News Core i9-14900K, Core i7-14700K CPUs Benchmarked

[bit_user]

The problem here is that you cannot extrapolate the 12900K out of the 12700K due to the disproportionate change in base turbo (2.4 vs 2.7 and 3.2 vs 3.6) and peak turbo (3.9 vs 3.8 and 5.1 vs 4.9 @ 241W vs 190W).

Base is merely a consequence of TDP. As long as your dissipation is below that (but still high), you can safely presume to stay above base. Base is basically the "minimum guaranteed frequency" for all-core loads. It doesn't control anything and therefore didn't taint their data.

As for turbo, yes their data is limited in the peak frequencies they go up to. 3.8 GHz for the E-cores and 5.0 GHz for the P-cores. It's likely my estimates would show a slightly taller "hump" in the middle of the graph, if I had data for that additional 3.9 GHz step. However, pay close attention to the frequency plots I provided. The E-cores don't even reach 3.8 GHz until package power hits 125 W and 143 W, in the 7zip and x264 cases, respectively. So, that's why I'm confident you would merely see a slight bump up from a little past that point, and later.

Similarly, if the P-cores had more "legs", the graph would merely extend further. However, the efficiency of the P-cores @ high frequencies is so poor that you really wouldn't see much more performance, if it did extend further out.

If you ran the numbers comparing 10p to 8p/4e I imagine it would look very similar to how the estimated 12p shows here in the x264 test.

Could be, but I just included the 12P data to give an idea of how an all-P approach scales.

I really wish chips and cheese had run the efficiency tests on both CPUs to see the difference.

Nah, it's just the upper limit that differs. Maybe the i9 is better-binned and would have a slightly different curve, but I rather doubt it.

Just like how I wish they would do the same on the 13700K and 13900K since those both have the same peak turbo I'd love to see how the scaling and efficiency plays out.

Yes, but they're operating on a shoestring budget. If you go on their Patreon, you can see their cumulative subscription revenue, and it's like not even enough to fund their hardware purchases. They accept donations by other means, but I think Patreon is probably the main one. I think they mainly do it as a hobby and don't mind getting a small subsidy to help out.

In general, I found it very disappointing that all the Raptor Lake reviews I've seen pretty much ignored the E-cores, as if it were a settled matter. For one thing, I'd like to see how much the ring-bus improvements and additional cache helped. Then, there's the fact of how many more there are. And maybe the process node enhancements improved their efficiency curve! Too many questions, and we're basically devoid of data on any of it!
: (

 

[thestryker]

Base is merely a consequence of TDP. As long as your dissipation is below that (but still high), you can safely presume to stay above base. Base is basically the "minimum guaranteed frequency" for all-core loads. It doesn't control anything and therefore didn't taint their data.

The base is a clear indicator that the power usage vs clockspeed is a fair bit different on the i9. The p-core difference surprised me more than the E-core difference as with double the e-cores one would expect those to drop a fair bit to maintain 125W.

As for turbo, yes their data is limited in the peak frequencies they go up to. 3.8 GHz for the E-cores and 5.0 GHz for the P-cores. It's likely my estimates would show a slightly taller "hump" in the middle of the graph, if I had data for that additional 3.9 GHz step. However, pay close attention to the frequency plots I provided. The E-cores don't even reach 3.8 GHz until package power hits 125 W and 143 W, in the 7zip and x264 cases, respectively. So, that's why I'm confident you would merely see a slight bump up from a little past that point, and later.

Similarly, if the P-cores had more "legs", the graph would merely extend further. However, the efficiency of the P-cores @ high frequencies is so poor that you really wouldn't see much more performance, if it did extend further out.

That's my point though raising the e-cores for ADL further is just a straight hit to efficiency as is p-cores, but if you're going straight p-cores you now have more power available to them than with lower SKUs, but not so much that they're running as high inefficiency.

Nah, it's just the upper limit that differs. Maybe the i9 is better-binned and would have a slightly different curve, but I rather doubt it.

I'd be surprised if going from 1 cluster to 2 of e-cores didn't change the curve behavior. Wouldn't be significant of course, but it'd be there and is likely part of the extra boost TDP.

Yes, but they're operating on a shoestring budget. If you go on their Patreon, you can see their cumulative subscription revenue, and it's like not even enough to fund their hardware purchases. They accept donations by other means, but I think Patreon is probably the main one. I think they mainly do it as a hobby and don't mind getting a small subsidy to help out.

Oh yeah definitely they have regular jobs and whatnot.

In general, I found it very disappointing that all the Raptor Lake reviews I've seen pretty much ignored the E-cores, as if it were a settled matter. For one thing, I'd like to see how much the ring-bus improvements and additional cache helped. Then, there's the fact of how many more there are. And maybe the process node enhancements improved their efficiency curve! Too many questions, and we're basically devoid of data on any of it!
: (

Oh absolutely that's bugged the crap out of me because ADL may have brought them but RPL showed why we should all want them. It's crazy to me that optimizations allowed them to increase cache, clocks and double the amount of e-cores at the cost of adding just 12W to the boost turbo. That alone should have warranted a lot of investigation.

 

[bit_user]

The base is a clear indicator that the power usage vs clockspeed is a fair bit different on the i9. The p-core difference surprised me more than the E-core difference as with double the e-cores one would expect those to drop a fair bit to maintain 125W.

Now that we have some actual data, you don't need to be constrained by your expectations.

On the x264 workload, ChipsAndCheese measured a quad E-core cluster at 10.3 W @ 2.7 GHz and about 8.1 W @ 2.4 GHz. That translates to the total budget for E-cores going from 10.3 W to 16.2 W, or a 3.9 W increase.

Granted, ChipsAndCheese is just one datapoint (or two, if you count the 7zip test case), and x264 is likely not as strenuous as what Intel uses to determine their base clocks. Still, it's probably off by only a fairly small scale factor.

That's my point though raising the e-cores for ADL further is just a straight hit to efficiency as is p-cores, but if you're going straight p-cores you now have more power available to them than with lower SKUs, but not so much that they're running as high inefficiency.

You lost me. If you look at my joint clock frequency graphs, the best strategy for clocking the E-cores @ high power follows a saw-tooth pattern. Giving them one extra frequency step just lets the points on those "teeth" go a bit higher, before taking the hit of increasing P-core clocks the next step.

The reason why the E-core clocks are following a saw-tooth is that it's more efficient to run them at max clocks than it is to run the P-cores above 4.4 GHz or so. So, they become a sort of intermezzo between P-core steps.

Oh yeah definitely they have regular jobs and whatnot.

I know at least one is a university student, though I forget which.

It's crazy to me that optimizations allowed them to increase cache, clocks and double the amount of e-cores at the cost of adding just 12W to the boost turbo. That alone should have warranted a lot of investigation.

It would be interesting to see all-core clocks at PL2 power levels, for different workloads. You seem to presume 12 W is all it took for them to hit their max turbo frequencies on all cores, but I don't think Intel ever said that.

It could be that 12 W is all they thought they could get away with raising them, to avoid power problems with some LGA1700 motherboards. We don't really know why they do what they do. We can only speculate, though having good data sure helps!

 

[thestryker]

It would be interesting to see all-core clocks at PL2 power levels, for different workloads. You seem to presume 12 W is all it took for them to hit their max turbo frequencies on all cores, but I don't think Intel ever said that.

It could be that 12 W is all they thought they could get away with raising them, to avoid power problems with some LGA1700 motherboards. We don't really know why they do what they do. We can only speculate, though having good data sure helps!

TPU changed their testing for RPL so there's a lot more useful data, but clocks seemed to be stable at stock (finding reviews that actually used stock turbo for 13900K is such a pain) and they recorded 26W diff between 12900K and 13900K for blender: https://www.techpowerup.com/review/intel-core-i9-13900k/24.html

 

[bit_user]

So, one thing I had the idea to look at was efficiency vs. clockspeed. The reason being that, in order to compare efficiency on the performance vs. power graph, you first have to imagine a line between the origin and a given datapoint. Then, you have to look at which side of it a datapoint on the other series falls. By computing perf/W and plotting that against frequency, you can just look at the Y-axis value to compare the cores' efficiencies at the same or different frequencies.

So, here's the data from 7zip:

​

And here's the same for x264:

​

I think this shows a lot better how you can balance frequencies between the different sets of cores, as you increase or decrease the power target.

 

[Crazyy8]

I'm delaying my build from this summer and waiting for this 14900K Raptor Lake Refresh.
Why not? It might be a little better. I can't wait for Arrow Lake to build (over a year away at least) and Meteor Lake is mobile only.

That aged like fine milk.

 

[Greg7579]

That aged like fine milk.

Yes, the Refresh is not a jump up at the i9 level, but I'm glad I waited because a fantastic new case just hit (Lian Li O11D EVO XL) as well as a refreshed ASUS motherboard (Maximus Z790 Dark Hero). I'm ordering parts now and will build in 2 weeks.