Tuesday, July 20th 2021
![Intel](https://tpucdn.com/images/news/intel-v1734446766296.png)
Intel Core i9-12900K Allegedly Beats AMD Ryzen 9 5950X at Cinebench R20
With qualification samples of the upcoming Intel Core i9-12900K "Alder Lake-S" processors and companion Socket LGA1700 motherboards hitting the black-market, expect a deluge of benchmarks on social media. One such that stands out makes a fascinating claim that the i9-12900K beats AMD's current flagship Ryzen 9 5950X processor at Cinebench R20, which has been AMD's favorite multi-threaded benchmark. At stock speeds, with liquid cooling, the i9-12900K allegedly scores 810 points in the single-threaded test, and 11600 points in multi-threaded.
To put these numbers into perspective, a retail Ryzen 9 5950X scores 641 points in the single-threaded test, and 10234 points in multi-threaded, in our own testing. The i9-12900K is technically a 16-core processor, just like the 5950X, but half its cores are low-power "Gracemont." The "Alder Lake-S" chip appears to be making up ground on the single-threaded performance of the "Golden Cove" P-core, that's a whopping 25% higher than the "Zen 3" core on the 5950X. This is aided not just by higher IPC, but also the max boost frequency of 5.30 GHz for 1~2 cores, and 5.00 GHz "all-core" boost (for the P-cores).Given the multi-threaded scores, it's safe to assume that either Intel or Microsoft has figured out a way to leverage the P-cores and E-cores simultaneously in peak multi-threaded workloads. This is possible when both the "Golden Cove" and "Gracemont" cores have the ISA capability needed by the workload, which in case of Cinebench R20, is AVX. "Gracemont" is Intel's first low-power core to support AVX, AVX2, and AVX-VNNI instruction sets. "Golden Cove" features a more lavish ISA that includes AVX-512 (select client-relevant instructions).
Sources:
OneRaichu (Twitter), VideoCardz
To put these numbers into perspective, a retail Ryzen 9 5950X scores 641 points in the single-threaded test, and 10234 points in multi-threaded, in our own testing. The i9-12900K is technically a 16-core processor, just like the 5950X, but half its cores are low-power "Gracemont." The "Alder Lake-S" chip appears to be making up ground on the single-threaded performance of the "Golden Cove" P-core, that's a whopping 25% higher than the "Zen 3" core on the 5950X. This is aided not just by higher IPC, but also the max boost frequency of 5.30 GHz for 1~2 cores, and 5.00 GHz "all-core" boost (for the P-cores).Given the multi-threaded scores, it's safe to assume that either Intel or Microsoft has figured out a way to leverage the P-cores and E-cores simultaneously in peak multi-threaded workloads. This is possible when both the "Golden Cove" and "Gracemont" cores have the ISA capability needed by the workload, which in case of Cinebench R20, is AVX. "Gracemont" is Intel's first low-power core to support AVX, AVX2, and AVX-VNNI instruction sets. "Golden Cove" features a more lavish ISA that includes AVX-512 (select client-relevant instructions).
155 Comments on Intel Core i9-12900K Allegedly Beats AMD Ryzen 9 5950X at Cinebench R20
You make a true statement, but fail in two regards. The biggest thing is your assumption about how much waste energy it consumes.
There are some references to Cobalt out there you can find. Bottom line though, it is a better conductor at sub 20nm than copper. This is because of the electro migration problem, basically that loss becomes an issue.
Less resistance = less heat which means greater density can be achieved.
Now, it may be that a chip running at 250W still produces the same heat. However, if that chip is getting 30% more work done, that is the win.
The argument that it is going to produce the same heat is sort of like the EPA / HP argument I see about cars these days. Cars are not getting significantly more MPG than they did in the mid to late 80s, that's a fact. However, they get slightly better MPG and have double the power. So yes, they are far far more efficient, we are just using that efficiency to make them go faster, not to save gas.
The same thing is likely to be true of Cobalt.
It is a fact.
A chip consuming 250W IS giving 250W of heat
There is no physical work done inside the CPU
ALL energy eventually change into heat inside your magic cobalt CPU.
And cobalt also has 1/4 of thermal conductivity of copper
So the heat produced inside the CPU will be much more "trapped" within.
It is Physics.
You cannot deny Physics.
You pack 250W into a smaller die
The heat density goes sky high
Your magic cobalt won't save you there
These new CPUs will most likely suffer with temperature spikes causing frequency drops.
Just like how Ryzen 3000 series suffered when they implement N7 node for the very first time.
How Intel's gonna use that 250W of energy is another story.
You might HOPE Intel suddenly produces the most efficient CPU on the planet.
But look at Intel's track record.
They DO NOT tune their mainstream CPU towards the efficient side.
But instead,
They pump whatever current they could find. Agreed
I think the voltage headroom are most likely to be used to compensate the frequency lost going 14nm to 10nm.
They will suffer the hotspot problem same as Ryzen 3000/5000
Combined with the trash windows scheduler (Already proven by how bad the EOL Lakefield performed)
I don't have high hopes for their first gen 10nm desktop parts.
- When it comes to manufacturing not too sure about what Intel has but other foundries all have power-optimized node variants. These are used mostly for mobile chips and have the caveat that they do not clock as high as the high performance variants that desktop/mainstream CPUs and GPUs are built on. Even if you get a noticeable efficiency boost and it helps with all core loads, the maximum frequency for single-core or a couple-core load will take a hit which puts you in a competitively bad position. For a recent real-life example - with a high-performance node but the same idea - see (early) Ice Lake.
- Other than that all you can play with is voltage and frequency. Unfortunately lower voltage also means lower frequency and competitive situation again dictates you want as high a frequency as possible.
There is a curve for efficiency and the most efficient point is easy to find - just check what big Xeons and EPYCs run at. :)
Edit:
They do not seem to be too far off one another:
Ice Lake Xeon 8358 apparently can run 32 cores at 2.6GHz in 250W TDP (note: tricky because in case of Intel this usually includes AVX-512).
Zen3 EPYC 75F3 does 32 cores at 2.95GHz in 280W TDP. Yes, hotspot problem is likely. On the other hand, this might prompt Intel to stop allowing/encouraging manufacturers to use indefinite Tau as default :)
Windows and its scheduler has been improved somewhat since Lakefield. Lakefield was as much for the purpose of getting started on the schedulers problem than anything else.
Looking at Tiger Lakes and the new Ice Lake Xeons, 10nm does not seem to be that far behind TSMC's 7nm. You are right, high frequencies may be different but we can hope :)
Intel really like short bursts (with 2 x TDP power) which throws the efficiency out of the window
Then deep dive the frequency to compensate for the "debt" of power / heat output
This behavior is really obvious in the mobile CPU segment.
Hey Xeons / EPYCs are going 270W too..
Not to the same degree as unlimited Intel CPUs but it is still pumping 140W into 8 cores for ~4GHz. Boost clocks need quite a lot more than that.
I would really love for those 810 ST points to be real because this would mean Alder Lake is 50% faster than Comet Lake, as it was supposed to be, but this will not translate to games.
That's what all the stuff I have read suggests!
I know its a new design but I still can't see how it can be twice as fast as rocket lake in multicore when it is only 25% or so faster in single threaded performance. It just doesn't add up which brings me back to some kind of intel trickery is at play here!
But they will be out in the next 6 months so we will get our answer soon enough!
But mother board manufactures especially asus "but not exclusively" usually give oc'ers options to use all resources and hopefully that includes all available cores even the small ones lol
Also most oc'er disable all background stuff anyway so these 8 small cores wouldn't be doing anything anyway or bare minimum on 1 or 2 small ones.
So I haven't forgotten anything lol
I apologize for my English. I did not break the 5.3 but I have run up a lot of electricity and I also changed the Corsair Hydro Series H100i Pro with an Arctic Freezer 240 because the temperatures reached the ceiling, so I have higher performance but at a higher cost
Anyway, that 3D V-Cache spells doom. Zen 4 is coming. I feel like Alder Lake won't be able to show off for long. ;)
All the fanboys in this thread as usual, including HenrySomeone of course. I don't understand why y'all only praise one company as if they care about you. For me, Intel being competitive again is great because in like 5 years when I make a new rig, I'll have more options and a huge performance boost over my current rig. I don't have to time to spend online arguing with people over which company's boots to lick. Have fun with the fanboyism now lol.
Alder Lake is supposed to come out late of 2021. Zen4 should be coming late 2022 which Intel should be countering with Meteor Lake at the same time frame.
My country has a very odd London, it actually sounds like fucked up Victoria 2 map. It's a small garden community (a small village like land with small territories for owning gardens, and small houses in eldership of Switzerland, which is in Jonava municipality. It only has 14 citizens nowadays. At its peak, it had 19 citizens and at its lowest point it only had 3 citizens. My city region is quite weird one. It also has Switzerland village. In my city I once found one place, where all street were named after Scandinavian countries for seemingly no reason at all. I know that one street is named French street, also for unknown reasons.
When retail silicon is reviewed then get excited or not.
If they have high ipc and low power they re expensive..dont worry users who dont care about price, science labs and wallstreet will pay for that chips they dont need save money.
A high end lab will never use amd if they have the money due to avx