• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Intel Arrow Lake-S Die Visibly Larger Than Raptor Lake-S, Die-size Estimated

btarunr

Editor & Senior Moderator
Staff member
Joined
Oct 9, 2007
Messages
47,288 (7.53/day)
Location
Hyderabad, India
System Name RBMK-1000
Processor AMD Ryzen 7 5700G
Motherboard ASUS ROG Strix B450-E Gaming
Cooling DeepCool Gammax L240 V2
Memory 2x 8GB G.Skill Sniper X
Video Card(s) Palit GeForce RTX 2080 SUPER GameRock
Storage Western Digital Black NVMe 512GB
Display(s) BenQ 1440p 60 Hz 27-inch
Case Corsair Carbide 100R
Audio Device(s) ASUS SupremeFX S1220A
Power Supply Cooler Master MWE Gold 650W
Mouse ASUS ROG Strix Impact
Keyboard Gamdias Hermes E2
Software Windows 11 Pro
As a quick follow-up to last week's "Arrow Lake-S" de-lidding by Madness727, we now have a line-up of a de-lidded Core Ultra 9 285K "Arrow Lake-S" processor placed next to a Core i9-14900K "Raptor Lake-S," and the Core i9-12900K "Alder Lake-S." The tile-based "Arrow Lake-S" is visibly larger than the two, despite being made on more advanced foundry nodes. Both the 8P+16E "Raptor Lake-S" and 8P+8E "Alder Lake-S" chips are built on the Intel 7 node (10 nm Enhanced SuperFin). The "Raptor Lake-S" monolithic chip comes with a die-area of 257 mm². The "Alder Lake-S" is physically smaller, at 215 mm². What sets the two apart isn't just the two additional E-core clusters on "Raptor Lake-S," but also larger caches—2 MB of L2 per P-core, increased form 1.25 MB/core, and 4 MB per E-core cluster, increased from 2 MB/cluster.

Thanks to high quality die-shots of the "Arrow Lake-S" by Madness727, we have our first die-area estimations by A Hollow Knight on Twitter. The LGA1851 fiberglass substrate has the same dimensions as the LGA1700 substrate. This is to ensure the socket retains cooler compatibility. Using geometrical measurements, the base tile of the "Arrow Lake-S" is estimated to be 300.9 mm² in area. The base-tile is a more suitable guideline for "die-area," since Intel uses filler tiles to ensure gaps in the arrangement of logic tiles are filled, and the chip aligns with the base-tile below. The base tile, built on an Intel 22 nm foundry node, serves like a silicon interposer, facilitating high-density microscopic wiring between the various logic tiles stacked on top, and an interface to the fiberglass substrate below.



A Hollow Knight also put out his die-area estimations for each of the individual tiles. The Compute tile is estimated to measure 114.5 mm². This is the largest tile by area, and is built on the 3 nm TSMC N3B foundry node, and contains the CPU core complex with an 8P+16E core-count. The second-largest tile is the SoC tile, estimated to measure 86.1 mm². This tile is built on the 6 nm TSMC N6 node. The I/O tile is an extension of the SoC tile, it is built on the same 6 nm node, and is estimated to measure 23.79 mm².

The Graphics tile is, interestingly, smaller than even the I/O tile, and measures 22.8 mm². Intel built this on the 5 nm TSMC N5 node. It contains the brains of the iGPU, with 4 Xe cores, and other graphics rendering machinery. The media-accelerators, display controller, and display PHY components, are located on the SoC tile. Then there are the two filler tiles, the more visible of the two is the one next to the I/O tile, measuring 17 mm², and a smaller one next to the Graphics tile, measuring 2.6 mm². The filler tiles ensure the clump of logic tiles ends up having a rectangular form and a uniform Z-Height.

Interestingly, a slender, rectangular portion of the base tile remains exposed, with nothing stacked on top. This is what contributes to the 300.9 mm² die-area measurement by A Hollow Knight. If you were to simply add up the areas of the logic tiles, subtracting this exposed portion of the base tile, and the two filler tiles, you end up with 247.2 mm², which is in fact smaller than that of the monolithic "Raptor Lake-S."

View at TechPowerUp Main Site | Source
 
Joined
Sep 17, 2014
Messages
22,638 (6.04/day)
Location
The Washing Machine
System Name Tiny the White Yeti
Processor 7800X3D
Motherboard MSI MAG Mortar b650m wifi
Cooling CPU: Thermalright Peerless Assassin / Case: Phanteks T30-120 x3
Memory 32GB Corsair Vengeance 30CL6000
Video Card(s) ASRock RX7900XT Phantom Gaming
Storage Lexar NM790 4TB + Samsung 850 EVO 1TB + Samsung 980 1TB + Crucial BX100 250GB
Display(s) Gigabyte G34QWC (3440x1440)
Case Lian Li A3 mATX White
Audio Device(s) Harman Kardon AVR137 + 2.1
Power Supply EVGA Supernova G2 750W
Mouse Steelseries Aerox 5
Keyboard Lenovo Thinkpad Trackpoint II
VR HMD HD 420 - Green Edition ;)
Software W11 IoT Enterprise LTSC
Benchmark Scores Over 9000
Filler tiles, a nonaligned mass of chips stuck to a substrate... man this looks like a hack job more so than a nice design

And to consider Intel was joking at AMD one time for 'glue'... tsk tsk
 
Joined
Jun 11, 2019
Messages
636 (0.32/day)
Location
Moscow, Russia
Processor Intel 12600K
Motherboard Gigabyte Z690 Gaming X
Cooling CPU: Noctua NH-D15S; Case: 2xNoctua NF-A14, 1xNF-S12A.
Memory Ballistix Sport LT DDR4 @3600CL16 2*16GB
Video Card(s) Palit RTX 4080
Storage Samsung 970 Pro 512GB + Crucial MX500 500gb + WD Red 6TB
Display(s) Dell S2721qs
Case Phanteks P300A Mesh
Audio Device(s) Behringer UMC204HD
Power Supply Fractal Design Ion+ 560W
Mouse Glorious Model D-
And to consider Intel was joking at AMD one time for 'glue'... tsk tsk
Well, that's marketing! If you hear "go for them" from the higher-ups, the slides trashing your competitors is what everyone get to see haha
 
Joined
Sep 5, 2023
Messages
422 (0.90/day)
Location
USA
System Name Dark Palimpsest
Processor Intel i9 13900k with Optimus Foundation Block
Motherboard EVGA z690 Classified
Cooling MO-RA3 420mm Custom Loop
Memory G.Skill 6000CL30, 64GB
Video Card(s) Nvidia 4090 FE with Heatkiller Block
Storage 3 NVMe SSDs, 2TB-each, plus a SATA SSD
Display(s) Gigabyte FO32U2P (32" QD-OLED) , Asus ProArt PA248QV (24")
Case Be quiet! Dark Base Pro 900
Audio Device(s) Logitech G Pro X
Power Supply Be quiet! Straight Power 12 1200W
Mouse Logitech G502 X
Keyboard GMMK Pro + Numpad
I do think it was smart to spread out the P-cores. It's going to be an interesting trade-off as thermal density continues to increase. While the parts should get more efficient, they want to keep pushing performance, so they're not reducing power much (mostly intel), and that means more heat coming from a smaller area. It doesn't really matter how good your cooler is if the heat can't get out of the die and into the IHS.

So spreading out the cores is a good idea here...but I'm sure that can come with penalties in latency due to layout complexities. They'll have to continue weighing up whether they want to push the cores harder and get the heat out, or make things as compact as possible.

Side note, this is why I looked at the 11th gen before it launched and had a hunch it would overclock well from a thermal perspective (which was correct). It was going to be easier to get the heat out of the larger die-area:

This idea doesn't translate as easy with Arrow lake as it's on a new process and the tile layout complicates things further so we don't yet know (other than leaks/rumors) how much power the N3B node can handle. Rumors indicate it can't do much more clock speed on P-cores. Plenty of e-core headroom to get Cinebench improvements from, but that's not my favorite game lol. I'm curious how easy all the other "new features" are going to be to figure out for people who don't want to spend months fine-tuning their OC. We'll have to see the impact of die-to-die OC, memory OC, "dual BCLK tuning", etc.
 
Last edited:
Joined
Mar 7, 2011
Messages
4,614 (0.92/day)
Filler tiles, a nonaligned mass of chips stuck to a substrate... man this looks like a hack job more so than a nice design

And to consider Intel was joking at AMD one time for 'glue'... tsk tsk
Even better is the fact that US taxpayers paid Intel billions to get these new chips made in Taiwan.
 
Joined
Sep 17, 2014
Messages
22,638 (6.04/day)
Location
The Washing Machine
System Name Tiny the White Yeti
Processor 7800X3D
Motherboard MSI MAG Mortar b650m wifi
Cooling CPU: Thermalright Peerless Assassin / Case: Phanteks T30-120 x3
Memory 32GB Corsair Vengeance 30CL6000
Video Card(s) ASRock RX7900XT Phantom Gaming
Storage Lexar NM790 4TB + Samsung 850 EVO 1TB + Samsung 980 1TB + Crucial BX100 250GB
Display(s) Gigabyte G34QWC (3440x1440)
Case Lian Li A3 mATX White
Audio Device(s) Harman Kardon AVR137 + 2.1
Power Supply EVGA Supernova G2 750W
Mouse Steelseries Aerox 5
Keyboard Lenovo Thinkpad Trackpoint II
VR HMD HD 420 - Green Edition ;)
Software W11 IoT Enterprise LTSC
Benchmark Scores Over 9000
Even better is the fact that US taxpayers paid Intel billions to get these new chips made in Taiwan.
The world is a village they say... :laugh:
 
Joined
Dec 31, 2020
Messages
999 (0.69/day)
Processor E5-4627 v4
Motherboard VEINEDA X99
Memory 32 GB
Video Card(s) 2080 Ti
Storage NE-512
Display(s) G27Q
Case DAOTECH X9
Power Supply SF450
Which is fine and nicely tucked under the TIM and IHS. it's just that the Pcores moved further down, off-center and closer together, needing a corrective bracket. So much for spreading. MCH moved to soc for added latency. There are 3 PCIE root complexes, two on CPU substrate one on motherboard. Two gpu tiles one holds media engine and display engine one holds separately the Xe cores, so much complexity brilliant, yes nobel prize.
 
Joined
Mar 11, 2024
Messages
79 (0.28/day)
I wonder if they could have used the filler tiles as L4 cache or something. It would drive the cost up sure but filler come on
 
Joined
Dec 31, 2020
Messages
999 (0.69/day)
Processor E5-4627 v4
Motherboard VEINEDA X99
Memory 32 GB
Video Card(s) 2080 Ti
Storage NE-512
Display(s) G27Q
Case DAOTECH X9
Power Supply SF450
I wonder what he did there by soldering the capacitor on JNR. Must be some overclocker's kitchen secret.
 
Joined
Jan 3, 2021
Messages
3,586 (2.48/day)
Location
Slovenia
Processor i5-6600K
Motherboard Asus Z170A
Cooling some cheap Cooler Master Hyper 103 or similar
Memory 16GB DDR4-2400
Video Card(s) IGP
Storage Samsung 850 EVO 250GB
Display(s) 2x Oldell 24" 1920x1200
Case Bitfenix Nova white windowless non-mesh
Audio Device(s) E-mu 1212m PCI
Power Supply Seasonic G-360
Mouse Logitech Marble trackball, never had a mouse
Keyboard Key Tronic KT2000, no Win key because 1994
Software Oldwin
Which is fine and nicely tucked under the TIM and IHS. it's just that the Pcores moved further down, off-center and closer together, needing a corrective bracket. So much for spreading. MCH moved to soc for added latency. There are 3 PCIE root complexes, two on CPU substrate one on motherboard. Two gpu tiles one holds media engine and display engine one holds separately the Xe cores, so much complexity brilliant, yes nobel prize.
And all the added complexity is just because Intel can't get enough mm² of sweet N3 silicon. They only get some when Apple allows them to.

Granted, "all" is a very strong word in the above sentence but ... that seven-lithic complexity is costly. It's not there to save a lot of money. N5 and N6 chiplets still cost a lot (price per transistor matters, not per mm²). Foveros bonding process can't be cheap, either. All in all, cost wise, it doesn't seem considerably cheaper than building a monolithic CPU. Or maybe, if N3 yields are still unacceptable, they could also choose to split the design in two chiplets, without the base tile. The AMD way basically but with a faster interconnect.
 
Joined
Jan 11, 2022
Messages
909 (0.85/day)
And all the added complexity is just because Intel can't get enough mm² of sweet N3 silicon. They only get some when Apple allows them to.

Granted, "all" is a very strong word in the above sentence but ... that seven-lithic complexity is costly. It's not there to save a lot of money. N5 and N6 chiplets still cost a lot (price per transistor matters, not per mm²). Foveros bonding process can't be cheap, either. All in all, cost wise, it doesn't seem considerably cheaper than building a monolithic CPU. Or maybe, if N3 yields are still unacceptable, they could also choose to split the design in two chiplets, without the base tile. The AMD way basically but with a faster interconnect.
Or n3b is still a low yield product and keeping tiles small ensures some yield
 
Joined
Feb 3, 2017
Messages
3,806 (1.32/day)
Processor Ryzen 7800X3D
Motherboard ROG STRIX B650E-F GAMING WIFI
Memory 2x16GB G.Skill Flare X5 DDR5-6000 CL36 (F5-6000J3636F16GX2-FX5)
Video Card(s) INNO3D GeForce RTX™ 4070 Ti SUPER TWIN X2
Storage 2TB Samsung 980 PRO, 4TB WD Black SN850X
Display(s) 42" LG C2 OLED, 27" ASUS PG279Q
Case Thermaltake Core P5
Power Supply Fractal Design Ion+ Platinum 760W
Mouse Corsair Dark Core RGB Pro SE
Keyboard Corsair K100 RGB
VR HMD HTC Vive Cosmos
And all the added complexity is just because Intel can't get enough mm² of sweet N3 silicon. They only get some when Apple allows them to.

Granted, "all" is a very strong word in the above sentence but ... that seven-lithic complexity is costly. It's not there to save a lot of money. N5 and N6 chiplets still cost a lot (price per transistor matters, not per mm²). Foveros bonding process can't be cheap, either. All in all, cost wise, it doesn't seem considerably cheaper than building a monolithic CPU. Or maybe, if N3 yields are still unacceptable, they could also choose to split the design in two chiplets, without the base tile. The AMD way basically but with a faster interconnect.
Density of anything but logic is not scaling down these days. RAM is not getting much - if at all - smaller on N3 compared to doing it on N6. Scaling of anything IO and PHY was brought out as problematic back when N7 was new. Doing all this on a smaller node is just a waste. Illustrated here by the IO and SOC dies and the separation of stuff to these. AMD is doing the same with IO Die manufactured on relatively older process.

What has a cost for Arrow Lake is the 22nm interposer and packaging on top of that. On the other hand 22nm should be dirt cheap today and Intel definitely can manufacture 22nm very effectively. I bet running data over silicon has quite a sizeable efficiency advantage over running the same lines over substrate though.
 
Joined
Jan 3, 2021
Messages
3,586 (2.48/day)
Location
Slovenia
Processor i5-6600K
Motherboard Asus Z170A
Cooling some cheap Cooler Master Hyper 103 or similar
Memory 16GB DDR4-2400
Video Card(s) IGP
Storage Samsung 850 EVO 250GB
Display(s) 2x Oldell 24" 1920x1200
Case Bitfenix Nova white windowless non-mesh
Audio Device(s) E-mu 1212m PCI
Power Supply Seasonic G-360
Mouse Logitech Marble trackball, never had a mouse
Keyboard Key Tronic KT2000, no Win key because 1994
Software Oldwin
Density of anything but logic is not scaling down these days. RAM is not getting much - if at all - smaller on N3 compared to doing it on N6. Scaling of anything IO and PHY was brought out as problematic back when N7 was new. Doing all this on a smaller node is just a waste. Illustrated here by the IO and SOC dies and the separation of stuff to these. AMD is doing the same with IO Die manufactured on relatively older process.

What has a cost for Arrow Lake is the 22nm interposer and packaging on top of that. On the other hand 22nm should be dirt cheap today and Intel definitely can manufacture 22nm very effectively. I bet running data over silicon has quite a sizeable efficiency advantage over running the same lines over substrate though.
Clearly, one can think of many reasons for more dies, and many reasons against more dies. For the two-chiplet solution I suggested however, a N3 + N6 combination would be more logical than N3 + N3.

Another thought ... the 22nm base die would actually be a suitable place for additional L3 cache, exactly because SRAM scaling is quite bad in the latest fab nodes. The SRAM cell size would be about 0.12 um² (Intel 22nm) against 0.02 um² (N3B). It's a simple structure compared to logic, requiring few metal layers. ~18 MB of cache would fit in a surface area 3x as large as the 36 MB of L3 on the compute die. That would leave enough space for all other signal and power wiring. Depending on how cheap 22nm is, and considering that L3 is not a big source of heat, it might even be economically justified.

Or n3b is still a low yield product and keeping tiles small ensures some yield
Yes. Low yields bring about two huge problems: high price and low availability. Here I'm just trying to guess which one is more huger for Intel at the moment.
 
Top