News Alleged Intel Raptor Lake Sample 20% Faster Than Alder Lake in New Benchmarks

[Admin]

A Chinese tech site has got an Intel 24C/32T engineering sample and pitched it against the Intel Core i9-12900K at equal clocks. In multi-threaded tests, the purported Raptor Lake chip appears to be as much as 20% faster than Alder Lake.

Alleged Intel Raptor Lake Sample 20% Faster Than Alder Lake in New Benchmarks : Read more

 

[-Fran-]

Well, as mentioned in the other previous leaks of the ES tests, just keep in mind these do not use final clocks and are, sometimes, full of bugs and need final microcode and stuff.

Reading a bit between the lines though... Doesn't seem like Raptor Lake has major IPC improvements over Alder Lake, but neither clock speeds, so I have a feeling their overall performance increase is not going to be big. Like under 10% most of the time for low threaded works and about 20% in MT workloads, ideally more since the small E cores were improved.

Time will tell.

Regards.

 

[KyaraM]

I'm a bit confused by the conclusion. It looks to me as if Raptor Lake is slower than Alder Lake in single-threaded workloads, why does the article talk about a lead? What did I miss?

 

[spongiemaster]

I'm a bit confused by the conclusion. It looks to me as if Raptor Lake is slower than Alder Lake in single-threaded workloads, why does the article talk about a lead? What did I miss?

What are the odds that adding a bunch of cache and increasing clock speeds is going to result in slower single-threaded performance?

 

[LuxZg]

From what I've read, Raptor could just as well suffer slight clock regression, so improved clocks aren't certain at all.

And some workloads don't improve with cache.

Thing is, article says one thing, pictures say something else entirely. AND that's at same clocks (underclocked Alder Lake). Both games and both single threaded pictures show lower results. And 20% MT is purely from more cores. 24 thread Alder Lake + 33.33% = 32 thread Raptor Lake. So that single threaded disadvantage even works against it in MT, otherwise we would see 25-30% better MT scores

Ok, we could talk till tomorrow defending each side. One could say they will have barely any ST advantage over AL, and ~25% MT purely due to extra cores, and little or no clock advantage. Basically a typical tick-tock strategy, where they will (finally) lower power requirements and get back in line. Other side could claim that this is all early sample, no microcode, wrong MBO/chipset, that they will get to 6GHz, with +10% ST, and +40% MT, and so on.

But this test does not look good for Raptor Lake at all, despite what the article text says. And looks equally bad for author and website publishing such texts without clearly saying it's all baseless guessing at it's best.

 

[TerryLaze]

And 20% MT is purely from more cores. 24 thread Alder Lake + 33.33% = 32 thread Raptor Lake. So that single threaded disadvantage even works against it in MT, otherwise we would see 25-30% better MT scores

That would work if all threads had the same performance but e-cores and p-cores don't have the same performance, even if they would run at the same clocks which they don't, and then there is HTT on top of that.

 

[spongiemaster]

From what I've read, Raptor could just as well suffer slight clock regression, so improved clocks aren't certain at all.

And some workloads don't improve with cache.

Thing is, article says one thing, pictures say something else entirely. AND that's at same clocks (underclocked Alder Lake). Both games and both single threaded pictures show lower results. And 20% MT is purely from more cores. 24 thread Alder Lake + 33.33% = 32 thread Raptor Lake. So that single threaded disadvantage even works against it in MT, otherwise we would see 25-30% better MT scores

Ok, we could talk till tomorrow defending each side. One could say they will have barely any ST advantage over AL, and ~25% MT purely due to extra cores, and little or no clock advantage. Basically a typical tick-tock strategy, where they will (finally) lower power requirements and get back in line. Other side could claim that this is all early sample, no microcode, wrong MBO/chipset, that they will get to 6GHz, with +10% ST, and +40% MT, and so on.

But this test does not look good for Raptor Lake at all, despite what the article text says. And looks equally bad for author and website publishing such texts without clearly saying it's all baseless guessing at it's best.

Not sure what rumor sites you frequent, but I have not seen any rumor that Raptor Lake would have lower clock speeds than Alder Lake. The most recent rumors from yesterday claim that it could reach 6GHz with special editions CPU's.

Intel XTU gets Raptor Lake update with “Efficient Thermal Velocity Boost” & per-core/package OC TVB support

From the Raptor Lake performance leak:

"In this case the P-cores are running at 1.4 GHz to 3.8 GHz range, while one cluster of E-Cores runs at 1.0 GHz and the other at 2.0 GHz. There is therefore some inconsistency in frequencies, as noted by Wccftech. "

So, there were issues normalizing the clock speeds. In actual use, neither eCore cluster is going to run at only 1GHz.

 

[PCWarrior]

The test is in POV Ray 3.7.1. In POV Ray the 12900K normally scores 10410 points while the 12700K scores 9013. So with 4 extra e cores as well as slightly higher clock speed the increase is 1397points. So, given the small differences in clock speeds between the 12700K and the 12900K, the contribution of 4 e-cores alone in 12900K is around 1005 points and the contribution of 8 p-cores is around 8400. Now if you have 16 E-cores you would expect their contribution to be 4020points. And if you have 8 P-cores with 10% performance regression their contribution would be 7560points(=0.9x8400). Thus, you would expect a total of 11580 points i.e. an 11.2% increase compared to the score of a retail 12900K. But here instead we get 20% higher than retail 12900K which would mean a score of 12492. That means that the MT performance of the e-cores in Raptor lake (in POV-RAY) improved by 22.7% (from 4020points contribution to 4932points). That's probably due to the larger cache.

It has to be said here that POV Ray is a test where in Alderlake the percentage contribution of the E-cores to the overall score is rather small compared to their contribution in other synthetic MT tests (e.g. Cinebench). So there was plenty of room for improvement which we see here. Now if we also account for the expectation that that the P-cores of a retail 13900K should see at least a 10% performance improvement (instead of a 10% regression) compared to a retail 12900K, the contribution of the P-cores should be at least 9240 points. Thus, a total 14172 points i.e. the 13900K should score at least 36.1% higher. If we see a 15% performance increase in the P-cores then we are looking at an increase of over 40%. With clock speed increases in the e-cores we exceed that increase even further.