,

which is why it can say itself It is the core of an AI computer.

It can accelerate AI applications with much lower power consumption and

can also run AIGC applications such as Stable Diffusion.

Of course, its focus is on energy consumption ratio rather than performance, and

it is more oriented towards noise reduction, etc. Cut out the picture. This kind of native, low-power AI application needs.

If you are really an AI practitioner,

NPU still cannot replace the work of independent graphics.

So here

our first review of Core Ultra 5 125H is coming to an end. In addition to

the CPU part

, Core Ultra basically continues the overall architecture since the 12th generation.

The small core has a small architectural evolution.

The core display part has made considerable progress.

The game performance can now achieve a 780M core display that is not inferior to the 7840HS.

It also supports XeSS, ray tracing and other functions.

I think the performance of the core display this time is very good. In

terms of battery life

, Core Ultra is significantly improved compared to the previous generation

, but it is obviously still far behind the battery life of Mac.

The addition of NPU is definitely a good thing

, because many AI applications that do not require high computing power can now be run directly on NPU

, so there are many small ones. AI functions can be integrated into software

and can be used on the move.

Although the current application scope is still very limited

, I believe this is definitely a good start for Windows PC.

At present,

although it has not yet reached Apple in this field of thinness and lightness The level of the M series

is still the best choice among current Windows thin and light notebooks.

The current price of the 2024 Xiaoxin Pro 16 using Ultra 5 125H

is 5999 yuan.

At the same time, the previous generation version using i5 13500H is currently 5699 yuan

. Do you think the current difference of 300 yuan is worth upgrading?

In general,

the Core Ultra of the Meteor Lake architecture

is indeed the generation that has changed the most in recent years

as Intel said , but you will find that the focus of the so-called "change" this time

has shifted from the CPU level to the SOC level.

Maybe next year's Lunar Lake will be this one. Under the completeness of the strategy,

we will also look forward to Intel's performance in subsequent products.

As far as the current Meteor Lake itself is concerned,

I think the changes are significant, but

the efficacy is not enough. We still need to take a bigger step. Okay, that

's it for this issue. All the content of the program.

If you like our program, don’t forget to press and hold the like button to give us a compliment.

Don’t forget to follow our channel.

Don’t forget to go to our Taobao store to see the clothes I’m wearing.

Then we'll see you in the next episode