Intel Core Ultra 9 275HX Crushes Core i9-14900HX In Cinebench R23 Multi-Core, While Losing Out in Single-Core

[GGforever]

The Intel Core Ultra 9 275HX was unveiled by Big Blue at CES 2025, packing 8 Lion Cove performance cores and 16 Skymont efficiency cores for a total of 24 cores and 24 threads, thanks to Intel doing away with Hyper-Threading support. As such, we already had quite low expectations for multicore performance from Arrow Lake-HX, although a recently listed Cinebench R23 run appears to confirm the contrary. However, considering the nature of leaks, be sure to accept this information with a fair amount of skepticism until more information comes our way.

To be honest, this result is quite a doozy, for the aforementioned reason. Like its predecessor, the Core Ultra 9 275HX will also boast the same number of performance and efficiency cores, without Hyper-Threading support. But somehow, the result indicates that the 275HX handily outperforms the Core i9-14900HX in multicore performance by a whopping 18%, despite losing out in single-core performance by the skin of its teeth. Clearly, something is going on here, that I am unable to wrap my head around. It is possible that Intel has massively improved multicore scaling, but that is not what we witnessed with a recent Passmark result, where the Core Ultra 7 255H portrayed healthy improvements over its predecessor in single-core, but not quite so in multicore.



View at TechPowerUp Main Site | Source

 

[Bobaganoosh]

It may be able to boost higher due to the lower power usage and heat-load. That tends to hold back the mobile 13th/14th Gen cpus a lot.

 

[TheinsanegamerN]

they should throw 128 old atom cores int here, to really take the MT crown.

 

[Vya Domus]

12% is not exactly "crushing" lets be real here.

 

[xSneak]

More treading of water.

Regarding the removal of HT, is Intel using the "rentable units' tech with these ?

 

[Darmok N Jalad]

It may be able to boost higher due to the lower power usage and heat-load. That tends to hold back the mobile 13th/14th Gen cpus a lot.

Yeah, I suspect ARL is able to maintain higher clocks for the duration. The E-cores are already clocked higher in ARL, so that might be entirely responsible for the performance gain. Cinebench is a fun number and all, but a program designed to max out all the cores is not all that representative of daily usage. Where I see P+E struggle is on mixed use, where a program only occasionally needs as much multithreaded power as it can get, and that’s usually when P+E falls on its face.

 

[ScaLibBDP]

12% is not exactly "crushing" lets be real here.

I've calculated the difference of the multi-threaded performance as follows: ( 31,854 * 100.0 / 35,481 ) - 100.0 = ~10.22%.

Note: In HPC world a difference greater than 5% is considered as a good improvement. On a Big Data set it will reduce processing time.

 

[Daven]

I like how there is different math from different people. From the same data set:

GGforever says 18%
Vya Domus says 12%
ScaLibBDP says 10%

Does anyone want to show math for 5%? 25%?

 

[ScaLibBDP]

I like how there is different math from different people. From the same data set:

GGforever says 18%
Vya Domus says 12%
ScaLibBDP says 10%

Does anyone want to show math for 5%? 25%?

I've described how I've calculated the difference. Next, I've just used the Microsoft Copilot to verify if my way is wrong. Here is a log of my "conversation" with Microsoft Copilot.

The result depends on what number is used as a base to compare with another number. So two possible cases:

...31,854 is lower than 35,481...
or
...35,481 is greater than 31,854...

Both results are Valid!

...
[ ScaLibBDP ] Calculate for how many percent 31,854 is lower than 35,481

[ Copilot ]
To find how much percent 31,854 is lower than 35,481, you can use the following formula:
Percent Lower=DifferenceOriginal Number×100\text{Percent Lower} = \frac{ \text{Difference} }{ \text{Original Number} } \times 100
Let's plug in the values:
\text{Difference} = 35,481 - 31,854 = 3,627 $$ $$ \text{Original Number} = 35,481
Now, use the formula to calculate the percent lower:
Percent Lower=3,62735,481×100≈10.22%\text{Percent Lower} = \frac{3,627}{35,481} \times 100 \approx 10.22\%
So, 31,854 is approximately 10.22% lower than 35,481

...
[ ScaLibBDP ] Calculate for how many percent 35,481 is greater than 31,854

[ Copilot ]
To find how much percent 35,481 is greater than 31,854, you can use the following formula:
Percent Greater=DifferenceOriginal Number×100\text{Percent Greater} = \frac{ \text{Difference} }{ \text{Original Number} } \times 100
Let's plug in the values:
\text{Difference} = 35,481 - 31,854 = 3,627 $$ $$ \text{Original Number} = 31,854
Now, use the formula to calculate the percent greater:
Percent Greater=3,62731,854×100≈11.39%\text{Percent Greater} = \frac{3,627}{31,854} \times 100 \approx 11.39\%
So, 35,481 is approximately 11.39% greater than 31,854

 

[ncrs]

I've calculated the difference of the multi-threaded performance as follows: ( 31,854 * 100.0 / 35,481 ) - 100.0 = ~10.22%.

Note: In HPC world a difference greater than 5% is considered as a good improvement. On a Big Data set it will reduce processing time.

Good improvement is Zen 4 to Zen 5, for example 9800X3D vs. 7800X3D under Linux where the improvement is ~25% using your formula.
Those 10% we're not even sure of because both Intel CPUs can scale from 45W to 160W, so unless both CPUs run at the same power limits with similar cooling it's not a useful comparison for laptop chips.

 

[GGforever]

I like how there is different math from different people. From the same data set:

GGforever says 18%
Vya Domus says 12%
ScaLibBDP says 10%

Does anyone want to show math for 5%? 25%?

The median for the 14900HX is ~30k, thats what I used....

 

[ZoneDymo]

ok but does anyone care?