Discussion Cinebench 2026

[Abwx]

Is there a rule that SMT gains must be equal between all renderers?

AMD leads in this benchmark regardless if the SMT gain is 35% or 44%.

Guess that it depend of the exact code, with CB R20/23 versions it was the only renderer where Intel was ahead, and the one with the lower SMT gain, also it seemed that CB 2024 was a lighter load than older CB versions for power usage, dunno for this one, some members could give us some hints.

EDIT : Computerbase has an article and publish their members scores here as well as their own for both CPUs and GPUs :

Cinebench 2026: Der Community-Benchmark-Test!

Maxon hat überraschend Cinebench 2026 veröffentlicht. Die Community weiß, was das bedeutet: Community-Benchmark-Test!

www.computerbase.de

 

[Jan Olšan]

Ryzen 9 7950X stock settings, air cooled, 32GB DDR5-6000(EXPO)

Multithread: 8167
Single Thread: 481
Single Core: 629 (!)
(Radeon RX 7700 XT 12GB: 35127)

BTW, I got 6333 MT score when limiting the processor to TDP:65W/PPT:88W. That's 77,5 % performance with 38,3 % of the maximum power, should be about 2x better power efficiency when only considering processor SoC power. And that's with the large fixed overhead of the IO Die's power consumption that AM4 and AM5 chiplet CPUs suffer from and which likely doesn't scale down when TDP is lowered, so the efficiency gain may be even bigger?

Let that be a reminder for people that think CPU cores have one given level of power efficiency. The reality is that it's a whole range where the progression is very non-linear and the single most important thing is which point (what clock/voltage combination out of the particular core's range) you are running at.

 

[FlanK3r]

crashing on load for me.... sad wanted to test.

I have the same, with Windows 10, Ryzen 7 9700X on stock, and Radeon rx6800....

 

[Z O X]

5700X3D
SMT on: 3241
SMT off: 2570

6800XT: 43025

 

[mikegg]

Let that be a reminder for people that think CPU cores have one given level of power efficiency.

No one thinks that.

 

[Markfw]

In 24 hours I can try this on my 9755 + 4080.

 

[511]

185H
1T - 424
1t on Crestmont E is 242

 

[MS_AT]

So did anyone try to observe if 1core setting acutally pins 2 sibling logical cores of single physical core for the whole duration? Or if it lets them migrate, does the pair always land on the same physical core?

Would try myself but won't have access to a PC for few more days.

 

[Det0x]

So did anyone try to observe if 1core setting acutally pins 2 sibling logical cores of single physical core for the whole duration? Or if it lets them migrate, does the pair always land on the same physical core?

Would try myself but won't have access to a PC for few more days.

Not exactly what you are asking about, but i saw something else when running the single core benchmark.
All the load is always get put on core #0, which means it dont follow the CPPC ranking (this core is ranked #4)



This is how it looks like running the single core bench on a 8core v-cache CPU with SMT enabled


Running the singlethreaded benchmark, the workload are jumps around between cores/threads


So this benchmarks single core benchmark kinda seems to have a little issue with the windows scheduling / following the correct CPPC core ranking
Anyone else getting the same results ?

 

[Jan Olšan]

No one thinks that.

I've seen people who may not realise but seem to assume that. There was that guy who claimed that Apple has a "7X" or "9X" something like that advantage on efficiency over AMD, that AMD has to "overcome". Yeah, an exact number, although I'm not sure if I remember the value right.

Almost every time this topic comes up and somebody compares a desktop CPU with usually an Apple SoC and uses their TDPs (or some guesses at power consumption in case of Apple) to make claims about efficiency of the microarchitectures in the processors, there's the implicit forgetting that the desktop chip is merely at one of its possible efficiency settings (usually close to the worst one).

 

[511]

I've seen people who may not realise but seem to assume that. There was that guy who claimed that Apple has a "7X" or "9X" something like that advantage on efficiency over AMD, that AMS has to "overcome". Yeah, an exact number, although I'm not sure if I remember the value right.

well X86 Camp likes to run cores higher Frequency way outside of optimal power/performance curve on DT and in general as well

 

[MS_AT]

So this benchmarks single core benchmark kinda seems to have a little issue with the windows scheduling / following the correct CPPC core ranking

Thanks for checking. It seems they have hardcoded the affinity to (logical) cores 0-1 then. While for single threaded benchmark they probably don't set the affinity at all to let Windows decide.

 

[poke01]

well X86 Camp likes to run cores higher Frequency way outside of optimal power/performance curve on DT and in general as well

they also run most of their laptop CPUs outside the optimal curve too.

 

[poke01]

M4 Max - This is a Mac Studio run.

Single thread- 715
MT- 7848

 

[Hulk]

I was so much hoping for a return of the swarm of bees work mode from the 2023 version. This is so boring to watch.

 

[jdubs03]

M4 Max - This is a Mac Studio run.

Single thread- 715
MT- 7848

View attachment 135878

That’s quite a ST score. The highest I saw for the M5 over in the computerbase forums so far has been 744.

 

[poke01]

That’s quite a ST score. The highest I saw for the M5 over in the computerbase forums so far has been 744.

Yeah I saw. The M5 is quite powerful for a tablet SoC, also beats the AMD HX 370 in the nT test.

 

[Gideon]

Here are my runs:

For my 2 work computers and home rig:

• 14" Mac M1 Pro (Late 2021 model)
• 14" Lenovo T14s, Zen 3 6850U (March 2022 model)
• Ryzen 5800X3D with RTX 4090 (Still rocking on a B350 Tomahawk from 2017)

The work macines are somewhat interesting as they are about the same form/factor and models from the same timeframe. Too bad neither CB2024 nor CB2026 GPU test run on the AMD iGPU, despite it being RDNA2. That would have been very interesting (as even the memory setup is similar)

	14" Macbook M1 Pro	Thinkpad T14s Zen3	Home Desktop
CPU	M1 Pro 10C @ 3.23 GHz	Ryzen 6850U @ 2.7GHz (ST 4.7 Ghz)	Ryzen 5800X3D, PBO undervolt
GPU	M1 Pro 16-core	Radeon 680M (RDNA2)	RTX 4090 24GB
Memory:	32GB LPDDR5 @ 6400 MT/s	32GB LPDDR5 @ 6400 MT/s	32GB DDR4 @ 3600MT/s
Cinebench 2026 GPU	11 157	-	140 327
Cinebench 2026 MT	3266	2176	3621
Cinebench 2026 SC	-	418	472
Cinebench 2026 ST	430	332	337
Screenshot	img	img	img
Notes:	Mode: - MT clocks: ~3.2 Ghz SC clocks: ~3.2 Ghz ST clocks: ~3.2 Ghz OS: Macos Sequia 15.6.1	Mode: High Performance MT clocks: ~3.1 Ghz SC clocks: ~4.2 Ghz ST clocks: ~4.35 Ghz OS: Windows 11 25H2	Mode: Balanced MT clocks: ~4.38 Ghz SC clocks: ~4.48 Ghz ST clocks: ~4.422 Ghz OS: Windows 11 25H2

Some notes:

• These are all runs in day-to-day rigs with years old OSes and unchanged thermal paste on the laptops. I just closed most of the resource consuming apps (and real-time protection in Windows Defender), but that was it.
• The 8/16 Ryzen 6850U barely finished the first MT run in 10 minutes (in fact it only did 1 run with Mode set to Balanced). In contrast the average run of the RTX 4090 was 8.35 seconds (doing 71 passes in total) No wonder people are saying CPU rendering is dead.
• ST load bounces around in windows thus lowering the max clocks vs the pinned SC load on my desktop
• The 5800X3D manages to beat the reference 5800X in MT but falsl behind in ST (due to relatively low clocks)

 

[eek2121]

So did anyone try to observe if 1core setting acutally pins 2 sibling logical cores of single physical core for the whole duration? Or if it lets them migrate, does the pair always land on the same physical core?

Would try myself but won't have access to a PC for few more days.

Not exactly what you are asking about, but i saw something else when running the single core benchmark.
All the load is always get put on core #0, which means it dont follow the CPPC ranking (this core is ranked #4)

View attachment 135872

This is how it looks like running the single core bench on a 8core v-cache CPU with SMT enabled
View attachment 135873

Running the singlethreaded benchmark, the workload are jumps around between cores/threads
View attachment 135874

So this benchmarks single core benchmark kinda seems to have a little issue with the windows scheduling / following the correct CPPC core ranking
Anyone else getting the same results ?

Yes, I read on the overclock forums that they pin the workload to core 0, which is less than ideal.

It would also be nice if we could get a Linux port. Alas…

 

[MS_AT]

which is less than ideal.

Well it could be worse. I mean on both Intel and AMD CPUs there is a high chance that physical core 0 is among the highest clocking cores. (On AMD the CCD0 will clock higher than CCD1 and variance within CCD is negligible most of the time. On Intel core 0 is bound to be Pcore). Otherwise they would need to write additional code to figure out which core clocks the highest.

I am not sure if Windows has an api to read this info out so might be they would need a custom solution and different one for each CPU vendor if Windows does not make this info accesible to 3rd party devs.

 

[Det0x]

Well it could be worse. I mean on both Intel and AMD CPUs there is a high chance that physical core 0 is among the highest clocking cores. (On AMD the CCD0 will clock higher than CCD1 and variance within CCD is negligible most of the time. On Intel core 0 is bound to be Pcore). Otherwise they would need to write additional code to figure out which core clocks the highest.

Well except the all the dual CCD X3D cpus where CCD1 is the higher clocking one 🤷‍♀️
Such as 7950X3D, 7900X3D, 9950X3D and 9900X3D i believe

And there seems to be very little if not zero benefit from the v-cache in this benchmark (?)

 

[Hulk]

Well except the all the dual CCD X3D cpus where CCD1 is the higher clocking one 🤷‍♀️
Such as 7950X3D, 7900X3D, 9950X3D and 9900X3D i believe

And there seems to be very little if not zero benefit from the v-cache in this benchmark (?)

No benefit from v-cache. Data is not being "needed" from the v-cache. It's more of a "fight for data in the L3" as more threads are added to the mix. But that is even pretty minimal with this bench. It's really a "straight through" number cruncher.

 

[lightmanek]

My contribution is from ASUS Vivobook S16 with Ryzen 7 365 and 32GB LPDDR5 7500MT.


iGPU to start, laptop plugged-in in Balanced power plan.

How nicely aligned my CPU is with Snapdragon 12C ... first run and Balanced power profile on my work W11 install with a ton of spyware ...


Finally, ST and SC scores are quite decent for a cheap laptop

 

[Hulk]

This is what the scaling looks like as you increase thread count in CB26. First it adds physical cores, then after the 16 physical cores it starts to add logical cores. Hence the two different scaling slopes. This is frequency normalized data. As you can see, you get about 35% from the logical cores in this bench.

 

[Hulk]

I don't want to get into this but Gideon's results were so well documented I couldn't resist charting them along with mine.