Discussion Intel current and future Lakes & Rapids thread

[João Bortolace]

That's gotta lend itself to a pretty accurate estimation of the die size of the chiplets, nah? Based off the package size...
Also weird that one of the IO chiplets is ever so slightly offset lol, but whatever

interesting because the emib bridges are perfectly aligned, they must be different io title

 

[Dayman1225]

interesting because the emib bridges are perfectly aligned, they must be different io title

I/O tiles are the same, it’s not uncommon for Intel to do asymmetrical layouts with EMIB for the same dies. Like the chip on the bottom right here.
possible that the Emib on bottom of the compute die is placed different from the Emib on the top of the die causing that on GNR

 

[Dayman1225]

MTL with on package memory

https://x.com/hms1193/status/1699452529300631855?s=46&t=Go9LM3iS24APoFi_ulHkGw

This also looks like it could be Sierra Forest

 

[Geddagod]

MTL with on package memory
View attachment 85432

https://x.com/hms1193/status/1699452529300631855?s=46&t=Go9LM3iS24APoFi_ulHkGw

This also looks like it could be Sierra Forest
View attachment 85433

One might be able to estimate the die size of that too, but angle looks kinda tricky ngl.

 

[igor_kavinski]

MTL with on package memory
View attachment 85432

So is that how they are solving the LPDDR5X latency problem, @adroc_thurston ?

 

[adroc_thurston]

So is that how they are solving the LPDDR5X latency problem

No it's just a packaging area play.
Trace length has zero impact on memlat.

 

[H433x0n]

Not super encouraging tbh for MTL tbh, PPC looks to be marginally worse than Zen 4. However, GB has typically had higher scores for AMD anyway, so not too surprised.
This would put single core ~5% higher for Raphael, and multi core (for what little that's worth based on the no power consumption given), appears to be higher for MTL than Raphael.

All of these models are the “Ultra 7” or lower. The Ultra 9 models will bring Geekbench 6 ST score to ~2600 (depending on rest of hardware).

Edit: Looking at other Geekbench 6 scores, it appears that it should outperform 13700H in MT and tie it in ST performance.

 

[Khato]

No it's just a packaging area play.
Trace length has zero impact on memlat.

Trace length does affect active power usage, so moving memory on package provides more benefits than just reduce total board area. Also simplifies the board traces a bit.

 

[SiliconFly]

Hardware Unboxed shows different results where Intel offers superior performance.

Also memory bandwidth limitation is a myth. 13900K would beat 7800X3D even with DDR4.
View attachment 85340

Zen 3 did really poorly here... Hard to believe that this is AMD sponsored game - maybe they didn't care about the CPU part at all. 13900K being 2x faster than 5950X...

Yep. It's kinda hard to believe. Too many negative reports: here, here & here. But thats just the tip of the iceberg. Like many AMD sponsored games, Starfield doesn't have DLSS (with no future plans either). But a dev has released a DLSS3 mod for NVIDIA GPUs.

Now here comes the crazy part... On NVIDIA GPUs, Starfield runs twice as fast with the DLSS3 mod compared to the default FSR offered by the game! Looks like AMD sponsored the game on purpose to save face!

 

[adroc_thurston]

Trace length does affect active power usage

Not really much if at all.

Also simplifies the board traces a bit.

And makes the overall SoC + MEM footprint a lot smaller which is the point.

 

[SiliconFly]

That's right, but Intel is so inefficient that even if they only reduce power consumption and keep the same performance, it's considered a 'success.' I don't like the direction Intel is going in. New generations of mobile and desktop processors practically offer minimal improvements in terms of performance. When I buy a new generation, I don't want it to only offer lower power consumption; I want better performance as well. Otherwise, I won't buy them just for the sake of lower power consumption. Intel is still searching, experimenting, trying to find the winning formula for chiplet design to break out of their rut, while AMD has a PhD in chiplet design and is moving full steam ahead. Also, when we talk about chiplet design, Intel's solution is much more expensive than AMD's, and we'll see what the prices will be. However, considering they have their own factories and produce much larger quantities, they can afford minimal profit per unit.

Ya. It's a bit too soon to tell. But all the leaked scores until now puts MTL's performance a bit lower than the 13th gen counterparts. Maybe the performance of final silicon (with bios, os, drivers, system & other h/w optimizations) might slightly exceed the 13th gen equivalents, but not by much I guess. I think MTL's main selling point is gonna be power efficiency.

 

[dullard]

Ya. It's a bit too soon to tell. But all the leaked scores until now puts MTL's performance a bit lower than the 13th gen counterparts. Maybe the performance of final silicon (with bios, os, drivers, system & other h/w optimizations) might slightly exceed the 13th gen equivalents, but not by much I guess. I think MTL's main selling point is gonna be power efficiency.

What leaked scores are you looking at to reach that conclusion? I tried my best to do a fair comparison (same platform, same memory size, same bottom of the i7 H line CPU, etc) and got same ST performance and 80% better MT performance with Meteor Lake.

Page 99 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

Page 99 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Geddagod]

What leaked scores are you looking at to reach that conclusion? I tried my best to do a fair comparison (same platform, same memory size, same bottom of the i7 H line CPU, etc) and got same ST performance and 80% better MT performance with Meteor Lake.

Page 99 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

Page 99 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Single core regression would be a embarrassing. Especially since most consumers are going to feel the single core regression more than a multi core one, or even a multi core gain (even if the regression is only by like ~5%).

 

[mikk]

I don't think it's that bad lol, is it?

Check out i7-1370p, the scores are all over the place: https://browser.geekbench.com/search?page=1&q=i7-1370p

Because there are so many we can pick out the fastest one. We don't have the luxury with the few Ultra 7 results.

 

[dullard]

Single core regression would be a embarrassing. Especially since most consumers are going to feel the single core regression more than a multi core one, or even a multi core gain (even if the regression is only by like ~5%).

I agree that you feel single core more, but honestly you usually can't feel anything within about +-10%. Heck with laptops the room temperature might impact the score more than any actual hardware single-digit CPU percent change.

With what looks to be a whole lot of other changes that will impact the performance FAR more, I don't think small single threaded regression (if any occur) will be that important. For example, if you do anything with graphics, Meteor Lake should easily win over Raptor Lake by a mile. If then even there was a hypothetical single core regression, that isn't what I'd call "embarrassing".

 

[Geddagod]

I agree that you feel single core more, but honestly you usually can't feel anything within about +-10%. Heck with laptops the room temperature might impact the score more than any actual hardware single-digit CPU percent change.

With what looks to be a whole lot of other changes that will impact the performance FAR more, I don't think small single threaded regression (if any occur) will be that important. For example, if you do anything with graphics, Meteor Lake should easily win over Raptor Lake by a mile. If then even there was a hypothetical single core regression, that isn't what I'd call "embarrassing".

Perhaps, but a gen on gen regression in single core performance is still a pretty bad look. I would say it's embarrassing even. Esp since you can't argue that there was an IPC increase to perhaps compensate for a frequency reduction (like ICL).

 

[Khato]

Not really much if at all.

Even if no other factors changed, then the transmission line losses are directly proportional to length. Of course bringing the memory on package is another factor that provides further improvements. The on package pJ/bit at same speed is probably around 1/4 of on board memory. Now sure the interface power for lpddr5 is probably only around 20% or so of memory power usage (annoying to find figures for lpddr5 so guessing based on figures from prior generations.) But that's still more than enough to have a notable impact.

 

[lightisgood]

a gen on gen regression in single core performance

Do you have any other evidence?
I simply don't think this will come in MTL.
We already know that high-end MTL model has 128MiB ADM cache.

 

[dullard]

Check out i7-1370p, the scores are all over the place: https://browser.geekbench.com/search?page=1&q=i7-1370p

Because there are so many we can pick out the fastest one. We don't have the luxury with the few Ultra 7 results.

The key with laptop chips is to compare the same actual laptop to each other. You can't compare a massive gaming laptop with great cooling to a thin-and-light with little cooling and expect to get any meaningful results relating to a CPU. That is because the CPU will be at turbo far more with the good cooling and a base speeds far more with the poor cooling.

Next compare other important factors. Just about every bad score (specifically Multi-Core under 10000) on the first page of your link was with Windows 10, half speed memory, or "invalid due to an issue with the timers on this system".

Once you control for all that, the scores are much more consistent (Multi-Core from 10600 to 13000). Unfortunately GeekBench doesn't always give you sufficient information. For example, it often likes to leave out memory size or speed in their results.

 

[Glo.]

I mean that package of MTL-P CPU is pretty much the whole subsystem of the computer, including the PCH, right?

Imagine how small NUCs can get with packages like those...

 

[dullard]

I mean that package of MTL-P CPU is pretty much the whole subsystem of the computer, including the PCH, right?

Imagine how small NUCs can get with packages like those...

And then Intel got out of the NUC business (due to budget cuts).

 

[adroc_thurston]

then the transmission line losses are directly proportional to length.

Depends on your PHY/encoding scheme/etc.
God I ain't doing any comparch 101 things here.
Onpkg memory is just PKG area play, that's it.

 

[igor_kavinski]

God I ain't doing any comparch 101 things here.

Don't disappoint us, Professor adroc.

Kindly elaborate.

 

[Glo.]

And then Intel got out of the NUC business (due to budget cuts).

They will still exist, with new products

 

[Khato]

Depends on your PHY/encoding scheme/etc.
God I ain't doing any comparch 101 things here.

Considering that such is squarely on the electrical engineering side, yeah, it's not covered in comparch 101.

Anyway, it's not like Intel is doing anything novel here. Exact same approach Apple used on their m series to minimize memory power consumption while sticking to cheaper lpddr. It'll be interesting to see whether Intel expands on package memory to other segments in the coming years... I can't help but amuse myself with the unpleasant thought of sku segmentation including memory, eg 32gb+ not available on anything below an i7.