News Intel Raptor Lake Refresh, Arrow Lake CPU Performance Projections Leaked

[Kamen Rider Blade]

Not just on-package but DIRECT-STACKED, the logical evolution of HBM: eliminate the base die by integrating whatever essential functionality is still necessary directly into whatever the raw DRAM stack gets installed on top of, such as the IO die that contains the memory controller. HBM already goes to 1TB/s per stack, direct-stacked memory should be able to go even higher by ditching the base die and practically eliminating wiring stubs between the DRAM dies and memory controller. DDR5-8000 on the other hand is still only 64GB/s per aggregate channel.

And you've solved for Thermals how?

1TB/s vs 64GB/s looks like a solid order of magnitude to me.

Fair enough if you can pull it off, but you don't need "Direct Stacked" for that.

HBM does that by being adjacent as is.

Once CPUs have enough direct-stacked memory to handle the active data set? Yes, because the extra memory beyond that is just a glorified swapfile, not seeing much traffic relative to how large it is.

Ok, you're only going to be able to fit a finite amount of Memory on Top of the tiny surface area for the ASIC portion, I don't think you're going to be able to fit nearly as much memory as you think and you still haven't solved the heat issue.

 

[InvalidError]

Ok, you're only going to be able to fit a finite amount of Memory on Top of the tiny surface area for the ASIC portion, I don't think you're going to be able to fit nearly as much memory as you think and you still haven't solved the heat issue.

The ~120sqmm IOD/SoC tile has enough surface area to slap 3GB DRAM chips on there and we're able to stack them up to 16-high, which means current tech could already pack up to 48GB.

As for thermals, what thermals? There is no problem that needs solving there: the IOD for AMD and SoC tile for Intel which contains the memory controller draw maybe 10W peak, add another ~10W for the two E-cores on Meteor Lake's SoC tile when they are under full load with the DRAM stack adding 2-3W itself. Cooling 20-25W/sqcm is a non-issue when you are already dealing with 150+W/sqcm for the CCDs and CPU tiles.

 

[Kamen Rider Blade]

The ~120sqmm IOD/SoC tile has enough surface area to slap 3GB DRAM chips on there and we're able to stack them up to 16-high, which means current tech could already pack up to 48GB.

As for thermals, what thermals? There is no problem that needs solving there: the IOD for AMD and SoC tile for Intel which contains the memory controller draw maybe 10W peak, add another ~10W for the two E-cores on Meteor Lake's SoC tile when they are under full load with the DRAM stack adding 2-3W itself. Cooling 20-25W/sqcm is a non-issue when you are already dealing with 150+W/sqcm for the CCDs and CPU tiles.

Did you forget that modern cIOD has the iGPU integrated?

AMD isn't going to "Undo" that decision moving forward.

AMD Ryzen 7000 integrated RDNA2 GPU overclocked to 3.1 GHz, up to 42% performance increase - VideoCardz.com

AMD Ryzen 7900 iGPU gets 30%+ higher performance Skatterbencher known for extreme overclocking attempts, puts AMD’s desktop CPU under a microscope. The latest review focuses on integrated graphics performance and its optimizations. One should remind our readers that Ryzen 7000 desktop series...

videocardz.com

The stock iGPU configuration offers 2.2 GHz clock speed at around 0.997 volts with Furmark software. Just as important is the memory clock, which in this case is 2.4 GHz. With such a setting, the CPU’s GPU and SOC (separate die on the package) consume around 38.5 watts. With manual overclocking, he was able to overclock the iGPU to 2949 MHz clocks at 1.2V, that also increase the power consumption to 63.5 watts. At the same time, memory clock was adjusted to 3.2 GHz.

The cIOD with 2 CU iGPU is consuming 38.5 watts & 63.5 watts when OCing. That's for the GPU portion alone.

The SoC or cIOD itself is consuming 23 watts.

For a grand total of either 61.5 watts under normal stress load or 86.5 watts when OCing the GPU on the cIOD.

So how are you going to integrate On-Package DRAM on top that has to draw more power and deal with that level of thermals on top?

 

[InvalidError]

So how are you going to integrate On-Package DRAM on top that has to draw more power and deal with that level of thermals on top?

Silicon has horrible thermal conductivity and the DRAM stack is coming out of the passivation layer thickness. The impact on thermal conductivity should be low enough to not matter much below 50W/sqcm. Even if the IOD goes up to 40W under stock settings at full load, that is only about 33W/sqcm. Again, non-issue.

 

[bit_user]

AMD bet the farm on zen and now the small cores are the same zen core just stripped down with half the cache...

IMO, it's worth reading a good analysis of Zen 4c, because they did a lot more than just remove half of the L3 cache. It seems like they made numerous design changes to make it smaller and more efficient.

​

Zen 4c: AMD’s Response to Hyperscale ARM & Intel Atom

Bergamo Volumes, ASP, Performance, Hyperscale Order Shift, Die Shot, Floorplan, Physical Design, and Future Use of Dense Core Variants Bergamo, AMD’s upcoming 128-core server part sets new heights …

www.semianalysis.com

​

Wasn't till Goldmont till Atom got OoO execution, tremendous performance improvment in chromebooks as much better media accelerators.

Actually, Silvermont had OoO. I'm pretty sure that was the first generation to have it (WikiChip confirms), but they definitely dropped SMT whenever they switched.

​

Silvermont - Microarchitectures - Intel - WikiChip

Silvermont (SLM) is Intel's 22 nm microarchitecture for the Atom family of system on chips. Introduced in 2013, Silvermont was the successor to Saltwell, targeting smartphones, tablets, embedded devices, and consumer electronics.

en.wikichip.org

​

 

[bit_user]

The wraith at 60% is about 130W power and that gives you 4.3 out of 5.2Ghz.

150 W is 65% of 230 W, but that's not what counts.

It's not power that counts. I already cited the performance graph, and that's what matters. We all (hopefully) know that power increases at a super-linear rate, with respect to performance. That's why you want to focus on performance, not power or even clock speed.

The Noctua cooler at 40% offers 97% as much performance as the Artic AIO. So, if you take that as representative of mid-market air cooler, it shows that air cooling is a very viable option for the 7950X.

The Wraith Spire they used is an entry-level cooler rated only for 95 W of cooling. It was never intended to be used with such a CPU, so was really meant to test the CPU at its extremes, rather than to be considered as any kind of serious option.

I can't even find an example of someone using a comparable cooler on Raptor Lake. The closest I could find was when Toms tested the 6-pipe Thermalright AXP120-X67, rated for 200 W, and it fared even worse than the less-capable Wraith on the the 7950X!

​

Again, this is the Raptor Lake i9-13900K:​

Source: https://www.tomshardware.com/features/intel-core-13900k-cooling-tested/2​

​

Furthermore, that's showing performance, rather than the power & frequency graphs you cited.

People in glass houses shouldn't throw stones.

 

[bit_user]

Another reason why the "THICC IHS" gets a bad wrap and it really is the cut-outs and the removal of Surface Area that is the issue, not so much the thickness.

Even with that bad IHS, @TerryLaze is grossly exaggerating the cooling requirements of the 7950X. Although the die does get hot and throttle, it's still able to deliver the vast majority of its performance before that point.

 

[bit_user]

The most common example of high memory usage in the mainstream is video editing. While 8k video editing may benefit from having 200+GB of RAM to cache clips for timeline scrubbing, that is only a cache to avoid going all the way back through the OS and file system to storage,

This example is fishy. High IOPS is achievable even, going through the OS and filesystem. I'd guess what's really happening is that the app is caching uncompressed (maybe downsampled) video frames in that extra RAM, and the responsiveness comes from avoiding decoding.

 

[bit_user]

Then you wouldn't mind having On-Package DRAM + Standard Expandable Memory via DIMM slots.

Plenty of LapTops already do this, why can't we just do this on DeskTop as well?

No laptops currently use HBM-class memory, like what @InvalidError is talking about. The closest example is Apple, and they don't have any DIMM slots.

I think you're confusing memory soldered on the motherboard with memory that's on the same interposer as the CPU. They're very different things.

 

[bit_user]

Did you forget that modern cIOD has the iGPU integrated?

AMD isn't going to "Undo" that decision moving forward.

Really? With so much memory bandwidth now available, it finally makes sense to have a big iGPU, like Apple is doing. At that point it would certainly be justifiable to move the iGPU onto its own tile.

If you still had a "toy" iGPU integrated into the IOD for desktops, then you could manage its thermal output just by keeping its clocks down. For instance, Intel's 24 EU iGPUs used < 10 W on Skylake.

 

[InvalidError]

No laptops currently use HBM-class memory, like what @InvalidError is talking about. The closest example is Apple, and they don't have any DIMM slots.

And even among laptops that do still have two DIMM slots instead of one channel worth of memory soldered on board, the configuration maxes out at 64GB except for the absolute highest-end models that may support 48GB DDR5 SO-DIMMs or more than two DIMMs worth of memory.

We're in $4000+ unicorn laptop territory at that point.

 

[bit_user]

And even among laptops that do still have two DIMM slots instead of one channel worth of memory soldered on board, the configuration maxes out at 64GB except for the absolute highest-end models that may support 48GB DDR5 SO-DIMMs or more than two DIMMs worth of memory.

We're in $4000+ unicorn laptop territory at that point.

Although we're getting way off topic, I just wanted to point out that Dell offers laptops using CAMM modules, which will supposedly scale to 128 GB. I know their Precision 7000 line uses these, and the cheapest one I see on their website is currently just under $2k. The specs are pretty much the same as my work laptop, and that's not bad at all.

 

[Kamen Rider Blade]

No laptops currently use HBM-class memory, like what @InvalidError is talking about. The closest example is Apple, and they don't have any DIMM slots.

And Apple use LPDDR5 and places it adjacent to the SoC.

Apple choosing not to offer DIMM slots is Apple being more & more controlling.

I think you're confusing memory soldered on the motherboard with memory that's on the same interposer as the CPU. They're very different things.

But it's the closest analogy in the current era that has soldered on RAM that is sold to the public outside of Apple.

And even among laptops that do still have two DIMM slots instead of one channel worth of memory soldered on board, the configuration maxes out at 64GB except for the absolute highest-end models that may support 48GB DDR5 SO-DIMMs or more than two DIMMs worth of memory.

We're in $4000+ unicorn laptop territory at that point.

That's just typical LapTop over-priced BS.

 

[bit_user]

And Apple use LPDDR5 and places it adjacent to the SoC.

It's not merely adjacent, it's on the same interposer. That's a big difference.

Apple M2 Ultra SoC Delidded, Package Size Bigger Than Intel's Sapphire Rapids Xeon CPUs

The Apple M2 Ultra SoC may not be the x86 killer it was deemed to be but it certainly is a massive chip on its own.

wccftech.com

Apple choosing not to offer DIMM slots is Apple being more & more controlling.

At least they have a strong technical argument for it, though.

But it's the closest analogy in the current era that has soldered on RAM that is sold to the public outside of Apple.

That analogy didn't apply, in the context where you were trying to use it.

 

[InvalidError]

Apple choosing not to offer DIMM slots is Apple being more & more controlling.

Apple choosing not to have external memory is Apple determining that its target audience will be fine with whatever memory options it tosses on its SoCs. Most of Apple's recent laptops have all memory soldered to the board and not user-upgradable either, no meaningful change as far as Apple's customers are concerned. Many cheaper x86 laptops and practically all 2-in-1s do the same too.

As memory gets cheaper, expect more laptop manufacturers to decide that it is more cost-effective and profitable for them to pack some more RAM on their motherboards across a few more models up their product stack than bother with DIMM slots. I bet they'd welcome CPU options with on-package DRAM enabling them to skip having to deal with DRAM altogether while saving board space and power.

 

[Kamen Rider Blade]

Apple choosing not to have external memory is Apple determining that its target audience will be fine with whatever memory options it tosses on its SoCs. Most of Apple's recent laptops have all memory soldered to the board and not user-upgradable either, no meaningful change as far as Apple's customers are concerned. Many cheaper x86 laptops and practically all 2-in-1s do the same too.

As memory gets cheaper, expect more laptop manufacturers to decide that it is more cost-effective and profitable for them to pack some more RAM on their motherboards across a few more models up their product stack than bother with DIMM slots. I bet they'd welcome CPU options with on-package DRAM enabling them to skip having to deal with DRAM altogether while saving board space and power.

Hopefully FrameWork & the User's right to Repair will start shifting things back in the other direction.

Sorry, I like having choice, not having my RAM decided for me.

FrameWork & the User's right to Repair is what I want, and I won't be spending any $$$ on ANY LapTop that has soldered in memory.

I'm going to make sure everybody I know avoids that type of LapTop.

Pre-Soldered in Memory = NO SALE!

I HATE the Pre-Soldered in CPU's in LapTops, that was one of the WORST decisions I see LapTop vendors go down, I want things to go back to user modularity. Just like DeskTop PC CPU's

Eff this Apple Mentality of chasing the never ending quest for "Thin-ness".

 

[bit_user]

Hopefully FrameWork & the User's right to Repair will start shifting things back in the other direction.

Right to Repair will have no effect on RAM upgradability. I highly doubt it'll even turn back the hands of time on laptop CPUs being soldered down.

As for Framework, they'll gladly sell you an entire new CPU module.

https://frame.work/marketplace/mainboards
​

Sorry, I like having choice, not having my RAM decided for me.

You'll have a choice. It will just be locked down at the time you select your CPU model. If you want to add capacity, then maybe you can get a laptop model with a CXL memory slot.

I won't be spending any $$$ on ANY LapTop that has soldered in memory.

I'm going to make sure everybody I know avoids that type of LapTop.

Pre-Soldered in Memory = NO SALE!

Okay, but it's hardly new a new development.

For my current laptop, I bought an "open box new" model on ebay for about $450. Most of Framework's motherboard modules cost more than that. I'll consider a Framework, for my next laptop, but it's looking unlikely I'll go that route.

 

[Kamen Rider Blade]

Right to Repair will have no effect on RAM upgradability. I highly doubt it'll even turn back the hands of time on laptop CPUs being soldered down.

As for Framework, they'll gladly sell you an entire new CPU module.

​

https://frame.work/marketplace/mainboards​

The entire industry is obsessed with Apple for the never-ending quest for "Thin-ness".

It's absurd, stupid, & pointless.

We can literally make CPU with LGA sockets that can have Direct Die Frames for cooling.

There's literally no reason for BGA Pre-Soldered other than to make it harder to repair.

All in the name of "Thin-ness" which isn't worth it IMO.

You'll have a choice. It will just be locked down at the time you select your CPU model. If you want to add capacity, then maybe you can get a laptop model with a CXL memory slot.

No DIMM or CAMM, no Sale.

Eff CXL

Okay, but it's hardly new a new development.

For my current laptop, I bought an "open box new" model on ebay for about $450. Most of Framework's motherboard modules cost more than that. I'll consider a Framework, for my next laptop, but it's looking unlikely I'll go that route.

Your choice, I want to practice what I preach, modularity & right to repair.

Now to push back for more choice & modularity like the old days.

 

[bit_user]

No DIMM or CAMM, no Sale.

Eff CXL

What I find funny about that is how you seem to appreciate the role of caches, in a memory hierarchy. And yet you somehow draw the line at having super-fast in-package memory and basically swapping in pages from a slower external DIMM. It's pretty much just an extension of the same basic cache hierarchy concept.

Your choice, I want to practice what I preach, modularity & right to repair.

Now to push back for more choice & modularity like the old days.

Hey, I like upgradable storage and RAM. My current laptop has DIMMs and a M.2 SSD and it's thin & light enough for me. If I'm not getting a benefit, like in-package RAM can provide, then I'll certainly prefer to have upgradable RAM.

When I bought my laptop, it had only one DIMM installed. So, I simply bought another one and that was that. I also planned to upgrade the SSD, but the existing storage has so far been more than enough.

 

[TerryLaze]

150 W is 65% of 230 W, but that's not what counts.

And today is wednesday which is the middle of the week...
Since we are making random statements.

130W is half of 260W which was me being generous with a potential 32core CPU.
You showed a performance graph with the noctua still giving 220-210W to the CPU but your argument in the previous post was that you would reduce the power to the cores to have them all still only use 230W.
So you have to take the performance you get at 115W to make your point and not the performance the noctua gives you which is still ~220W, that would make a 32core that would use 440W , we where at this point several posts ago already but like always you have to chew the other persons ear off with nonsense until they are too bored to talk to you anymore at which point, I guess, you believe you won internet points or something.

 

[Kamen Rider Blade]

What I find funny about that is how you seem to appreciate the role of caches, in a memory hierarchy. And yet you somehow draw the line at having super-fast in-package memory and basically swapping in pages from a slower external DIMM. It's pretty much just an extension of the same basic cache hierarchy concept.

I have no problem with "In-Package-Memory" as long as you don't arbitrarily take away my DIMM slots that are directly attached and try to sell me on CXL as the solution.

CXL is a nice "Extendeable Option" once you run out of DIMM slots that are direct attached, but it shouldn't be a replacement.

Hey, I like upgradable storage and RAM. My current laptop has DIMMs and a M.2 SSD and it's thin & light enough for me. If I'm not getting a benefit, like in-package RAM can provide, then I'll certainly prefer to have upgradable RAM.

When I bought my laptop, it had only one DIMM installed. So, I simply bought another one and that was that. I also planned to upgrade the SSD, but the existing storage has so far been more than enough.

But I want all of it to be upgrade-able.

If I want a new MoBo, just buy one and drop it in, just like a DeskTop PC.

If I want a new CPU, just buy a new CPU and stick it in.

It's not like AMD & Derbauer haven't experimented with Direct Die frames, that's all you really need for LapTops thin cooling solutions to work.

 

[bit_user]

But I want all of it to be upgrade-able.

If I want a new MoBo, just buy one and drop it in, just like a DeskTop PC.

If I want a new CPU, just buy a new CPU and stick it in.

I understand what you're saying and Framework is the only current example, to my knowledge, where you can swap out the motherboard.

Unless you want to run a desktop CPU inside a laptop, and suffer all the consequences of doing so, you're probably always going to have to live with swapping the motherboard + CPU as a single unit.

One nice thing about Framework is that you can get desktop shells, in case you want to reuse an old motherboard for some non-portable purpose (e.g. HTPC, micro-server, etc.).

 

[Kamen Rider Blade]

I understand what you're saying and Framework is the only current example, to my knowledge, where you can swap out the motherboard.

Unless you want to run a desktop CPU inside a laptop, and suffer all the consequences of doing so, you're probably always going to have to live with swapping the motherboard + CPU as a single unit.

One nice thing about Framework is that you can get desktop shells, in case you want to reuse an old motherboard for some non-portable purpose (e.g. HTPC, micro-server, etc.).

It wasn't that many years ago that LapTops had modular socketable CPU's that weren't DeskTop CPU's shoved into a LapTop body.

With modern Direct-Die frames & LGA sockets, you can still make "Pretty Damn Thin" LapTops that aren't "Super Model" Thin that Apple loves to do.

It's "Good Enough" IMO and makes things more repair-able and gives you choice.

And given AMD's penchant for supporting a platform for 3x generations, I think it would be good for the eco-system to go that route on the LapTop side.

Also developing one LGA socket that works for LapTop & DeskTop would be kind of nice.

Spoiler: Z-Height of LGA 1700 Socket + IHS

Imagine how tall a Direct-Die frame attached to the CPU for LapTops would be in Z-Height.
Probably about 50% Z-Height @ 3-4 mm high with a directly attached screw mount.

Make use of that Nano-ITX form factor. 120mm x 120mm is plenty of space, and you definitely needs screw holes to make sure your Cooler has enough pressure to clamp down on your CPU/APU for good contact anyways.

We can use properly torqued screws to attach the Direct Die Frames & Flat Heapipe / Vapor Chamber setups that are designed to screw into a AMD defined hole patterns.

Don't be afraid to use screws to mount everything, even torque limited screws.

Hell, Derbauer designed a Direct Die Frame & AMD was already testing them internally.

We know for a fact that Intel uses a screw/QD mount on frame to easily test LapTop CPU's.

Spoiler: Intel's BGA Interposer test rig

If Intel refuses to sell taht kind of Quick Attach System to the consumers, then we can go back to LGA + Direct-Die Attached CPU's.

 

[bit_user]

130W is half of 260W which was me being generous with a potential 32core CPU.

That wasn't clear to me, and also doesn't align with the "60%" comment, but we can leave that aside.

You showed a performance graph with the noctua still giving 220-210W to the CPU but your argument in the previous post was that you would reduce the power to the cores to have them all still only use 230W.
So you have to take the performance you get at 115W to make your point

Not all of the power goes to the cores. If you assume the I/O die uses 20 W, then we ought to be able to use a TDP of 230 W by doubling the 125 W performance:

​

However, PPT seems to be about 35% above the TDP, meaning a 230 W CPU would have a PPT of 311 W. If we're instead targeting a TDP/PPT of 170/230 W, then the per-core performance should be somewhere in between the 65 W and 105 W cases. If we pessimistically assume half-way in between, then that's about 2100 points per core. So, if you scaled that linearly, you'd get an estimated score of 67154.

Now, there's way too much conjecture and hand-waving in that estimate. Linear scaling not only depends on not prematurely throttling but also having enough memory bandwidth and the software being implemented in a scalable way. Far too many unknowns. So, I'm not going to hew to any actual performance predictions, other than that to say if we merely doubled the performance of the 7950X at TDP of 85 W, it would be compelling.

like always you have to chew the other persons ear off with nonsense until they are too bored to talk to you anymore

What matters is that there's a coherent and relevant claim that's made with sound supporting data and logic. Unless and until those conditions are met, of course we're going to iterate.

at which point, I guess, you believe you won internet points or something.

Internet points are a lie.

I don't know why you posted that again, because as I said it's misleading. Performance doesn't scale linearly with power, as any overclocker knows and as was clearly demonstrated in Anandtech's TDP scaling experiments (above).

That's why I posted the actual performance data they achieved with 4 different cooling configurations. Unless you're posting in bad faith, you shouldn't be trying to mislead.

 

[bit_user]

It wasn't that many years ago that LapTops had modular socketable CPU's that weren't DeskTop CPU's shoved into a LapTop body.

With modern Direct-Die frames & LGA sockets, you can still make "Pretty Damn Thin" LapTops that aren't "Super Model" Thin that Apple loves to do.

It's "Good Enough" IMO and makes things more repair-able and gives you choice.

And given AMD's penchant for supporting a platform for 3x generations, I think it would be good for the eco-system to go that route on the LapTop side.

A Framework-like laptop motherboard doesn't have very much on it. This should hopefully make the actual board rather simple and inexpensive. So, perhaps it should be an acceptable compromise merely to make the cooling solution resuable and offer motherboard modules for sale without it.

We know for a fact that Intel uses a screw/QD mount on frame to easily test LapTop CPU's.

Intel's BGA Interposer test rig:

Except a real laptop needs to be light, durable, and shock-resistant. Their lab solution probably would make some compromises in those areas. Also, I wonder how much more power you'd burn by not soldering it down.