Question Raptor Lake - Official Thread

[Hulk]

Since we already have the first Raptor Lake leak I'm thinking it should have it's own thread.
What do we know so far?
From Anandtech's Intel Process Roadmap articles from July:

Built on Intel 7 with upgraded FinFET
10-15% PPW (performance-per-watt)
Last non-tiled consumer CPU as Meteor Lake will be tiled

I'm guessing this will be a minor update to ADL with just a few microarchitecture changes to the cores. The larger change will be the new process refinement allowing 8+16 at the top of the stack.

Will it work with current z690 motherboards? If yes then that could be a major selling point for people to move to ADL rather than wait.

 

[pakotlar]

If the 12900K loses MT performance the longer the Render it goes, you can bet Rocket lake will lose it sooner than Alder Lake.

Raptor Lake. I doubt Alder Lake loses performance when not experiencing thermal buildup and throttling; could you share more detail about the behaviour you’ve seen and the temps over time?

 

[nicalandia]

Raptor Lake. I doubt Alder Lake loses performance when not experiencing thermal buildup and throttling; could you share more detail about the behaviour you’ve seen and the temps over time?

Pudget System tested Alder Lake with a 10 minute MT CBR23 Render and the performance was lower due to the CPU adjusting for Temp and TDP. The performance went down from 25799 (Windows 11, DDR5-4800 32GB) to 23784 Points.

 

[dullard]

Pudget System tested Alder Lake with a 10 minute MT CBR23 Render and the performance was lower due to the CPU adjusting for Temp and TDP. The performance went down from 25799 (Windows 11, DDR5-4800 32GB) to 23784 Points.

And what happened to the power? Or do you only want to discuss one of the two important details?

 

[nicalandia]

And what happened to the power? Or do you only want to discuss one of the two important details?

Since it was not OC, you can bet that power draw was caped at stock levels which is kinda high and then multiply that by 10 minutes, you can be sure that it an ADL 12900K not only loses to a 5950X on a 10 minute Render, but it will consume much more watts per minute.

 

[Det0x]

Hmm.

Core-for-Core, Golden Cove vs Zen 3 at the same reasonable frequency of 4GHz how do they compare in terms of performance and power?

I have no doubt a 16 core Zen 3 clocked at 4GHz is more efficient than a 12900K clocked within an inch of it's life for equal performance. But that's simply due to transistors/cores.

Not sure about a heads up Zen 3 vs GC comparison of performance and efficiency though.

Why this fascination with 4000mhz exactly ?
Its way too high for Zen3's maximum efficiently range.. Lets just say there is a reason why 64core epyc's dont run allcore workloads at 4ghz..

4ghz would be fine for Zen4, but with Zen3 you should try to stay in the 3.0ghz to 3.6ghz range if your looking for highest possible performance/watt (and this is what the servers also are running)
A stock 5950x dont even hold 4ghz allcore in Cinebench..

BUT i did a test run at 4000mhz/0.875v with 6 cores just to look at the load numbers while running Cinebench:

CPU cores = ~29watt (~4.4w each core @ 4ghz)
SOC = ~11watt
CPU+SOC = ~40watt
PPT = ~54watt
Score = ~10.3k

 

[dullard]

Since it was not OC, you can bet that power draw was caped at stock levels which is kinda high and then multiply that by 10 minutes, you can be sure that it an ADL 12900K not only loses to a 5950X on a 10 minute Render, but it will consume much more watts per minute.

What I meant was, when it is thermally limited, the frequency goes down. So both performance and power consumption drop. Mentioning one without the other is misleading.

 

[nicalandia]

What I meant was, when it is thermally limited, the frequency goes down. So both performance and power consumption drop. Mentioning one without the other is misleading.

True, Also AMD CPUs performance go down, but not as much as Alder lake, for example the 5950WX goes from 24,635 to 24,255 points at 10 minutes which is higher than Alder lake 10 minute 23,784 Points

 

[Markfw]

True, Also AMD CPUs performance go down, but not as much as Alder lake, for example the 5950WX goes from 24,635 to 24,255 points at 10 minutes which is higher than Alder lake 10 minute 23,784 Points

This is exactly what I see with my own Alder lake and 5950x. When fully loaded, the 5950x stays at 3.7 ghz,142 watt. My 12700F is beaten consistently at 2x the performance at about 162 watt, what it seems to draw under constant 100% load. Now that is 8c/16t since my e-cores are disabled, so, the P-cores are about even with the 5950x, but at a higher power level. And considering its only 8 cores instead of 16, thats even higher yet.

I suspect Raptor lake will use even more power.

 

[scannall]

They are more efficient than Zen3 cores at the same clocks and when both at their minimum voltages.

For example I did test 12600K with only 6P cores enabled at 4ghz and 0,835Vcore and compared it with someone's 5600X at 4ghz 0,95V.

The result
6P 4ghz 0,835Vcore 49W package power 12425MT CBR23
6 Zen3 cores 4ghz 0,95Vcore 56W package power 10200MT CB R23

To even match my score with 5600X you would need 4,7ghz+ all core oc and 100W package power

I have my Ryzen 5800 set in eco mode. Cinebench 23 gets 13593 drawing 59.8 watts.

 

[Hulk]

Why this fascination with 4000mhz exactly ?
Its way too high for Zen3's maximum efficiently range.. Lets just say there is a reason why 64core epyc's dont run allcore workloads at 4ghz..

4ghz would be fine for Zen4, but with Zen3 you should try to stay in the 3.0ghz to 3.6ghz range if your looking for highest possible performance/watt (and this is what the servers also are running)
A stock 5950x dont even hold 4ghz allcore in Cinebench..

BUT i did a test run at 4000mhz/0.875v with 6 cores just to look at the load numbers while running Cinebench:

CPU cores = ~29watt (~4.4w each core @ 4ghz)
SOC = ~11watt
CPU+SOC = ~40watt
PPT = ~54watt
Score = ~10.3k
View attachment 64460

3.5GHz, 3.0GHz, whatever.
I'm sorry I posted. I just have a hard time allowing misleading generalizations to go unchallenged.

 

[pakotlar]

I have my Ryzen 5800 set in eco mode. Cinebench 23 gets 13593 drawing 59.8 watts.

8 cores are probably better able to get good PPW than 6 (can afford lower clocks, lower voltage, so quadratic reduction in power)

 

[Schmide]

quadratic reduction in power

I don't think quadratic is the word you're looking for.

 

[Henry swagger]

Impressive from Intel to get 5.8ghz on the same node.. the cobalt in the node can take high frequencies without to big heat

 

[Markfw]

Impressive from Intel to get 5.8ghz on the same node.. the cobalt in the node can take high frequencies without to big heat

Where did you see that a Raptor lake did 5.8 ghz ? And was this under LN2 ?

 

[uzzi38]

Where did you see that a Raptor lake did 5.8 ghz ? And was this under LN2 ?

It seems to be with something called eTVB. We don't know yet what conditions eTVB includes, but regular TVB was mostly temperature based. E.g. if the main core is under 90c then +100MHz, if it's under 70c then another 100MHz etc etc.

The actual temperatures and number of stages vary by product.

 

[uzzi38]

They are more efficient than Zen3 cores at the same clocks and when both at their minimum voltages.

For example I did test 12600K with only 6P cores enabled at 4ghz and 0,835Vcore and compared it with someone's 5600X at 4ghz 0,95V.

The result
6P 4ghz 0,835Vcore 49W package power 12425MT CBR23
6 Zen3 cores 4ghz 0,95Vcore 56W package power 10200MT CB R23

To even match my score with 5600X you would need 4,7ghz+ all core oc and 100W package power

That depends on the clocks you run them at. There's a crossover point where Golden Cove takes over on efficiency, as for where that is idk but it's somewhere in the 3 - 3.5GHz range. Below that Zen 3 is more power efficient.

Also, locking voltages is a meaningless test as the chips wouldn't ever ship like that. You should reduce PL2/PPT to a specific level to test this instead. Let the default boost algorithms determine how the chips will perform, not silicon lottery.

 

[dullard]

Where did you see that a Raptor lake did 5.8 ghz ? And was this under LN2 ?

Here is an article about a 5.8 GHz video (two cores supposedly reach that speed). I didn't watch the video, so I do not know the cooling method.

https://www.tomshardware.com/news/intel-core-i9-13900k-benchmarked-in-chinese-video-delivers-10-35-perf-increase

5.8 GHz has been the rumor for months, although yes there are other rumors so it is impossible to know at this point which is true. https://www.pcgamer.com/intels-flagship-raptor-lake-cpu-could-run-at-up-to-58ghz/

I suspect Raptor lake will use even more power.

In the image above you see the rumored power levels. There is a 12 W PL2 increase. So, in most cases it will use more power. The very temporary max PL4 power level is reduced.

 

[Henry swagger]

Here is an article about a 5.8 GHz video (two cores supposedly reach that speed). I didn't watch the video, so I do not know the cooling method.

https://www.tomshardware.com/news/intel-core-i9-13900k-benchmarked-in-chinese-video-delivers-10-35-perf-increase

View attachment 64500

5.8 GHz has been the rumor for months, although yes there are other rumors so it is impossible to know at this point which is true. https://www.pcgamer.com/intels-flagship-raptor-lake-cpu-could-run-at-up-to-58ghz/


In the image above you see the rumored power levels. There is a 12 W PL2 increase. So, in most cases it will use more power. The very temporary max PL4 power level is reduced.

Mayb intel are using dvlr to keep the power down.. raptor lake has 8 more cores but 12w increase in pl2 🤔✅

 

[dullard]

Mayb intel are using dvlr to keep the power down.. raptor lake has 8 more cores but 12w increase in pl2 🤔✅

We will find out soon. Raptor Lake might be announced at the Sept 27/28 Innovation Event. https://www.intel.com/content/www/us/en/events/on-event-series/innovation.html

A 5% increase in power for a 10% to 35% increase in performance in that video is a good step in the right direction. Intel needs more E cores or a better node to really help on the power/performance metric. While that is an important metric to many users, it is not the only important metric. I do personally like the route of P and E cores so that you can eventually get the best of both worlds: snappy single threaded performance and good multi threaded performance.

 

[pakotlar]

I don't think quadratic is the word you're looking for.

watts = volts² / ohms

In an 8 core processor, each core can be fed lower voltage to achieve iso performance in MT workloads (than in the 6 core version, assuming same binning), which will lead to a quadratic reduction in power usage.

 

[pakotlar]

Where did you see that a Raptor lake did 5.8 ghz ? And was this under LN2 ?

That’s the rumor, and ES3 already hits 5.7.

 

[dullard]

watts = volts² / ohms

In an 8 core processor, each core can be fed lower voltage to achieve iso performance in MT workloads (than in the 6 core version, assuming same binning), which will lead to a quadratic reduction in power usage.

Squared is not correct (thus quadratic is incorrect). CPU power is dominated by the dynamic term Pdynamic = C * V^2 * f where C is the capacitance, V is the voltage, and f is the frequency of the CPU. https://en.wikipedia.org/wiki/Processor_power_dissipation#Sources The key that you are missing is that frequency is roughly proportional to voltage. When overclocking you often turn up the voltage to be able to hit higher frequencies. So in reality, CPU power is proportional to V^3. Try the word "cubed" instead. But also more cores means a physically larger CPU so the capacitance goes up linearly. The end result is a mixture of linear and cubic effects.

Your point is correct though. More cores -> Lower frequency per core -> less power used for iso-performance in multi-threaded tasks.

 

[pakotlar]

Squared is not correct (thus quadratic is incorrect). CPU power is dominated by the dynamic term Pdynamic = C * V^2 * f where C is the capacitance, V is the voltage, and f is the frequency of the CPU. https://en.wikipedia.org/wiki/Processor_power_dissipation#Sources The key that you are missing is that frequency is roughly proportional to voltage. When overclocking you often turn up the voltage to be able to hit higher frequencies. So in reality, CPU power is proportional to V^3. Try the word "cubed" instead. But also more cores means a physically larger CPU so the capacitance goes up linearly. The end result is a mixture of linear and cubic effects.

Your point is correct though. More cores -> Lower frequency per core -> less power used for iso-performance in multi-threaded tasks.

Good point about ignoring frequency, but that will have a small effect. The overall relationship will be nearly quadratic. Btw, the relationship between voltage and frequency is only linear in a regime well below what the consumer chips are clocked at. Above that frequency, the relationship is exponential. For Ryzen that threshold is somewhere below 4Ghz.

So it’s worse than quadratic (which FYI is commonly used to refer to square relationships, i.e quadratic big O complexity is N^2), but it won’t be cubic, because the voltage effect is larger than the frequency effect. Between linear and cubic, but it will be much closer to quadratic than linear or cubic, except for small changes in the parameters, I think. Also, don’t appreciate your condescension (try the word cubic).

You’ll notice that, assuming 2 different voltages, the relationship is nearly quadratic (we are talking about the ratio of power):

V1^2*Xf / V2^2*f is closely approximated by V1^2/V2^2 when X is small relative to V1^2/V2^2, which it will be, especially above the elbow in the power vs frequency curve.



Another way of writing that is (V1/V2)^2 * X. In a big O sense, this is N^2, we ignore linear terms, though you’re right they are important.

 

[dullard]

So it’s worse than quadratic (which FYI is commonly used to refer to square relationships, i.e quadratic big O complexity is N^2), but it won’t be cubic, because the voltage effect is larger than the frequency effect. Don’t appreciate your condescension (try the word cubic).

In that region of your graph, you could fit voltage to frequency with a quadratic term. Thus, in that region power is proportional to V^4. Try quartic? [Note: quartic is only valid in that high frequency region of the graph]

Pdynamic = C * V^2 * f

f ∝ V^2 in the region of your graph, thus

Pdynamic ∝ C * V^4

 

[pakotlar]

In that region of your graph, you could fit voltage to frequency with a quadratic term. Thus, in that region power is proportional to V^4. Try quartic?

It’s not V^4. Read my edited post.

edit: I see your edit. The relationship is (V1/V2)^2, roughly. F1 and F2 doesn’t matter a whole lot here, except in its effect on V1 and V2, you now see that. Call V1 V2^N, then you have ~ V2^2N, in that exponential regime.