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.

 

[Carfax83]

Latest computerbase.de review pitting the 13900K against the 7950x in gaming using an RTX 4099 with several new titles:

High-end benchmark duel: Ryzen 9 7950X and Core i9-13900K play on RTX 4090

 

[Hitman928]

That review is totally GPU bottlenecked. As I and many others have said before, you can't take 13th gen or Zen 4 game performance reviews with anything less than an RTX 4090 seriously. They are just too bottlenecked.

Just look how bunched up those scores are.

It’s 1440p, not sure why he didn’t post the 1080p results which are more interesting.

 

[Hitman928]

Latest computerbase.de review pitting the 13900K against the 7950x in gaming using an RTX 4099 with several new titles:

High-end benchmark duel: Ryzen 9 7950X and Core i9-13900K play on RTX 4090

It’s still with just officially supported memory speed and timings for both platforms, so these results are not that interesting either.

 

[igor_kavinski]

It’s 1440p, not sure why he didn’t post the 1080p results which are more interesting.

Happy?

 

[igor_kavinski]

A bit disappointing that they did not upgrade the IMC for DDR4. Not that 4000 MT/s is bad, but still.

The latency advantage with DDR4-4000 CL15 is signficant. Should translate to better game performance in latency sensitive game engines.

 

[In2Photos]

Latest computerbase.de review pitting the 13900K against the 7950x in gaming using an RTX 4099 with several new titles:

High-end benchmark duel: Ryzen 9 7950X and Core i9-13900K play on RTX 4090

I think we found where Harry's 9 key went!

 

[nicalandia]

Look, if you have a 4090 You will not skimp on RAM so.... Go Full Raptor, 4090 and get 7200 DDR5 Sticks.

 

[Grimnir]

Look, if you have a 4090 You will not skimp on RAM so.... Go Full Raptor, 4090 and get 7200 DDR5 Sticks.

That's a big if...

 

[ondma]

I agree, especially with the quoted part. I might have posted this before, but I too keep Excel calculations with various scenarios going. See the image below. This particular one is just a simple take from Chips And Cheese's analysis of efficiency (https://i0.wp.com/chipsandcheese.com/wp-content/uploads/2022/01/image-151.png?ssl=1). I took a major assumption that isn't quite correct, that 4 E cores and 1 P core were the same area. But, it is pretty close to accurate.
View attachment 70252

In all power cases, given roughly equal area, you want as many E cores as possible. I hear a ton of people wanting an all P-core CPU and you can see that it would absolutely suck on Intel's node for anything multithreaded. This is pretty much the same as your conclusion if you look at your 164W data, the ones with the most E cores win in every single grouping.

But, there is a big tradeoff that you mention: an all E core CPU sucks at single threaded performance. Which, despite the lack of attention it gets in reviews, ST is what most people use in their day-to-day tasks. So, a hybrid approach given Intel's limitations is the best move that they have. Once we get to 32+ E cores it will be all very clear. Much more clear than the 8+4 ugly stepchild that we have in some Intel chips now.

Wont they eventually need more than 8 P cores though? Perhaps for gaming, as gpus get stronger and stronger? I understand the hybrid approach, and it seems to have worked so far to compensate for the node and core size disadvantages, but 32 E cores? I would hope that eventually with a tiled approach and better process, they can increase P cores as well. Just a wild idea, but if MTL has a 6P 8E tile, how about putting 2 of those together for 12P/16E?

 

[Grimnir]

View attachment 70265

The latency advantage with DDR4-4000 CL15 is signficant. Should translate to better game performance in latency sensitive game engines.

Aye, the lower latency is nice.

Either way I would expect a pretty big upgrade in CPU performance, from my 8086K at 5Ghz. I frequently run into some CPU limitations, so it would be a very noticeable upgrade.

 

[igor_kavinski]

Look, if you have a 4090 You will not skimp on RAM so.... Go Full Raptor, 4090 and get 7200 DDR5 Sticks.

The risk of house burning down is just a bonus

 

[LightningZ71]

Wont they eventually need more than 8 P cores though? Perhaps for gaming, as gpus get stronger and stronger? I understand the hybrid approach, and it seems to have worked so far to compensate for the node and core size disadvantages, but 32 E cores? I would hope that eventually with a tiled approach and better process, they can increase P cores as well. Just a wild idea, but if MTL has a 6P 8E tile, how about putting 2 of those together for 12P/16E?

In my time of following game development and benchmarks, it is highly typical that games becomes performance constrained on roughly four threads or less. This is why, in most reviews, you see very little clock normalized performance difference between 6 and 8 core processors with HT. Typically, when you find a notable difference, it's often because the 8 core part on Intel has more L3 cache. This is also why you can almost always get PLAYABLE frame rates on 4/8 processors with a decent video card in most games. Those quad core processors typically come with a notable frequency deficit against the 6 and 8 core parts and half the L3 cache, so there's a lot going on to hurt their performance, but having 4 real cores is usually enough to keep things. This is also why you find that the dual core celerons and pentiums have been taking a real beating in the benchmarks and are usually just stutterfests to use in real life. They often don't have markedly less L3 than the quad cores, but having to swap multiple main threads it a back breaker.

All of that to say that a 6P/24E processor with other modest improvements, more L3, and higher clocks would likely have even better game performance than the 13900K does, and is why I don't doubt that meteor lake will perform quite well.

 

[Tuna-Fish]

Has anyone measured the L3 latency of a 13900k? For some reason I'm having a hard time finding it. Possibly I'm just dumb.

 

[Carfax83]

It’s 1440p, not sure why he didn’t post the 1080p results which are more interesting.

I honestly don't think there's much of a difference. The 13900K is known to be bottlenecked by an RTX 4090 at 1080p high settings in many games, so you can imagine how much of a limitation the RTX 3090 is.

To mitigate the GPU bottleneck, he should have tested at 720p low settings, but then I'm sure he would have gotten complaints from less knowledgeable folk who don't understand how bottlenecking works.

 

[Carfax83]

It’s still with just officially supported memory speed and timings for both platforms, so these results are not that interesting either.

I agree that higher speed memory would have been better, but these results are still valid and computerbase's test methodology is always top notch.

It just goes to show that when you add an RTX 4090 into the equation, the 13900K takes off like a rocket; especially in games that utilize the CPU heavily. AMD is definitely going to need the Zen 4 3D parts to take the gaming crown from Intel.

 

[nicalandia]

Has anyone measured the L3 latency of a 13900k? For some reason I'm having a hard time finding it. Possibly I'm just dumb.

I got this from the Web

 

[Carfax83]

View attachment 70265

The latency advantage with DDR4-4000 CL15 is signficant. Should translate to better game performance in latency sensitive game engines.

DDR5 has latency mitigation tech though which allows for superior memory parallelism compared to DDR4.

Speaking of DDR5, I picked up my DDR5 7200 today! My motherboard is supposed to arrive on Saturday, but I probably won't be able to install everything until Tuesday next week.

Can't say I'm looking forward to it. It's going to take hours

 

[ondma]

I honestly don't think there's much of a difference. The 13900K is known to be bottlenecked by an RTX 4090 at 1080p high settings in many games, so you can imagine how much of a limitation the RTX 3090 is.

To mitigate the GPU bottleneck, he should have tested at 720p low settings, but then I'm sure he would have gotten complaints from less knowledgeable folk who don't understand how bottlenecking works.

I understand how bottlenecking works, but I am not sure a cpu bound scenario at 720p is a reliable predictor of how a given cpu will perform at higher resolution with a more powerful card.

 

[nicalandia]

I understand how bottlenecking works, but I am not sure a cpu bound scenario at 720p is a reliable predictor of how a given cpu will perform at higher resolution with a more powerful card.

720 is still too high of a resolution to test for Bottle-neck

480P is the absolute best scenario, in that regard AnandTech is the only(as far as I am aware) review site doing such tests.

 

[igor_kavinski]

Can't say I'm looking forward to it. It's going to take hours

Because it's been years since you built a PC? I was afraid that I would mess things up too when checking the i5-12400 with an ASUS mobo. Thankfully, it booted on first try. But the 40+ seconds it took to display the screen was unnerving. My previous build was an i7-5775C maybe 6 months before and before that, an i3-2100 in 2013 (upgraded to a used i7-3770 in 2019 and had the seller install it on the same mobo).

 

[igor_kavinski]

Either way I would expect a pretty big upgrade in CPU performance, from my 8086K at 5Ghz.

Don't sell it. It might be worth a lot someday.

 

[Grimnir]

Don't sell it. It might be worth a lot someday.

You think it would turn into a collector's item? I'm not so sure, but I won't be selling it. The plan is to use it in my home server, replacing an ancient i7 950.

 

[itsmydamnation]

View attachment 70265

The latency advantage with DDR4-4000 CL15 is signficant. Should translate to better game performance in latency sensitive game engines.

you cant actually view it that way because you can have more concurrency with DDR5 , so 4 queued memory requests with DDR4 = 19ns , 4 with DD% = 7.5ns.

 

[In2Photos]

You think it would turn into a collector's item? I'm not so sure, but I won't be selling it. The plan is to use it in my home server, replacing an ancient i7 950.

Careful now, my daily driver is an i7-920.

 

[Hulk]

I agree, especially with the quoted part. I might have posted this before, but I too keep Excel calculations with various scenarios going. See the image below. This particular one is just a simple take from Chips And Cheese's analysis of efficiency (https://i0.wp.com/chipsandcheese.com/wp-content/uploads/2022/01/image-151.png?ssl=1). I took a major assumption that isn't quite correct, that 4 E cores and 1 P core were the same area. But, it is pretty close to accurate.
View attachment 70252

In all power cases, given roughly equal area, you want as many E cores as possible. I hear a ton of people wanting an all P-core CPU and you can see that it would absolutely suck on Intel's node for anything multithreaded. This is pretty much the same as your conclusion if you look at your 164W data, the ones with the most E cores win in every single grouping.

But, there is a big tradeoff that you mention: an all E core CPU sucks at single threaded performance. Which, despite the lack of attention it gets in reviews, ST is what most people use in their day-to-day tasks. So, a hybrid approach given Intel's limitations is the best move that they have. Once we get to 32+ E cores it will be all very clear. Much more clear than the 8+4 ugly stepchild that we have in some Intel chips now.

I'm trying to make a valid P vs. E power consumption analysis but the E's are slippery. If I set the P's to 5GHz and the E's to 4GHz and then test with 16 and 4 E's activated it should be a simple matter to use subtraction of the two power levels and divide by 12. All well and good except that if you go with 0 and 4 E's you get a different result. And another different result with 4 E's and 8's. It's confounding.

I do know that an E at 4GHz uses somewhere between 6 and 8 Watts. But those E are like electrons and seems to follow the laws of quantum mechanics and just can't be accurately measured if you isolate them! They seem to behavior differently in different situations. They are their own little double slit experiment.