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.

 

[Kocicak]

Wow, GB Biscuits are a mess. Asus SP rating probably as well.

 

[Hulk]

Please don't respond to this if you believe the SP rating is an imaginary number assigned to processors randomly by motherboard manufacturers.

But if you do think the SP rating is computed from the VID data in a processor, which itself is related to the quality of the silicon as far as overclocking, undervolting, etc.. then have a look at this chart. Using the Igor's Lab data I broke the Raptor Lake out by series.

If you have an idea or theory as to why all of the highly rated parts are 13900 series I'd love to hear it. Again, if you don't know please don't berate the question.

It seems as though there should be a fair number of highly rated parts in the 13600/13700 series that simply have defects in the cores that make them unusable thus bumping them down from 13900's. Otherwise I'm wondering perhaps if there is a correlation between defective cores in a die and the quality of the die in general, which would account for the fully operational 8+16 dies having the highest ratings.

 

[jpiniero]

Unlocked parts, I'm guessing.

 

[Just Benching]

View attachment 77723

Please don't respond to this if you believe the SP rating is an imaginary number assigned to processors randomly by motherboard manufacturers.

But if you do think the SP rating is computed from the VID data in a processor, which itself is related to the quality of the silicon as far as overclocking, undervolting, etc.. then have a look at this chart. Using the Igor's Lab data I broke the Raptor Lake out by series.

If you have an idea or theory as to why all of the highly rated parts are 13900 series I'd love to hear it. Again, if you don't know please don't berate the question.

It seems as though there should be a fair number of highly rated parts in the 13600/13700 series that simply have defects in the cores that make them unusable thus bumping them down from 13900's. Otherwise I'm wondering perhaps if there is a correlation between defective cores in a die and the quality of the die in general, which would account for the fully operational 8+16 dies having the highest ratings.

It has to do with wafer quality. Or what we call batch. A very high quality batch usually doesn't have defective parts, so no CPUs that are going to turn into 13600/ 13700 etc. A bad quality batch might have a mix of defective cores and low SP ratings.

SP is definitely not random. It's based on the VF curve.

 

[Hulk]

It has to do with wafer quality. Or what we call batch. A very high quality batch usually doesn't have defective parts, so no CPUs that are going to turn into 13600/ 13700 etc. A bad quality batch might have a mix of defective cores and low SP ratings.

SP is definitely not random. It's based on the VF curve.

That makes a lot of sense and explains the data very succinctly.

 

[Hitman928]

It has to do with wafer quality. Or what we call batch. A very high quality batch usually doesn't have defective parts, so no CPUs that are going to turn into 13600/ 13700 etc. A bad quality batch might have a mix of defective cores and low SP ratings.

I'm not sure that there is such a strong correlation between parametric yield and defect density as you are suggesting. . .

 

[Kocicak]

... SP is definitely not random. It's based on the VF curve.

So why not to directly reveal voltages for e.g. 3 frequencies typical for each model as 5200, 5500 and 5800 MHz for 13900K. Not that 5200 is typical in any way, but it has the same step as the other two frequencies.

For example string 5,2-1,20; 5,5-1,35; 5,8-1,45 is a little bit longer than 2 digit score, but it is real data.

 

[Hulk]

13900KS seem to be sold out at all Microcents around me, and I have a lot of them within 100 miles. Not that I'm looking to buy but just curious.

Is the 13900KS simply as rare as hens teeth or is it selling that well?

At first I was thinking what a ripoff. But after looking at those SP ratings it's kind of an ok deal if you want a guaranteed good rated chip.

 

[Hitman928]

13900KS seem to be sold out at all Microcents around me, and I have a lot of them within 100 miles. Not that I'm looking to buy but just curious.

Is the 13900KS simply as rare as hens teeth or is it selling that well?

At first I was thinking what a ripoff. But after looking at those SP ratings it's kind of an ok deal if you want a guaranteed good rated chip.

Since it is a specialty top tier binned chip, I'm guessing the supply is really low. It's just a guess, though.

 

[Mopetar]

I'm not sure that there is such a strong correlation between parametric yield and defect density as you are suggesting. . .

Defects can be somewhat separate from other quality measures which don't make the silicon defective, but would relegate it to a lower bin.

I had t really thought about it before, but there is a certain amount of sense to the idea that if you've just diced your wafer and are going through the binning process that if the first die you test qualifies for the i5 bin that the probability of seeing more i5 dies from that wafer increases. It could come down to a particular machine not being calibrated or some process step going awry. Things like that will affect the wafer as a whole.

Since it is a specialty top tier binned chip, I'm guessing the supply is really low. It's just a guess, though.

Probably a little bit of both. Lower and limited supply coupled with more demand for it as a halo product. Same reason the 7950X3D is sold out, but a 7900X3D is still available. Same story with GPUs where a 4090 or a 7900 XTX are hard to get, but a 4080 or 7900 XT are available.

 

[Hitman928]

I had t really thought about it before, but there is a certain amount of sense to the idea that if you've just diced your wafer and are going through the binning process that if the first die you test qualifies for the i5 bin that the probability of seeing more i5 dies from that wafer increases. It could come down to a particular machine not being calibrated or some process step going awry. Things like that will affect the wafer as a whole.

I'm assuming you are referring to i5's get binned on performance rather than defects forcing the cut down of a SKU. If so, then this is correct. There will be wafer to wafer variation and even intra-wafer variation but the wafer to wafer variation should dominate what the "quality" of the silicon is. The defects are from particles in the environment that end up on the wafer at various stages causes shorts/opens/other damage and isn't really related to the quality of the wafer itself from what I know.

 

[Exist50]

Defects can be somewhat separate from other quality measures which don't make the silicon defective, but would relegate it to a lower bin.

I had t really thought about it before, but there is a certain amount of sense to the idea that if you've just diced your wafer and are going through the binning process that if the first die you test qualifies for the i5 bin that the probability of seeing more i5 dies from that wafer increases. It could come down to a particular machine not being calibrated or some process step going awry. Things like that will affect the wafer as a whole.



Probably a little bit of both. Lower and limited supply coupled with more demand for it as a halo product. Same reason the 7950X3D is sold out, but a 7900X3D is still available. Same story with GPUs where a 4090 or a 7900 XTX are hard to get, but a 4080 or 7900 XT are available.

There's definitely a strong correlation between functional defects and parametric defects.

 

[CU]

For example string 5,2-1,20; 5,5-1,35; 5,8-1,45 is a little bit longer than 2 digit score, but it is real data.

Wish there was a large collection of VF curves for different chips online. It would be nice to know how your own chip compares.

 

[Hulk]

Wish there was a large collection of VF curves for different chips online. It would be nice to know how your own chip compares.

From what I've read the Asus SP value is the v/f curve programmed into the chip "boiled" down to a single number with higher numbers being better and ostensibly implying lower v/f curves.

Sure v/f curves printed on the boxes would be ideal but then we'd have scalping of the chips. Or make fine graduations in pricing from the lowest SP in a series to the highest, perhaps you'd pay $5 per extra SP rating or something. Never happen of course.

Seems like the relatively wide variation in SP ratings would open the door for a "silicon lottery" type of online shop to open up. Plus the shop wouldn't even have to test overclocking of a CPU, worry about customer, cooling, ram, mobo quality, etc... You want 120 SP? That's what you get.. for a price.

Obviously Asus knows how they generate the SP rating from the VID of a CPU but we'll never know how it's done.

 

[Timur Born]

On my 13900K the cores in red boxes are (made to be) prioritized by Windows scheduler before the other cores in order to keep space in between active cores. Putting the same load on adjacent cores leads to considerable temperature increases vs. putting it on non adjacent cores.

 

[Mopetar]

I'm assuming you are referring to i5's get binned on performance rather than defects forcing the cut down of a SKU. If so, then this is correct. There will be wafer to wafer variation and even intra-wafer variation but the wafer to wafer variation should dominate what the "quality" of the silicon is. The defects are from particles in the environment that end up on the wafer at various stages causes shorts/opens/other damage and isn't really related to the quality of the wafer itself from what I know.

We already know that most silicon isn't binned strictly on defects. The defect density is nowhere near high enough to force binning based strictly on physical defects. This has been true for quite a while now as there are plenty of historical examples of people finding out that they could reenable hardware to get additional performance out of a product.

What I'm hypothesizing is that if you take a random die from a wafer and put it throw binning process and find that it goes into a lower bin, that the likelihood of having other chips that go into that bin will increase. In other words the defects due to stray particles that would only affect a single die aren't the only factor that goes into silicon quality.

Even a high quality wafer that yields mainly top bin chips may still have a few with physical defects.

 

[Hitman928]

We already know that most silicon isn't binned strictly on defects. The defect density is nowhere near high enough to force binning based strictly on physical defects. This has been true for quite a while now as there are plenty of historical examples of people finding out that they could reenable hardware to get additional performance out of a product.

What I'm hypothesizing is that if you take a random die from a wafer and put it throw binning process and find that it goes into a lower bin, that the likelihood of having other chips that go into that bin will increase. In other words the defects due to stray particles that would only affect a single die aren't the only factor that goes into silicon quality.

Even a high quality wafer that yields mainly top bin chips may still have a few with physical defects.

Sounds like we're on the same page

 

[Timur Born]

If you have an idea or theory as to why all of the highly rated parts are 13900 series I'd love to hear it. Again, if you don't know please don't berate the question.

I could well be that the SP rating drops for CPUs that have cores "disabled", because the "average" drops. That is, if the SP rating is always calculated on the basis of a full core count 13900 then "disabled" (defect) cores would count as 0 (or at least lower) for the average. In that case the functioning cores of 13600 and 13700 may not be worse than the 13900 cores, their lack of full core count would just drag down the final SP number.

Obviously this is speculation, because we don't know how the proprietary SP algorithm works.

 

[Hulk]

I could well be that the SP rating drops for CPUs that have cores "disabled", because the "average" drops. That is, if the SP rating is always calculated on the basis of a full core count 13900 then "disabled" (defect) cores would count as 0 (or at least lower) for the average. In that case the functioning cores of 13600 and 13700 may not be worse than the 13900 cores, their lack of full core count would just drag down the final SP number.

Obviously this is speculation, because we don't know how the proprietary SP algorithm works.

Of course one opinion is as good as the next but I like the logic of how "Just Benching's" explanation explains the data. Basically that there is a variation in wafer quality and as Exist 50 wrote that there is a correlation between functional and parametric defects as both of these support the data of the nearly 700 tested chips, which should be enough to show a trend.

Boiling down what they wrote inter-wafer quality may be larger variable than intra-wafer quality for these tested cpu's, meaning a good wafer will not only have a lot of fully functional cores but also very high performing ones. A not so good wafer on the other hand will have many not fully functional cores and they will also not be great performers.

If we got the raw data from Intel we'd know if this is true or there is some better reason(s) or a combination. Unfortunately we'll never know. I thought the data was worth posting/discussing in our never ending quest of understanding how sand can be organized for doing useful things.

Or Intel is binning the good 13600's and 13700's to release as "KS" parts. Just kidding of course.

 

[Timur Born]

Wasn't it that the outer diameter of wafers usually produce lower quality than the center parts? That's somewhat loosely in the back of my head from reading somewhere at some time, so I am not sure.

 

[Hulk]

Speaking of yields and binning, what do you think the Raptor Lake Refresh will bring?

We have two Intel refreshes to look back on to see what Intel considers a refresh.

Kaby Lake Refresh was a mobile only refresh doubling cores (2 to 4)and slightly increasing max clocks by 300MHz. Coffee Lake Refresh simply added 2 cores to the top of the stack Coffee Lake sku.

It is unlikely Intel would increase Raptor Lake core count but would they re-arrange P and E counts a bit? Perhaps 6+24? This would be a relatively easy way to combat Zen 5 as most non well optimized MT applications that work well with "around" 8 cores would also do well with 6 (like games) so not much would be lost there. But the all important for advertising ridiculously MT apps/benches like CB would do better with 6+24 vs 8+16. Especially if another + on the 10nm process could crank up E clocks a bit.

The other and more likely possibility is a line of Raptors that have higher all core boost clocks. Say 6GHz for the 13900KS, 5.8GHz for the K's, 5.6GHz for the 13700K, and 5.3GHz for the 13600K.

I'm curious to know where they're going with this. Absolutely no leaks or information at this time other than it's coming.

 

[DAPUNISHER]

Steve has the Raptor tuning and scaling vid up

 

[Hulk]

I've had the 13600K for a few weeks now. How is it?

I don't really notice a performance difference from my 13900K. Perhaps with big video encoding jobs but honestly it runs cooler (stock settings) and never restarts. Voltage maxes ~1.25 under load but is generally ~1.2. I like not having the stress of finding settings that work around the restart issue of my 13900K.

I still want a 13900KS or Raptor Refresh if brings something interesting to the table and works with my mobo. But for now I'm just getting work done.

 

[igor_kavinski]

Steve has the Raptor tuning and scaling vid up

DDR5-6000 tuned is good enough for me.

 

[A///]

I've had the 13600K for a few weeks now. How is it?

I don't really notice a performance difference from my 13900K. Perhaps with big video encoding jobs but honestly it runs cooler (stock settings) and never restarts. Voltage maxes ~1.25 under load but is generally ~1.2. I like not having the stress of finding settings that work around the restart issue of my 13900K.

I still want a 13900KS or Raptor Refresh if brings something interesting to the table and works with my mobo. But for now I'm just getting work done.

Why'd you drop down to a 600?