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.

 

[nicalandia]

I measured two more power limits of 130 and 210W.

View attachment 69233

Compared to 250W, at 130W at nearly half the power draw you lose just 17,6% performance. That is the point at which you could end up if you wanted the chip to run very efficiently but at the same time you wanted high performance.

Here are the decreases in power and performance compared to 250W:

View attachment 69235

Thanks for sharing.

 

[RTX2080]

I thought Raptor was released but I just realize it is not LOL. So this might be a news for this thread. 13600k/13700k being tested with 7700X:

https://twitter.com/x/status/1581294442475966465

 

[Thunder 57]

This "node behind" stuff in relation to AMD is getting very tired. Raphael is more than just the compute chiplet(s) and that "more" is on the same class of node as Intel. This will only get worse with the release of something like Meteor Lake, with an even bigger smorgasbord of nodes in it. Or Raphael-X for that matter, SRAM stack ain't "5nm class" either AFAIK.

Core size v core size is more relevant, at least to most and in the context of roughly similar performance (otherwise what's the point?). And last I checked, TSMC N5 ain't 2x as dense as Intel 7... not even close. And AMD can fit in more than just 8 of 'em.

Yea, funny how back in the day no one cared that Intel usually had a node advantage over AMD.

 

[DrMrLordX]

Yea, funny how back in the day no one cared that Intel usually had a node advantage over AMD.

They did once AMD stalled out on 32nm SOI. 28nm didn't work out for GF all that well. AMD was running on fumes until GF licensed 14nm from Samsung.

 

[Thunder 57]

They did once AMD stalled out on 32nm SOI. 28nm didn't work out for GF all that well. AMD was running on fumes until GF licensed 14nm from Samsung.

I was talking about earlier like the K7/K8 days. Nobody mentioned nodes when bulldozer was around, it just sucked and that was it.

 

[DrMrLordX]

I was talking about earlier like the K7/K8 days. Nobody mentioned nodes when bulldozer was around, it just sucked and that was it.

Ah okay. In general it seems like your average forum dweller didn't know as much about process nodes until AMD hit a brick wall.

 

[Hulk]

I measured two more power limits of 130 and 210W.

View attachment 69233

Compared to 250W, at 130W at nearly half the power draw you lose just 17,6% performance. That is the point at which you could end up if you wanted the chip to run very efficiently but at the same time you wanted high performance.

Here are the decreases in power and performance compared to 250W:

View attachment 69235

I realize that you're probably "benchmarked out" but if you have the time to do one more I'd really like to see how the 13900K does on our Handbrake benchmark found here: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...tem-new-benchmark-criteria.2588294/?view=date

We have a 7950X score on top right now. All of the instructions are in my first post of the thread as well as full results for everyone here that has run the test in the 2nd post.

 

[Hulk]

Anybody want to run Cinebench R23 at these same power levels and report the scores? I'll add that data to the chart and we can actually compare efficiency in a very specific manner.



This next chart is interesting. As you can see the 13900K is most efficient at 35W. Would efficiency decrease at some point if power was reduced further?

The point here I think is that given approximately equal architectures and process technology it's the number of transistors doing the work that is most important if you want to achieve really high efficiency. And of course transistors are expensive, which is why we stress over die sizes.

Imagine 4 13900K's on one substrate, running 140W total with a score of 65,000+. Suddenly Raptor Lake would be very efficient even on Intel 7. Of course that is a magical part because it would be very inefficient when it comes to economics! It is of course a balancing act for AMD and Intel.

One huge advantage of a smaller process is the ability to more cost effectively put more transistors down, which leads to high efficiency, especially in "ridiculously" multithreaded tasks.

How many transistors does the 13900K contain vs the 7950X?

 

[Kocicak]

I realize that you're probably "benchmarked out" but if you have the time to do one more ...

I am now limited by 210W heat output - I have a small air cooler on the CPU now. If you think that the CPU will not overheat if I run that benchmark, I could try it on higher power limit than 210W. How long does the benchmark last?

 

[Hulk]

I am now limited to 210W heat output - I have a small air cooler on the CPU now. If you think that the CPU will not overheat if I run that benchmark, I could try it on higher power limit than 210W. How long does the benchmark last?

Fastest score we have is 91 seconds. I don't think Handbrake will load up all the cores 100% so thermals might not be an issue.

 

[mikk]

Raichu gets 38.7k with newest Bios at 253W, it's roughly 2k more than 1.5 months ago.

https://twitter.com/x/status/1581636997994586114

 

[Kocicak]

The CPU has some minimal power draw so that it can function properly. The curve probably looks like I added. Maximum of the ratio of both variables can be found like I indicated. So the point of maximal efficiency could be between 15 and 20W.




And the other graph looks like this:

 

[Hulk]

The CPU has some minimal power draw so that it can function properly. The curve probably looks like I added. Maximum of the ratio of both variables can be found like I indicated. So the point of maximal efficiency could be between 15 and 20W.


View attachment 69251

And the other graph look like this:

View attachment 69253

Anyone want to test their 7950X at the same power levels for a heads up comparison?

 

[Doug S]

The CPU has some minimal power draw so that it can function properly. The curve probably looks like I added. Maximum of the ratio of both variables can be found like I indicated. So the point of maximal efficiency could be between 15 and 20W.

However much power it takes to operate all the E cores at the minimum clock rate they are capable of, plus the uncore overhead, would be that figure.

 

[BorisTheBlade82]

The CPU has some minimal power draw so that it can function properly. The curve probably looks like I added. Maximum of the ratio of both variables can be found like I indicated. So the point of maximal efficiency could be between 15 and 20W.


View attachment 69251

And the other graph look like this:

View attachment 69253

I did something similar with my 4700U Renoir a while ago. The pattern is similar.

12w is the efficiency Sweet Spot. Below it gets worse.

 

[Det0x]

Anybody want to run Cinebench R23 at these same power levels and report the scores? I'll add that data to the chart and we can actually compare efficiency in a very specific manner.

Think i can do it later tonight if i get the time

 

[Kocicak]

First I make the CPU to run a benchmark with insufficient air cooler on it and the CPU bakes. Then I want it to run with so little juice, that it refuses and asks for more. Honestly I think that this CPU thinks that I am a complete idiot and it would like to leave the socket in my computer, if it could.



Anyway, it seems that the minimal possible power draw at 800 MHz is the most efficient, nothing below that is possible, at least with 100% load. I will not experiment with smaller loads and power draws.

BTW the computer is not snappy anymore at 8W.

 

[ZGR]

First I make the CPU to run a benchmark with insufficient air cooler on it and the CPU bakes.

Have you tried installing Intel XTU and add negative 50mV increments; and maybe 100 MHz less turbo until you find the right settings for your cooler? Would be cool to see if your CPU can do -50mV or -100mV at stock clocks.
I think your air cooler is doing quite well considering the power draw.

I've linked this video before but in a different thread. Ignore the title since the timestamp is for 12900k undervolting. (TLDW - 12900k is at 4.2 GHz P cores, 3.0 GHz on E cores at ~0.9v competing very well against 5800X3D in performance AND power):

 

[Det0x]

Anybody want to run Cinebench R23 at these same power levels and report the scores? I'll add that data to the chart and we can actually compare efficiency in a very specific manner.

Alrighty, i have completed gathering the data for a tuned 7950x
4800MT/s memory with auto timings were used for all the runs.
SOC draw alone is around ~10w (important for the lower PPT runs)

PPT 250w, 160A TDC and 225A EDC limit = 40360 points (OBS: CPU only pulled 230w maximum)


PPT 210w, 140A TDC and 200A EDC limit = 40085 points (OBS: CPU only pulled 204w maximum)


PPT 160w, 115A TDC and 175A EDC limit = 39068 points


PPT 130w, 100A TDC and 160A EDC limit = 37344 points


PPT 100w, 80A TDC and 150A EDC limit = 34574 points


PPT 65w, 70A TDC and 130A EDC limit = 28387 points


PPT 50w, 65A TDC and 125A EDC limit = 22832 points


PPT 35w, 60A TDC and 120A EDC limit = 144667 points


PPT 25w, 40A TDC and 80A EDC limit = 4180 points (cores were running at a wooping 500mhz each while rendering cinebench lol)


At the lower end of the PPT scale the 7950x is limited by the chiplet design with SOC on a separate die drawing ~10w alone
At PPT limit 65w the cores had ~3.44w each
At PPT limit 50w the cores had ~2.5w each
At PPT limit 35w the cores had ~1.56w each
At PPT limit 25w the cores had ~0.93w each

A monolithic APU should be much better suited here..

*edit*
added 50w PPT run

 

[Hulk]

 

[Markfw]

View attachment 69285

Yes, 7950x is faster at the same wattage, it just won't go over 230 for any reason (all-core)

 

[poke01]

Yes, 7950x is faster at the same wattage, it just won't go over 230 for any reason (all-core)

I would be very shocked if it wasn't.

 

[LightningZ71]

Alrighty, i have completed gathering the data for a tuned 7950x
4800MT/s memory with auto timings were used for all the runs.
SOC draw alone is around ~10w (important for the lower PPT runs)

PPT 250w, 160A TDC and 225A EDC limit = 40360 points (OBS: CPU only pulled 230w maximum)
View attachment 69276

PPT 210w, 140A TDC and 200A EDC limit = 40085 points (OBS: CPU only pulled 204w maximum)
View attachment 69277

PPT 160w, 115A TDC and 175A EDC limit = 39068 points
View attachment 69278

PPT 130w, 100A TDC and 160A EDC limit = 37344 points
View attachment 69279

PPT 100w, 80A TDC and 150A EDC limit = 34574 points
View attachment 69280

PPT 65w, 70A TDC and 130A EDC limit = 28387 points
View attachment 69281

PPT 50w, 65A TDC and 125A EDC limit = 22832 points
View attachment 69284

PPT 35w, 60A TDC and 120A EDC limit = 144667 points
View attachment 69282

PPT 25w, 40A TDC and 80A EDC limit = 4180 points (cores were running at a wooping 500mhz each while rendering cinebench lol)
View attachment 69283

At the lower end of the PPT scale the 7950x is limited by the chiplet design with SOC on a separate die drawing ~10w alone
At PPT limit 65w the cores had ~3.44w each
At PPT limit 50w the cores had ~2.5w each
At PPT limit 35w the cores had ~1.56w each
At PPT limit 25w the cores had ~0.93w each

A monolithic APU should be much better suited here..

*edit*
added 50w PPT run

At some point, it probably makes more sense to entirely shut off one CCD and channel the remaining power to the remaining one.

 

[Hulk]

View attachment 69285

If anyone with a 5950X or 12900K wants to benchmark these power points I'll add those CPU's so this graph as well. I might be interesting to see current and past flagship parts on one chart.

 

[Hulk]

Based on the two tests from members of this forum here is a comparison of Cinebench R23 MT scores at various power levels. Except for 35W Zen 4 is as much as 17% more efficient than this Raptor Lake sample at all other data points. By 17% I mean it has 17% more throughput while using the same power. Of course if I included Raptor Lake power levels up to 350W the margins for Zen 4 would be even larger but since Zen 4 really doesn't use more than 210 or so Watts and both parts seem to top out around 40,000 in CB R23 that calculation would have to be made using some power points like 210 W vs 350 W in which, of course Zen 4 would be wildly more efficient than Raptor.

But in my opinion after about 210 W Raptor becomes increasingly nonlinear in terms of what you get out for what you put in and the only reason to push that much power into the CPU would be to attain a score to call "victory" over Zen 4. Then we get into the arguments of how inefficient Raptor is from the Zen guys, and the Raptor guys say yeah but it's still a win.. blah, blah, blah.

If we are to assume the data from these CPU's is representative then at least for this benchmark there is no denying that Zen 4 is 9 to 17% more efficient than Raptor at what I would call sane power levels for Raptor Lake. Yes, Zen 4 is as fast as Raptor Lake in this test and more efficient.

I would be interested in seeing how these same results for the ST bench because as we know not a lot of actual software is as parallel as Cinebench. Still, it's an impressive result from Zen 4 and not as bad a performance from Raptor as I had expected given all of the doom and gloom I've been reading. I don't think the Raptors will be causing large scale blackouts around the world and I think most people would be wise to cap them at perhaps 230W or so as there doesn't seem to be much gained after that point anyway except for posting numbers.