Meteor Lake - Can Intel leapfrog AMD?
Vložit
- čas přidán 20. 05. 2024
- By clicking my link www.piavpn.com/HighYield get 83% discount on Private Internet Access! That's just $2.03 a month, and also get 4 extra months completely for free!
Join me on a deep-dive into the architecture of Meteor Lake. We will discuss Intel's Foveros 3D stacking technology, the new Tile Based Design and explore the secrets of Meteor Lake, including Redwood Cove and Crestmont Intel 4 CPU cores.
Support me on Patreon: www.patreon.com/user?u=46978634
Follow me on Twitter: / highyieldyt
0:00 Intro
1:57 Part 1: The Foundation / monolithic vs modular
4:25 Private Internet Access
6:10 Intel Foveros 3D Stacking
7:48 Part 2: The Tiles / Meteor Lake overview
9:45 CPU Tile / Redwood Cove + Crestmont
11:41 GPU Tile / Intel Xe iGPU
13:02 I/O Tile
13:35 SoC Tile / Intel AI VPU + LP Crestmont
16:44 Base Tile / Interposer / Adamantine L4 Cache
21:02 Part 3: The Strategy / Intel VS AMD - Věda a technologie
By clicking my link www.piavpn.com/HighYield get 83% discount on Private Internet Access! That's just $2.03 a month, and also get 4 extra months completely for free!
Edit: the GPU tile info might be wrong. Apparently there is/was a 192EU tile produced in TSMCs N3B, but the smaller tiles are N5. Let's see if we will get Meteor Lake SKUs with the large 3nm GPU tile or not.
It's difficult to argue for a general maximum acceptable power draw for a GPU.
But I know that I don't want a space heater, and I don't want to waste energy. So I'm all for efficiency-focused designs.
Most reviews of this chip show that its power draw is at least twice as much as AMD. That's very bad for mobile devices.
czcams.com/video/Ig1P1v7pyB8/video.html
Energy efficiency improvements are very useful for desktop chips, because reaching 350W is unacceptable for a GPU, let alone a CPU
Keep in mind that end result efficiency and power draw is still heavily based on how hard a chip is being pushed. Yes, all the connections and everything may be highly power efficient techniques, but if they're still having to clock the cores to the gills in order to be performance competitive on desktop, then it's going to lose much of its efficiency potential at load. That said, the move to Intel 4 will finally put themselves on par with Zen 4 in terms of process, which will help a lot in direct comparisons(at least for MTL and laptops). A lot will depend on just how quickly AMD move to Zen 5 and then again how quickly Intel can move to 20A in 2024 as promised(I'm doubtful).
There's still efficiency benefits in a product like a 4090 if it can complete a set of task like 3d rendering much faster
Higher performance at set power
Gpus have way bigger dies
@@RayanMADAO true and yet it's unacceptable. GPUs should max out at 225-230W and CPUs at 140-150W like they used to for a very long time.
I just want to note that Meteor Lake is not Intel’s first chiplet design. It’s basically Lakefield 2.0 with high-performance designs. Sapphire Rapids is also selling today with 2 different tiles, that combine in a 4 tile design. That’s 2020, and 2022 technology respectively.
As for why the IO die may exist and not be directly integrated in the SOC tile, is probably for reusability. Mobile and desktop processors have different IO requirements, so by separating it from the SOC tile Intel doesn’t have to design another SOC tile. It’s cheaper to design a smaller and simpler processor than modify a more complex processor. As for why Intel is using separate IO tiles, probably for power since not having elements is more efficient than power-gating them.
Also zen1 had chiplets aswell (epyc). Zen2 just took it to a larger level.
He also got it wrong on the AMD side, zen1 threadripper was AMD's 1st chiplet architecture, not zen2 ..
pentium d was intels first chiplet desktop cpu. AMD had just come out with the opteron 165 that was monolithic and said intel "glued" their cpus together.
@@dgillies5420 Magny Cours was AMD's first chiplet CPU.
Intel has also had several years of experience with EMIB connected tiles on their FPGAs. They also build the Kaby Lake G chips with in-package HBM and GPU tiles.
It'll be interesting to see if they manage to pull it off on time and at scale.
Coming this year in October. May be they unveal something in computex 2023.
It's already late. Meteor Lake was originally slated for H1 of 2023. But sometime earlier last year, it was clearly getting delayed til later. I'm confident it will come this year, but yea, Intel still is not able to execute on time with basically anything these days. They've gotten better since the 10nm debacle, but that shouldn't be a high bar to clear...
On time is the biggie. Their constant ability to not release products on time or just not release them at all has fucked them in the last 5 years. Where the hell is battlemage?
@@How23497
Battlmage was always planned to release in Q1 to Q2 2024.
Wait for Q2 2024.
The additional CPU cores in the SOC tile are interesting, not just for energy efficiency reasons, but because it could allow some programming of SOC behaviour - i.e. implementing QoS or using them as additional accelerators that transparantly handle memory compression or encryption.
efficient compression/decompression or any accelerator for the matter, requires specialized hardware, not weak low power generic CPU cores. those two cores are there to barely show the desktop and run some background services imho.
The idea seems compelling enough that perhaps AMD should add a couple of their own LP cores to the IO-die. They would have to design them first, of course. But if they manage to use them to be able to disable the infinity fabric between the io die and the core complex, perhaps they could greatly improve their idle power consumption. I think that the only market in which would make a large difference is on their desktop replacement laptops, but it would still be a good thing for everyone but servers perhaps.
@@marsovac AMD can offload SEV and memory encryption to the tiny ARM core in the SOC chip that handles Trustzone and platform initialization - and yes, that tiny core got an accelerated fast path for AES.
Doing the same with an x86 core and making sure that a select few instructions can run fast, while resorting to microcode emulation for others, is standard practice by now. Besides - at least accelerated cryptography is needed anyways due do to always-on/always-connected nature of new devices.
Its a good day whenever a new High Yield video is out. I love learning new stuff with you! I really hope Intel execute their chiplet strategy properly and see some more competition, especially in the high core count space where AMD are basically unrivaled at the moment.
I agree heavily, everything was a efficiency choice because performance is great at the moment and no need for that added complexity just yet. My eyes are still on meteor lake regardless of canceled desktops parts, the architecture has been a long time coming and it’s with great privilege to see it in action soon. I think the L4 is going to be a great addition, especially when arrow lake arrives.
I never expected such a deep dive analysis from you! Even MLID never provided anything ike that, and he calls himself an analyst! you're a true analyst here
MLID is a hack, expand your horizons.
MLID doesn't actually understand that much about hardware and rests heavily on having a less informed audience he can bullshit better.
@@maynardburger true, he was so random in the early days but when you do it for 4 years then you at least learn with during the process, he has had amazing guests and definitely improved his knowledge
@@Raven-lg7td He blankets everything, shills for AMD (6500xt shit), then no matter what he can basically go "well at one time I was right", because he literally parroted everything.
He’s mainly about spewing out info from good sources, which I appreciate. Been following since vega fe, and pretty much most things are on track
But in terms of analysis, it’s basically up to our own interpretation and research. He’s pretty harsh on his viewers, at one point quoting “that’s a stupid question”(tho it was kind of a bad question)
Still a lot better than gamermeld, basically a rumour mill.
I'm finally at a point when I can confidently say that I feel equally as indifferent about either of the two companies. Let's go competition
Previous engineers: lets not put south and north bridges under same heatsink as the cpu, that will be to hot.
Intel: lets integrate north bridge
Amd: lets make whole additional I/O die
Intel: I will then gather it all together with as small distance between them as possible
That's not exactly right. In the old days, the NB and SB were different chips because lithography was relatively primitive. Smaller, cheaper chips was the main drive. The NB was eventually integrated as the silicon budget granted by smaller processes allowed for it, but that was something done by AMD first back with the Athlon 64 launch in 2003. Then, AMD has been selling interposer based tile products since 2015. The story of AMD and Intel is an interesting one since AMD has long been doing a large part of the innovation that eventually becomes "standard" in people's minds once Intel does the same thing.
You earned yourself a subscription with this vid.
The GPU tile is TSMC N5. There were some Taiwanese media spreading the news that Intel somehow would be the 2nd customer after Apple to use N3 2 years ago. Then everyone thought the GPU would be N3
it is N4 not N5. Also i don't remember anything saying the gpu will be alchemist based. it could be the first outing of battlemage.
@@quantumdot7393 You're ALL making the same weird mistake - confidently claiming what it *will* be before we actually know anything for sure. You're all just using different rumors and thinking your rumor is the correct one for some reason. :/
Yes, but this is not the fault from Intel and don't count them, TSMC N3B was suppose to be ready last year, iPhone 14 suppose to use it but that is a problem, the media and the outlet will not say there is anything to see here in TSMC, TSMC always on time.
If TSMC N3 is available last year and Apple will not have 100% of capacity this year, so GPU was planned to use N3 (maybe).
@@maynardburger Meteor lake GPU using N5 is confirmed by Intel staff in the Intel vision event last year. It is always N5.
When Intel first shared slides about Meteor Lake the slides show that they would use an external Fab for the GPU Tile and didn't even say it was TSMC they just put 3nm. The same slides confirmed Meteor Lake was on Intel 4 and Arrow Lake on Intel 20A. It is Intel who made everyone speculate that the GPU would be based on TSMC 3N. TSMC have had major problems with their latest advance nodes so it 8s not just Intel who struggle with manufacturing. If Intel stuck with TSMC 3N then Meteor Lake would not be coming out on December 2023 (really Q1 2024). Apple M3 is still yet to be announced for the same reason.
Quite interesting analysis. Intel could very well stop the bleeding in the laptop market with this IF it comes out on time. That's been the big question with Intel over the past four or five years: Will they get it done in time?
The delay in MTL laptops is due to the complexity of integrating new things while a bit extra time can work it out. It turns the extra time is about two months. On the other hand, many delays in the past is due to some bottlenecks that can quagmire for a long time. I now have much more confidence on Intel's forecasts and plannings. Pat made a huge difference.
Arrow Lake has a lot of interesting potential
AMD doesn't use PCIe for chiplet to chiplet comms. It's a specialized infinity fabric. It may be similar to how PCIe works, but it's not connecting chiplets via the same PCIe bus that goes out to the MB. It's independent.
The first point you talk about, particularly improvements within the cores are better called microarchitectural changes, architectural changes are generally at a higher level and encompass things such as AMDs Infinity Fabric and partitioned L3, or Intel's evolution from linear buses in the first 3 gens of cores to the mesh topologies they use on higher core count parts today. Architectural design also encompases ISAs and instruction types like SIMD, whereas microarchitecture involves *how* those instructions are implemented in the core.
This might be your best video yet. GJ high yield!
love the breakdown! As we have seen in SPR, complexity isnt conducive to a timely roll out. Hopefully they have got things figured out now and can produce at scale. Will be interesting to see the power and performance available!
Thanks! Very informative, well explained!
if Intel can finally change their stance on motherboard longevity (probably won t because of their relation with their partners) I will gladly go check them, but right now with AMD you can just drop an X3d on a old motherboard and get like twice the perf that you got at the start even with the same amount of core
This is only for mobile. Desktop just gets a refresh of the current chips, which clocks 100-200 MHz higher.
Not even close. A 5800x3d is maybe 15% better than a 5800x.
@@lostpacket It's atleast 15% usually, sometimes it's 50% or more - look to well optimised simulation games like factorio where x3d chips are dominating across generations
@@aravindpallippara1577 the x3d is only faster at gaming than the intel chips. productivity and multicore workload is now owned by intel 13th gen.
@@iequalsnoob Not necessarily, the 13900k has the crown of about 10 to 15% lead in multicore with 50% more power draw
You can't use that power without having insane cooling to go with it, which is why 7950x is still the better actual processor for sustained multiore workload, since it doesn't get thermal throttled by an nhd15 or the average 240mm cooler, especially where I live with an ambient temperature of 35C to 38C
The intel p core single thread is very very good, but I mostly want that for gaming, where there is x3D, and for multicore I want an nhd15 coolable cpu, which isn't 13900k
This seems like a fascinating change not only for laptops but also potentially for the new portable handheld segment
Great video. It explained the technical details nicely, very few channels do this and just talk about buzz words.
You ideas make SO MUCH SENSE!!! I think you hit the nail on the head!!!
Your vids are so informative!
This video got me excited, this will probably be better than any presentation Intel will give form their PR team. subscribed!
I'm also excited about how fast Intel is transitioning from FinFet transistors to Gate All Around transistors.
Same here. Intel 20A has 2 times more transistor density than tsmc 3nm.
So excited for all this, especially arrow lake, which might be my next upgrade. I wonder if they have considered multiple cpu tiles for scaling to more cores on desktop. And what about server/workstation? Are they going tiled there too?
Excellent Video. Very interesting. Thank you an regards from Germany.
Meteor lake Sounds promising but like the i7 G design with Vega on the same chip what will be the cost of production. I’m worried intel has made a complex chip to compete with Amd/apples monolithic small mobile chip.
All these chips are complex and Intel is on the second and soon third generation of foveros. What once seemed complex will seem ordinary
Are there software strategy (videos) that aligns to the silicon strategies of these chip competitors?
Meteor Lake sounds great if they deliver on these claims. May be my next laptop.
Meteor Lake appears to be incredibly impressive. I can understand why they're opting for a name change after such a long period of time. However, it seems to be quite costly and complex, which may lead to even more delays. Considering it will be competing against Zen5, it remains uncertain whether this will be Intel's best moment.
Pretty sure it can beat zen5. Amd's strength is x3d and efficiency due to chiplet + newer node. Intel is doing similar thing by adding a cache tile and is moving up in node.
@@yuvanraj2271 There is no way it can beat 16 Zen5 cores in performance. Strix point is even rumored to have a zen 5 and a zen 4d chiplet. So you get 8+16 cores.
It'll be good for the mobile sector, but it fell short for desktop which is why we'll be getting a raptor lake refresh as the desktop 14th gen. It won't be until Arrowlake that the desktop sector gets the new architecture.
@@Psi-Storm Pretty sure, it can. Raptor lake already beat zen 4 in pure performance. Both gaming and productivity. The x3d is what made zen 4 to still get a win in cache specific games. This is with 10nm. The efficiency can be attributed by node difference. With their own version of x3d, Intel can sure win in performance in both gaming and productivity.
@@yuvanraj2271 "if we dismiss this new awesome technology from our competitor then we win" lol. Also, RPL doesn't beat AMD even without 3D cache, it trades blows in gaming and productivity while using double the energy. Stop being a fanboy.
Very cool analysis and an excellent channel, thx! Meteor Lake is clear improvement for Intel especially with the 3D packaging and possibility of even including cache in the active interposer. Also a capability of turning off tiles which aren't needed in certain cases while keeping SOC with the 2x E-cores active is also something very cool to see in action.
Intel is clearly looking forward by improving its 3D stacking/chip packaging technology and not just chasing lowest possible nm. I still wonder how Meteor lake SoC would perform per watt in comparison to the Apple's M3 if it would be completely made of the latest 3nm tiles.
This was a very interesting clip, thank you. In my opinion, Intel returned to inovation with gen12, which will be stepped up by the tile technology. Having big+small cores was quite a big step ahead, but I cannot wait to see the first laptop chips of gen14. My gen12 laptop will be here for a few more years, but the future looks great.
Most times, analog I/O and Serdes designs on leading edge nodes gets done later thanks to the complexity of a new node - and there may not be any area savings at all thanks to having to support the same operating voltage for external communication protocols. Therefore, selecting the I/O tile such that it contains well proven I/O makes sense from an engineering time perspective
I/O doesn't scale with process node, as well as many other parts, as long as there is no performance penalty in moving elements off x86 core to auxiliary elements then it's worth doing it
Seriously an underrated channel....great videos!!
coming from a z170 system with a overclocked i7 7700k i now using a z690 system with a i5 12600kf@5p/4eGHz. i like the Adler Lake chip very much. as i saw the weird behavior in taskmanager with parking cores and with that literally the Thread Director switching Hyper Threading on and off for the first time i had to Google it because all my other chips never did anything like that. and with a fresh installation of W11(23H2) the whole thing runs very nice. i think Meteor Lake and everything what follows will be amazing.
Probably E-LP cores are on the SoC so it has a low power mode in which you can disable the other cores, and thus you don't need to spend energy to maintain that part of the interconnect until you decide to wake it up.
Problem is I heard Meteor Lake is only for mobile, not desktop. Also, all that 3d stacking stuff is gonna cost a fortune. So Meteor Lake is probably meant to be premium chips for premium laptops only.
For desktop Intel will release a refresh of Raptor Lake later this year, and I believe they're gonna release Arrow Lake later next year or early 2025
That’s true, Meteor Lake is Laptop only
well, i am listening, as always i was stuck on the idea of using Intel as it was fastest I switched to AMD, 7600 ryzen 5 and a 6800 NON-XT this is a powerhouse, I will compare it to this! GL!!
This is a quite excellent rumor analysis. I'm excited to try out MTL systems, honestly, after all these power-hungry years. One note: 16:42, the thread counts don't match up. 16C: 6P + 8E + 2LE. But then, SMT is only on Redwood, no?
SMT hyper threading seems deprecated and on it's way out.
great video. Thanks.
Wow, great job! Its crazy to think Intel not even using their own fabs for their products these days.....
what a time to be an enthusiast...Intel vs AMD has never been so exciting, but Arrow Lake is even more exciting due to MLID
What is MLID?
@@sam-pd7su Moores Law is Dead youtube channel
@@Raven-lg7td Please stop listening to that guy. He doesn't really know anything.
Do keeperlake or seeperlake, I know friends from Chennai, they always suggest me these names , their relatives work for intel and they wish the projects to be named after them. Very talented individuals from top institutions , got several awards while being employed
What a time to be alive
My long story short:
***
Intel is cpu core, you are interconnect, while viewers(me) are "cache" 😅
Large heaps of customers can make things faster.
??? ----------------???
Not sure if Meteor lake would come out at whole lineup or just high end products of 14th gen and rest would be just rebranded or cut physically Raptor lake.???
If it gets just into laptop s, that would be little bit dissapointment for desktop hardcore fans!
It's nice to have translator, mediator for chip companies designs, which are sophisticated but very interesting.
That deep sleep another level of energy efficiency mode over current one could be interesting for desktop PCs too.
->
Having PC at never turn off would probably got rid off my already low 6s boot to Windows time. Nothing needed, but it would be logical
next step.
Thank you for this Video. Now I understand why the latency is there for AMD due to the parasitic capacitance that has to do with the silicon substrate. With foveros it somewhat avoids all the fun stuff since it is metal and much shorter paths. Less chance of noise, thank you for the images of the Zen2 pathways it almost looks like art 😊
I really hope so. We need Intel to come back and Smack AMD around a bit. More Competition is great for the Market. Unlike the GPU market.
Really hope Intel come back strong, we will all benefit from good competition between AMD and Intel
Yes, only AMD dumb fanboy will hate to see Intel coming back.
Are they gonna go into RiscV?
Great channel
I don’t think there is a chance that Adamantine is coming to Meteor Lake. All rumors are pointing towards Adamantine still being at least a year or two away.
if anyone knows the true reason (not official public statement) that each quad E-core tile has shared L2, I would like to know. Recall that in Conroe (65nm) and Penryn (45nm), 2 cores shared L2, but Nehalem client, each of four cores had dedicated L2. The reason stated then was that: two cores sharing L2 was not excessively complicated or burdensome (technically and in terms of latency), 4 cores with shared L2 would have been. What has changed?
One advantage of tiles is the ability to mix processes. The SoC is on a back process (22/16) because driving off chip signals involves buffering up the current (from 20nm signals at the lowest level on die to 0.1mm (100um or 100,000nm) bumps off chip. there is only limited scaling in trying to build SoC on a newer process while the older process is much cheaper?
The E-Cores are based off of Atom chips not the full processor family tree like the P-Cores.
I can’t believe this channel doesn’t have more subs. You make amazing content. Thank you
I really hope it ends up coming out for back-to-school like Moore's Law is Dead says he thinks it'll be, cuz i need a new laptop and holy crap meteorlake lookin nice
ML sounds expensive with all the packaging complexity to produce in large volume, especially for mid to lower skus (likely why its skipping desktop and there's no 8+x core options).
It also looks like a big boon for TSMC given they provide most of the tile silicon. How does this keep Intel's fabs full?
It would have been good to compare ML to AMDs Strix also, which it will more directly compete against given likely release timing.
Is it possible to do a video comparing Intel 3D stacking vs. what TSMC does for MI300? AMD seems to be saving the most expensive designs/packaging for the highest end product (huge accelerator) where pricing can more than cover the costs with huge margins.
very interesting, I want Intel to succeed so much and be back in the game. Because there are still much more Intel laptops available but their CPUs are so bad compared to AMD APUs, Meteor Lake i3 in $400 laptop that has iGPU equal to 1650 would be such needed revolution
If Meteor Lake performs really well, dont expect them to come in at budget prices. Intel will continue to produce older processor generation which might well cover more low end options. I think Meteor Lake is going to be more expensive to produce than many are thinking.
@@maynardburger you have a point, Intel is going for efficiency instead of cost like AMD, but we didn't get any info about the cost of this packaging...
Energy efficiency is one thing. But... what if Intels Battlemage gets integrated into the CPU! Sure this version may not be efficient nor small. But this chip may overthrow NVIDIA's monopoly!
*PLEEEEAASE* Intel compete with AMD Ryzen APU integrated graphics and give us laptop and desktop CPUs with GTX 1060 and 1080-class "ARC Xe" iGPUs 🙏🙏
Love your channel.
As for Meteorlate apu from Intel
Congratulations, Intel, 8 months late, but now you have a comparable apu to AMDs7000 series apus. Too bad hawk point will decimate your late to market meteorlate apu
Wasn't it already stated that Meteor Lake will omit models with 8 P-cores on desktop?
Stated where? Certainly not by Intel. So until then, it's all just rumors. Though this has seemed likely for a long time now, given Intel has literally only ever used the same 6c package when showing the Meteor Lake silicon(which they've been doing since 2021...).
It does look like they are going for the efficiency price point. AMD seems to be sticking with Lowe manufacturing cost and performance in gaming. I doubt they want to stay that way, but they prob know they can't come out the door with chiplets AND beat AMD in performance. They will prob only get a little more performance than their current gen if any, but if they can show a large efficiency jump, then it may be palatable to the market. The they can.work on getting performance. However, given what AMDs new/projected APUs are showing I have a problem we have only been seeing "prototypes" up til now(especially RDNA3). I am hopeful they will ride the momentum they already have and keep making an excellent product. Intel is going after an open market, though, as there are many smaller low power/batt powered devices that do not require good graphic capabilities and prefer low power usage while being active. Going to be weird for reviewers and the comparisons will be getting more "apples to oranges".
Anyway, thanks for the great update! I was not even aware of this. I knew I could count on you to keep me informed!!
Excellent analysis! Thank you for this video.
I like what Intel is doing, however I can still appreciate the simplicity in how AMD scales their chiplets, but Intel's approach will have benefits in the long run and by then I'd imagine AMD will simply copy Intel like they've historically done.
AMD is already there. Just look at the mi300. 6 cdna gpu chiplets, 3 zen 4 chiplets, io-die, a bunch of cache and hbm memory.
I mean, intel did copy AMD with their 64bit ISA
After all there is a reason it's called x86_64 or AMD_64 (reading out of the activity monitor in macbook)
Everybody copies everybody - that's how good ideas benefit everyone
@@Psi-StormAMD needs to implement this in their HX chips then, since their battery life is really suffering there because of it
I wonder what Intels logic behind the 2 extra efficient cores in a different tile is.
Google/Amazon voice commands still being handled whilst the laptop is in sleep mode like some laptops can already do but at even lower power usage? Or perhaps they're betting big on it and just disabling the entire CPU tile whenever the load is low enough and let those 2 cores handle things (probably enough power still for 99% of people for the majority of the time). Then the question becomes if they can move data around fast enough to handle tasks over to the CPU block when the power is needed.
That's also a surprising amount of TSMC produced die-space. They must have bad yields themselves but either the performance is great or they don't want an external company handling that intellectual property.
Another great video!
Shud be around arc a380 for the igpu
Lower than that is considered failure to catch-up considering AMD rdna4 atleast with radeon880M will target 3050/1660super
The cache in the base tile could yield some nice performance gains, esp. in gaming. When Broadwell was released with 128 MiB of eDRAM, it performed better than the following Skylake CPUs that didn't have L4 cache. Having a ton of data cached may also help with masking the ever increasing memory latency from new generations of throughput orientated RAM. I bet 0,01€ that it'll be a victim cache.
That L4 cache is still a rumor, by the way. I appreciate High Yield's videos, but he stated 'as fact' several things in the video which have not at all been confirmed, and are only hinted at or rumored. All we know is that there's potential for that base die to house cache with active silicon, not that it absolutely will in an actual consumer product. Personally, it sounds monstrously expensive.
Could you explain what the Meteor Lake chip design will mean to an average user who isn't that tech savvy? Will there be newer functions facilitated by this design?
The return of the north and south bridge 😱
9:10 ok, you blew my mind. How are they mixing and matching different foundries? Wtf? And nodes? OK.... Intel is ready to play baby. Let's go.
Going through reviews and noticing that people don't acknowledge the fact that this chip is done by multiple assembling lines in multiple nodes.
Redwood Cove and Crestmont seem very disappointing in terms of performance. It’s mostly equal to Raptor Cove and Goldmont.
The GPU is still TSMC N5, and the base tile is Intel 16.
MTL-U gets 64 execution units, while MTL-H gets 128 execution units.
It also turns out Intel shoved Thunderbolt 4 and PCIe 5.0 into the I/O die. MTL-U get PCIe 5.0 and half the Thunderbolt 4 axed.
The package shots you used for 6P+8E+2LP-E is actually MTL-U which is 2P+8E+2LP-E. The LP-E cores appear to just be for stuff like watching Netflix. No always-on stuff aside from the usual Modern Standby.
Adamantine cache appears to not be a thing and the base tile is supposedly completely inactive.
I'm 2 minutes in and it's my first time watching your content, you deserve a like 👍
Given the actual power efficiency delivered and disappointing performance even at the 35W level, with Phoenix not really bested despite the review samples having unusually fast expensive LPDDR memory this video calls out for a follow up.
Can Intel turn things around by ARL or are the Intel tile design over complicated and inefficient squandering the benefits of process node advantages? MI300 may also have had production issues.
Absolutely agree that Intel finally is on the right path to contend or pass AMD. In just a few years, Intel has managed / planned some big innovations in chip design. The 12th, 14th and 15th are major milestones. I especially like the idea of having 2 LPE cores in SOC tile in the MTL, that's what Intel and AMD should have done for a long time. The power savings will benefit a vast swath of laptop users. Cannot wait to get hold a 14th-gen laptop.
I am wondering about the lifespan of non monolithic Designs.
Considering that those xeons from 2010 where running heavy load nonstop for like 5 to 10 years and are still running fine today.
I really hope this will be all true, as intel needs a kick for a while, they were selling 2018 product for 5 years now. I hope tech to turn off "chiplet/tiles" when on low load, will be shared soltion for all players on market, as we could halve idle power usage for everyone, and that is worth chasing.
Intel's own "glued together" CPU architecture.
Has a lot of advantages like fewer waste
Zen1 threadripper was the start of AMD chiplets, not zen2 ...
good report thank you. mb
the L4 cache could be an FPGA specific this or that function . . . it may not be a buffer or cache buy function memory? mb
functional programmable memory . . . could b mb
tiles butt up against each other to prevent dice from sagging off this way or that from their interconnect that is different from AMD multi-chip module on fabric. The titles cant shift and the way they are packaged means to precent shift this way or that in production. mb
extra i/o cores = management processors; microcontrollers, Intel has always had management processors in Xeon. mb
prevent shift
Ok but when do we get 3d stacking that's more then 2 layers becouse 2 layers are still pretty flat
Im curious how they will deal with heat on such a dense chip not just horizontally but also vertically.
Heat will be somewhat mitigated by the fact that different parts of the package can be off or in a low power state when not in use. Long running all core workloads are just not that common except when benchmarking.
@@LouisDuran There is gaming of course the activity where a laptop actually has to put some effort into
Otherwise any damn chip will work for most of the stuff, look to fanless apple macbook air
Vertically there is very minimal heat distribution. All of the hot tiles are stacked on top and directly transfer heat into a IHS or cooler. The only tile trapped below is the base tile, which shouldn't have thermal issues. It's like the substrate used in processors with HBM attached to it.
My concern is that this will be very expensive, to the point where even a lower performing AMD counterpart is the better buy.
Can Meteor Lake come in the socket 1700!? Dont feel like buying a new motherboard....
Lol.... than don't buy intel
All those energy improvements are nice, but the main issue piwer consumption is P cores boosting up to 5GHz@1.3V every time there is some background activity and if scheduler moves those task across all cores fast enough then ALL cores will be idling at 5GHz and 1.3V.
What they need to do is come up with a plan to prevent this while still keeping boost clocks for tasks that really matter.
About standby power consumption: The solution is so easy: Get rid of that stupid "connected standby" on laptops. Nobody needs this stuff on a device that you do not carry in your pocket. What we had 10 years ago was perfectly fine.
Blame OEMs and Microsoft . S3 sleep is still in both Intel and AMD micrcode, but needs to be enabled as an option by OEMs, who were strong armed by Microsoft not to do so.
I was actually looking for info on meteor lake last few days! So happy that you had this video in store for us.
A few notes: GPU tile remains a mystery. Some sources (MLID, anandtech, sometimes Intel itself) claim it's Arc Battlemage, so next-gen architecture evolved from Arc Alchemist (which would make sense at TSMC 3nm), however other sources (common sense, Intel itself sometimes) claim it's Arc Alchemist, so the same microarch as the desktop GPUs, manufactured at TSMC 5nm.
It doesn't make sense that Alchemist would be manufactured at 3nm since it would take a huge effort to port it to the new node. And I find it a bit surprising that they would release Battlemage on laptops a whole year before Battlemage on desktops is out. Also, not even Apple has its 3nm products out yet, and we all know apple gets the newest node due to being the biggest friend of TSMC.
I am excited for actual deep dives into this product once it's available. CPUs are in one of the most exciting states they have been in years! Yen 5 will be equally amazing, and arrow lake and Zen 6 even more so, hopefully.
With CPU cores in the SOC die, I would guess that we will see low end chips (Atom, Celeron, whatever) that have only the SOC (and maybe GPU) dies.
Love your video ❤
Thank you!
Hey, do you have a german channel, too? :) thanks for the videos!
Sorry, hab leider nur den Kanal in englisch, für zwei mal filmen und schneiden habe ich leider keine Zeit :(
@@HighYield verständlich! :)
Wish Intel would totally embrace making RISC-V for the embedded microprocessor markets so there could be a reliable North American manufactured source of such chips. Is a huge problem to have so much of industry ultimately dependent upon all manner of lower tier chips as supplied by China.
Where does x86-64 go? ARM seems to be much more efficient. Even Microsoft is moving to ARM. All the relevant OSs will be running on ARM: MacOS, Android/Chrome OS , Linux and Windows. And what about upcoming Risc-V? Can AMD and Intel switch to their own Risc-V cores?
Will these meteor lake crush RDNA3 IGPU's power equal to 1060? In other words do we have a new APU competitor?
It's hard to say. If the large 192EU 3nm GPU tile actually does exist, it would certainly beat RDNA3 iGPUs like in Phoenix. Smaller 5nm tiles will compete well, but most likely not "crush" AMD:
@@HighYield Very exciting developments
thank you so much brother for ure usefull reply. appreciate it.@@blue-lu3iz
I seem to remember Intel's first chiplet based product, it was called Pentium II.
Nice.
I don't consider AMD to be superior to intel except for the process node advantage with tsmcs euv nodes. Intels products have better quality and better single thread performance than any amd product as of now. The only disadvantage of intels products is the process node disadvantage that causes leakage and power consumption. Power consumption is what limits the multi threaded performance. With the e cores intel have already covered the multi threaded performance disadvantage on desktop processors.
Too your point, if the 13900K was manufactured on TSMC 5nm, it would crush the 7950X in every regard.
There are some information might needed more clarification from you.
I don't think the GPU tile is going to be TSMC N3B I think it will be TSMC N5 so the amount of EU should be low end of the estimate in meteor lake, Apple is 100% on the fab, even if Intel wants it, they has to have it NOW, and I don't think this is the case, I think Intel Quantity is at least 20% of Apple (consider Apple make 100% from TSMC and Intel GPU title look very small to me).
Some funny things from MLID said that Arrow Lake will have a CPU tile version made by TSMC.
I am not sure how Intel will set up the Fab, but it sound like funny. As CPU tiles is not very big, I think is only 30% to what Raptor lake is full monolithic title so to achieve good yield is easier as the die is smaller. Anyway, MLID is bad.
Intel's PVC shows they can do 47 tile GPUs. They also have 408MB of L2 SRAM in PVC ... comparable to 3d V-cache. Systems withe these are being sold by SuperMicro. So, I don't understand any comments about AMD's tile capabilities being more advanced. In any case, Intel can also use TSM mfg if they choose.
what is the bus that connects the chiplet tiles? AMD has infini fabric. Nvidia has nvlink-c2c. so far I think nvlink-c2c is the future
Intel uses different protocols to connect the individual tiles.
CPU-SoS uses IDI, GPU-SoC uses iCXL and SoC-I/O uses IOSF