ZOMG LOOK DIS PROCESSOR HAZ 19027831097589263874523094235094762948374523094283905867 CORES! *fires up video game, 4 cores being used in a very inefficient matter*
I only care for people not to exaggerate massively fanboying. It used to be intel, it's amd now. It can affect sales and online comentators do hold a lot of power there, lately masking cheap speculation as journalism.
@@backupplan6058In my headcanon Wendell secretly has access to a Tesla teleporter/cloner from The Prestige and he's pulling a magic trick on all of us, sans the water suicide tank
With Lunar Lake looking to be decent, Xeon finally getting competitive again, and Arrow Lake rumored to be far more efficient than RPL it looks like we're finally seeing the fruits of not having an accountant at the helm. Maybe they'd return to the HEDT realm to reign in some of that Threadripper pricing.
Maybe after Nova Lake Core Ultra on desktop shall be segmented into all-E core configs and all-P-core only configs with HT again, targeting different workloads. P-core-only needs dual CPU tiles on regular Core Ultra 7/9 up to 16C32T, shall free up 24C48T ones in Xeon w (still tiled, more CPU tiles and one SOC-IO-PCH tile as a HEDT/mainstream workstation SOC, no additional chipset needed on motherboard) This time all-P-core-only shall target absolute gaming and content creation performance with optional Foveros Direct wafer on wafer stacked L3 cache on top of CPU tile proper as Intel indicates going forward
@@Zapdos0145 Aaaaaand nothing whatsoever I was talking about has squat to do with Raptor Lake, but go ahead and bash Intel. After all, it's the favorite past time of the AMD fanboy.
Underestimated! "Let's make 3834 SKUs differentiated by very specific set of super niche features, but our top SKUs are focused on something no one cares about" - Intel, probably, since 2012.
It must be some sort of psy-op by Intel. Lure some of us back, give us access to the accelerators, then next gen, BAM, back to the delusions of paying extra for another meg of cache here, maybe even by the month for QAT there.....they're up to something.....
@@rich1051414 I doubt it. The accelerators are a key differentiator from the competition. If Intel wants them widely adopted, they need to be present on all their chips, so developers can experiment with them, and software vendors can rely on them. Similarly with IO: Epyc is winning designs where PCIe lanes matter, even on lower core count parts. 88 isn't quite the number as SP6 offers, but it's close enough. 64 can be given to NVMe or GPUs, with 16 for high speed networking, with 8 left over system drives for instance. Intel isn't ahead of AMD right now, they've only become competitive again with this release, and AMD isn't holding still. Plus there is ARM competition. They have to get to the point where their only competition is themselves again before they can pull those shenanigans again and that won't happen any time soon.
@@rich1051414we'll see, he's only been in the top job for a couple years and all the products up to now would have been at some part of their design pipeline when he joined as CEO. This could just be a return to having an engineer CEO and the product that entails.
I like Wendell's optimism about any piece of tech he get his hands on. Everywhere else those CPUs would be presented a lot more pessimistically "cause Epyc is better". And even if it's true, those Xeons look quite appealing and have some pretty nice features. Yay competition! 👏
Normally would agree, but Intel putting accelerators on all of the chips and not trying to sell them twice and/or at a monthly subscription or some other stupid, stupid, accountant scheme, really is something more akin to Scrooge waking up and finding out he can still make sure Tiny Tim doesn't die.
The Forests are better in Performance per watt, so they shall compare apples to apples in benchmark that includes watts in equation like score/W, score/1kW.
For us and our datacenter, it's all about consistent performance and power consumption per rack unit. Multiply times the number of racks and that's your capacity, base operating cost, and how much you need to forecast expanding to more racks. Preferably, we use fewer racks with less power draw, which can result in savings of tens of thousands per rack per year.
As a former employee at a hyper scaler I have to say: Power to cooling requirements are the most important. Space or power are concerns as well but potential for DCs are measured in capacity of cooling actually
Look at that! Sane use of tiling, instead of the insanity of building the same die but flipped for Sapphire Rapids. It is still astonishing to me that Sapphire Rapids even shipped.
To be fair to SPR, that design was the most optimal way to have the EMIB connections all face each other. Why completely redesign a die when you can just mirror it? That being said, EMR is a noticeable improvement with the larger tiles and less interconnect power losses.
Is it too much to hope for Xeon 6 HEDT? With Epyc Genoa-x 9184X giving 40MB of 3d-vcache/L3 per core, I’m disappointed in Threadripper offerings. Maybe if Xeon 6 HEDT enters the market, Zen5 Threadripper could be more… epic.
Intel still struggles to compete with threadripper both in price and performance (I know that I should be bringing up Epyc, but it's threadrippers and ARM-processors that are dominating Intel rn).
There are so many homelabbers hoping for something like that from either Intel or AMD. Don't need a ton of cores but ECC support and a decent number of PCIE lanes and low power, yes please!
I know I'm speaking from ignorance, but as soon as Intel released Alderlake my impression was that what was truly interesting and innovative was the possibility of deploying many "e" cores. It is impressive how far atom architecture can be taken.
Intel is back and hitting hard. It took awhile to turn the ship around but Intel is looking great for 2024 and beyond, especially if they deliver on the expectations that they will match or surpass TSMC nodes.
You should make a diy superchip. All p-cores in one slot, all e-cores in the other slot. No clue what to do with it, but having a whole machine mimic a mixed core CPU just sounds fun
@@mddunlap03 If you casually ignore the 288 Core SKU that is coming on the 12 channel platform, yeah. But I believe you'd have to have pretty selective memory to make such a statement.
Maybe this is easy to find out, but what makes an e-core different from a p-core? Is it there is less cache and less branch prediction? I know branch prediction can cause a lot of heat.
I've always been intrigued with Intel Xeon. Any word on what the street price for hardware like board and CPUs will run? My dual 22 core Xeons are looking a bit long in the teeth these days, though 1TB of DDR4 RAM makes up for it.
E1.S can be up to 8 lanes. I just don't know if any drives are available with that many. Also many are still PCIe 3.0 or 3.1. Just imagine if they did have 8 lanes and were PCIe 4 or 5 how quick they could be. 8 would be double a typical M.2 slot so there are a lot of possibilities there. If you ever saw E1.S it looks a little bit like M.2 except it is not directly compatible with an M.2 slot without an adapter. I have seen E1.S adapters for PCIe slots and they are very similar to the ones you can get to use an M.2 drive in regular PCIe slot. Of course either adapter card has minimal components on it since they mostly just allow the PCIe slot to be used with those drives. It seems that you can get an adapter for nearly anything these days and they are usually not expensive. If you have an E1.L drive known as the ruler form factor that is just a longer E1.S drive.
The 64 core and 96core E core cpus cost around $4.1k each. Those X14 motherboards from supermicro cost around $600 for the single socket and $1.1k for the dual node motherboards. That was the price for the X13motherboard. The rest of the chassie components are around 1.5k usually. +$ per ram stick Around $7k -$8k for the system.
The n100 is such a cool homelab cpu. I’d love to see a 32 core e core only cpu on the deaktop. If it was only 2-3 watts per core you could cool it to only a single socket with a tower cooler.
One emerging problem with 2S systems in cloud is network bandwidth per vCore. CSPs like 2S systems because the $/core is better than 1S because you amortize more shared resources (case, PSU, boot drive, etc.) over more cores but the customer perf/core is better with 1S systems.
Doth mine eyes deceive me? Did they just put all the accelerators on all the chips? Where is Intel and what did you do to make it act like it wanted to provide a product their clients want and not simply what they should be satisfied with having the honor of paying Intel for?!!
is it normal to have such a narrow IO (height dimension but wider in X-AXIS) die on XEON 6th? On the surface comparison, it seems that AMD IO is a centralised piece of die with identical X and Y axis measurement?
Heh, that Task Manager shot of 288 CPU cores on a two socket machine - I remember when the biggest shared memory multiprocessor box you could buy was a 64 (unicore) processor Starfire E10000 (simultaneously, the largest x86 SMP box was a Unisys ES7000 with 16 processors, later upgraded to support 32). One of the other potential uses for these processors is certain types of DB workloads (especially graphDB’s) where you get the query engine, on deep traversal, generating a lot of recursive queries with dependent data. In those scenarios, more cores across fewer NUMA nodes is advantageous. NUMA node affinitization and cross node performance issues are still a big optimization problem in large memory, high-concurrency workloads. I’d really like to see them also do a KnightsLanding style “near” memory (possibly as a large L4) of 16-32GB in package, though I suppose a “son of Phi” is a lot to ask (as I was probably one of five people who actually had a useful workload for the later Xeon Phi processors).
It's nice to see competition in the server space again after Epyc had been embarrassing Xeon on pretty well every metric for over 5 years now. Annoyingly, the Dell Poweredge range was still skewed heavily towards Xeons during that period, despite being significantly worse value. It makes me think Intel were throwing large bags of money to Dell to keep their range 90% Xeons (I'd often see large discounts on Dell Xeons only to keep them competitive - Intel Development Fund money or something?). Xeon 6 looks very promising, but let's wait for the Zen 5 Epycs before deciding who is the 2024 server winner - I'm just glad it's a close race again because customers win when there's strong competition. As Wendell said, it does make picking a "side" trickier depending on your workload. It might come down to pricing if the perf/power consumption is close between the two.
@Level1Techs Do you heard something about the Intel HEDT/WS platform's refresh for LGA4677? In February there was news that new LGA4677 processors would be announced. Maybe you heared something at Now Computex 2024? Intel told about new Xeon 6 platform with new socket but it is servers segment not WS and mamy people are waiting for something new for LGA4677. Is there still a chance we'll see something new this summer for 4677?
That's a weird thing to say. Though multithreading well really was not common back then (and frankly, it's still lacking in some areas to this day, almost 20 years later). And to give the teacher credit, if 4 core high-end variants would be out on the market, you can basically do all the basic and medium-sized needs on that. A 4 P-core at 6GHz from Intel or one at 5+ GHz with X3D cache from AMD should be able to reach very near top FPS in most games. For basic things like browsing and Office, that will be taken care no problem too. And programming can be done as well, though in some circumstances it will show its limitations. Heck, except Chromium, which takes ages, all other programs compile in decent (for me) times on my almost-8-year-old i7 6700HQ, running Gentoo. Last time Firefox compiled in exactly 1 hour. Not sure about video/audio/media work, but I suspect it's going to be good there too.
@@User9681e Yeah, true! But also, I think people today have taken a taste of what things like machine learning can do. And we also have the promise of quantum computing, so maybe we kinda know that no amount of computing "power" can ever be enough! We could probably simulate the entire human brain with enough computing power! And make true AI
@@-_Nuke_- there is already supercomputer with the same processing power but it's not just processing power it's how our mind processes information About quantum you need liquid nitrogen so zero consumer computing About ML I think that it uses different instructions from my understanding not every ML runs on the npu some require more complex operations Btw why people call ML as AI lol
@@User9681e That's why I said that you get AI if you manage to simulate the human brain's complexity in the amount of neural connection it has. Right now we are nowhere near that. And yea, maybe not even quantum computing will be enough for true AI
Every node name is just pr since the 16nm days. Almost nothing in a 3nm chip is actually only 3nm across. Intel3 is at similar density to N3 or N4 depending on the libraries used. Intel's node renaming was to better match up with the competitive densities offered by other foundries, as they previously used their own different system.
Intel is cooking good again, i'm getting excited. might get a simple 6 P-core CPU for my workloads after i'm done with AM4 (might be a while), the E-core scheduling is good but just a layer of complexity i don't need the pricing is just not competitive though, Ryzen 5 5600G vs Core 5 12400, the latter is 50% more expensive
7:45 I see the GPD Win Max 2. Can't wait for GPD to get their hands on the rumored 40CU Strix Point APU, can GPD build a heat sink good enough for a 120w TDP? I hope they try.
They might have to make something the size of a regular laptop, but I'd bet they still try to build a controller into it somehow. Maybe deplayable joy-con style things that slot into the sides?
Intel will over price this massively. There is a duopoloy between AMD and Intel, there will never be real competition and they will price fix the market and buy out competition indefinitely.
The noisy neighbor issue (one vCPU consuming more than it's fair share of memory bandwidth, or other shared resource) is a big issue in public cloud (Azure, AWS, GCP) so it looks like Intel's solution is a little superior in this regard.
If thier all e-cores then is it an e-core based CPU. if every one is sick then no one is normal. If nothing is more performative on die then it's just a more "efficient" Xeon version right???
If you don't have many light workloads or one that parallelises good/easily, then you wouldn't buy one of these anyway. I didn't got to see which type of E-core it is, but could it be Skymont ? The same that's in Lunar Lake ? That would make it somewhat on par with Raptor Cove IPC.
It is pretty insane where we are with core counts and all the fabrics and interconnects and hell, throw in some secret sauce... so much complexity for customers to determine the best optimized systems for their workloads.
When Intel started this big little cores I was sceptical, but owning a i3 12100 and a N100 with the Ecore being over 50% power saving the N100 is a lot cheaper to run and it is no my always on PC/Home lad. Also E-core can run passively on small core counts too. The E-cores and more of them is the way to go with energy costs effecting earnings. I watching more out of interest of what might make it to the mini micro PC's and Hackers boards.
Maybe, if you can get the board or server chassis cheap too? These new xeons will 20x to 50x the price of that cpu plus DDR5 is still double the price of DDR4
@@reubenmitchell5269 Thanks. Yeah, the price of mobo + RAM + custom heatsink/cooling will def. increase the price rapidly. Looking to upgrade my Skylake Xeon (Xeon W 2140). Need more cores, and would like a bit more futureproof as well. One option is buying a new Xeon W (up to 18 cores). But worried that it won't last long. Doing more and more VM's because of IT operations studies, and would like my server to be more performant. Was thinking about Epyc Milan with 24/32 cores as well, but didn't find a good deal for ATX/E-ATX mobo + CPU so far. Could reuse the RAM (128GB) from my current setup then, but would need a new cooler.
Highly parallel CPUs used to be popular for supercomputers as you just need to run x86 code. Matlab, for example uses x86 by default and it's a pain to code existing scripts to use Cuda. It could run software ray tracing I guess. I think their advantage, compared to GPUs is that they can run massively parallel non numeric operations pretty quickly. GPUs are still slower with that.
This would be a beast for software ray tracing. Almost definitely not fast enough for gaming though. But Cinebench and Blender are basically just software ray tracing.
I will never forget the words of Intel CEO when he took the stage with Steve Jobs to introduce Intel Macs: “we are committed to the relentless pursuit of Moore’s Law.” That was peak Intel and it would have been hard to imagine that just a decade later Intel would be years behind the competition.
If it replaces at a 3:! ratio of performance per watt than you get 200% more rather than 66% more. Or to say it another way it is 300% of the performance of the baseline, whereas 66% more would be 166% of the performance of the baseline.
Reminds me of the Phi series of cpus... a LOT of cores but the cores themselves only as powerful as atom cpus. Interested to see how this plays out for Intel in the long run. Effeciency is pretty important in data centers so not sure how I feel about this.
It still wonders me who this is for. Either you are a hyperscaler, someone who owns their own software stack or use opensource software to afford the potential licensing costs per core.
144 cores, Microsoft Licensing Department Must be salivating so hard now
Broadcom VMware too
Most servers actually run on Linux
@@chenyansong really depends on the enterprise. Also, you still have to pay for your linux support
@@Eugensson You'd be a fool to run Vmware on this lmao. That's not what this is for.
@@MrHav1k I'd be curious to know why 🤔
Crazy that 144 cores are pulling less than the 14900k
Intel considers 14900K excessive power consumption as feature for the customer to keep them warm in case of a blizzard!😂
If your 14900K is pulling 1000 watts, you might have a problem...
Arrow lake should correct some if not all of Intel's power problems with a 3nm compute tile node.
There is a big difference between a $400 processor and a $11500 processor.
@@vomitkermit3446yeah. the depth of the customers report wallet.
I don’t care what people are hating on, remember ‘competition is good’. I love that both Intel and AMD are going at each other as hard as they can.
ZOMG LOOK DIS PROCESSOR HAZ 19027831097589263874523094235094762948374523094283905867 CORES! *fires up video game, 4 cores being used in a very inefficient matter*
@@TheeGlocktopus yes, computer performance doesnt really matter if the software running on it is unoptimized and runs bad nonetheless
When intel and AMD compete hard, we win. Big. Regardless of which side we buy.
I only care for people not to exaggerate massively fanboying.
It used to be intel, it's amd now.
It can affect sales and online comentators do hold a lot of power there, lately masking cheap speculation as journalism.
@@tteqhu Literally all this comment section is full of frothing at the mouth AMD fanboys.
Wow, wendell flew back from computex to the studio and back to taipei just to show us Xeon 6 launch. Even Taylor swift can't beat that.
Clearly one is the real Wendell and the others AI Wendell. There had to be a contingency in the event it was a Boeing flight.
His real name is Twoendell, the "T" is silent.
@@backupplan6058In my headcanon Wendell secretly has access to a Tesla teleporter/cloner from The Prestige and he's pulling a magic trick on all of us, sans the water suicide tank
@@backupplan6058 ooooofff... double whammy!! AI Wendell (the world needs one) and a 🔥Boeing joke. Well done, my dude.
With Lunar Lake looking to be decent, Xeon finally getting competitive again, and Arrow Lake rumored to be far more efficient than RPL it looks like we're finally seeing the fruits of not having an accountant at the helm. Maybe they'd return to the HEDT realm to reign in some of that Threadripper pricing.
Right on bro!
Maybe after Nova Lake Core Ultra on desktop shall be segmented into all-E core configs and all-P-core only configs with HT again, targeting different workloads. P-core-only needs dual CPU tiles on regular Core Ultra 7/9 up to 16C32T, shall free up 24C48T ones in Xeon w (still tiled, more CPU tiles and one SOC-IO-PCH tile as a HEDT/mainstream workstation SOC, no additional chipset needed on motherboard) This time all-P-core-only shall target absolute gaming and content creation performance with optional Foveros Direct wafer on wafer stacked L3 cache on top of CPU tile proper as Intel indicates going forward
Completely agree.
aaaaand 14900ks are degrading. it was good while it lasted
@@Zapdos0145 Aaaaaand nothing whatsoever I was talking about has squat to do with Raptor Lake, but go ahead and bash Intel. After all, it's the favorite past time of the AMD fanboy.
I appreciate that Intel is no longer hypersegmenting functionality.
Underestimated! "Let's make 3834 SKUs differentiated by very specific set of super niche features, but our top SKUs are focused on something no one cares about" - Intel, probably, since 2012.
It must be some sort of psy-op by Intel. Lure some of us back, give us access to the accelerators, then next gen, BAM, back to the delusions of paying extra for another meg of cache here, maybe even by the month for QAT there.....they're up to something.....
@@FroggyTWrite I have a feeling this is a single generation thing to regain some market share, and then back to business as usual.
@@rich1051414 I doubt it. The accelerators are a key differentiator from the competition. If Intel wants them widely adopted, they need to be present on all their chips, so developers can experiment with them, and software vendors can rely on them.
Similarly with IO: Epyc is winning designs where PCIe lanes matter, even on lower core count parts. 88 isn't quite the number as SP6 offers, but it's close enough. 64 can be given to NVMe or GPUs, with 16 for high speed networking, with 8 left over system drives for instance.
Intel isn't ahead of AMD right now, they've only become competitive again with this release, and AMD isn't holding still. Plus there is ARM competition. They have to get to the point where their only competition is themselves again before they can pull those shenanigans again and that won't happen any time soon.
@@rich1051414we'll see, he's only been in the top job for a couple years and all the products up to now would have been at some part of their design pipeline when he joined as CEO.
This could just be a return to having an engineer CEO and the product that entails.
10 years ago i wouldnt have even dreamed of this level of efficiency
It's only because arm was brought into servers so x86 been in a race making simpler cores
I like Wendell's optimism about any piece of tech he get his hands on. Everywhere else those CPUs would be presented a lot more pessimistically "cause Epyc is better". And even if it's true, those Xeons look quite appealing and have some pretty nice features. Yay competition! 👏
Normally would agree, but Intel putting accelerators on all of the chips and not trying to sell them twice and/or at a monthly subscription or some other stupid, stupid, accountant scheme, really is something more akin to Scrooge waking up and finding out he can still make sure Tiny Tim doesn't die.
Good point I agree 👍
The Forests are better in Performance per watt, so they shall compare apples to apples in benchmark that includes watts in equation like score/W, score/1kW.
For us and our datacenter, it's all about consistent performance and power consumption per rack unit. Multiply times the number of racks and that's your capacity, base operating cost, and how much you need to forecast expanding to more racks. Preferably, we use fewer racks with less power draw, which can result in savings of tens of thousands per rack per year.
As a former employee at a hyper scaler I have to say:
Power to cooling requirements are the most important.
Space or power are concerns as well but potential for DCs are measured in capacity of cooling actually
That perf per watt is incredible and this looks to be a real beast on that front. Looks like Intel is back!!
All mention of “E” cores should have a voice over of mini-me saying “E”.
Look at that! Sane use of tiling, instead of the insanity of building the same die but flipped for Sapphire Rapids. It is still astonishing to me that Sapphire Rapids even shipped.
To be fair to SPR, that design was the most optimal way to have the EMIB connections all face each other. Why completely redesign a die when you can just mirror it? That being said, EMR is a noticeable improvement with the larger tiles and less interconnect power losses.
@@DigitalJedibut they had to use two different mirrored dies instead of the same die mirrored and rotated 4 times.
144 cores is pretty nice at about ~330 watts only with this performance.
😂
I think aws would love it in their ec2 configs.
Now, imagine clearwater forest on 18A with the new improved ecores! 😯
4:42 Wendell: "...there are off roadmap SKUs..."
My gearhead ears: "...there are off-road map SKUs..."
Let me remind you of XKCD 37
Thanks Dude for testing the docker usage case, nice and helpful information. Cheers
144 E-cores would make one heck of a NUC! 😁
All we need are Xeon compatible itx boards! Most people would say "why?" I would say "want!"
@@yensteel +1
@@yensteelASRock did a couple for Epyc, we can hope!
As someone who worked on GNR/SRF validation, I can't wait to watch your review!
Is it too much to hope for Xeon 6 HEDT? With Epyc Genoa-x 9184X giving 40MB of 3d-vcache/L3 per core, I’m disappointed in Threadripper offerings. Maybe if Xeon 6 HEDT enters the market, Zen5 Threadripper could be more… epic.
Nice to see Intel back in the game and it will be interesting to see what AMD comes up with in the future.
They already have Turin.
I mean AMD could drop Zen 5c
This is competing with 12 month old Bergamo and even then its a toss up depending on workload.
192 zen5c, not 256
Intel still struggles to compete with threadripper both in price and performance (I know that I should be bringing up Epyc, but it's threadrippers and ARM-processors that are dominating Intel rn).
would love to see a prosumer version of this oops all E-cores platform on an itx board to replace the old avaton/denverton stuff
There are so many homelabbers hoping for something like that from either Intel or AMD. Don't need a ton of cores but ECC support and a decent number of PCIE lanes and low power, yes please!
32 ecore with 2 pcores...
Exactly, I was really looking forward to this release, hoping for 12-32 e-cores, ECC, qsv/igpu, ipmi/vPro :/
I know I'm speaking from ignorance, but as soon as Intel released Alderlake my impression was that what was truly interesting and innovative was the possibility of deploying many "e" cores. It is impressive how far atom architecture can be taken.
I am curious of how 144 cores render a blender.
These the e cores same as lunarlake e cores or meteorlake?
Closer to meteor lake than lunar but improvements since meteor
@@Level1Techs cool thx for answering!
Intel is back and hitting hard. It took awhile to turn the ship around but Intel is looking great for 2024 and beyond, especially if they deliver on the expectations that they will match or surpass TSMC nodes.
Lunar Lake, their new mobile chips are actually produced on TSMC nodes.
You should make a diy superchip. All p-cores in one slot, all e-cores in the other slot. No clue what to do with it, but having a whole machine mimic a mixed core CPU just sounds fun
I’m really excited to see Intel being competitive, let the accelerator wars begin!
I mean it's a toss up depending on workload vs amd last Gen? LOL let's see how this ages in the next 6 months
@@mddunlap03 If you casually ignore the 288 Core SKU that is coming on the 12 channel platform, yeah. But I believe you'd have to have pretty selective memory to make such a statement.
30% ppw is kinda huge. this is impressive, and i say that as a ryzen fanboy
Maybe this is easy to find out, but what makes an e-core different from a p-core? Is it there is less cache and less branch prediction? I know branch prediction can cause a lot of heat.
Excuse me Wendell, did you say that you can use two different CPU's in one INTEL socket???
wait what, I missed the news that hell froze over
Price?
But will there ever be a consumer solution to using these like past xeon CPUs?
I've always been intrigued with Intel Xeon. Any word on what the street price for hardware like board and CPUs will run? My dual 22 core Xeons are looking a bit long in the teeth these days, though 1TB of DDR4 RAM makes up for it.
Jesus Christ I wish I had that
E1.S can be up to 8 lanes. I just don't know if any drives are available with that many. Also many are still PCIe 3.0 or 3.1. Just imagine if they did have 8 lanes and were PCIe 4 or 5 how quick they could be. 8 would be double a typical M.2 slot so there are a lot of possibilities there. If you ever saw E1.S it looks a little bit like M.2 except it is not directly compatible with an M.2 slot without an adapter. I have seen E1.S adapters for PCIe slots and they are very similar to the ones you can get to use an M.2 drive in regular PCIe slot. Of course either adapter card has minimal components on it since they mostly just allow the PCIe slot to be used with those drives. It seems that you can get an adapter for nearly anything these days and they are usually not expensive. If you have an E1.L drive known as the ruler form factor that is just a longer E1.S drive.
Competition benefits consumers, yes! You have to love competition.
I really would like to see an all e core SOC with at 55w it can be 16 core or 32 cores. It needs to be a BGA and supports ecc.
Hey, this actually looks pretty good! Not quite at parity with Turin-Dense (or even really Bergamo) but this at the very least a compelling offering
How much does the system you showcased here cost?
The 64 core and 96core E core cpus cost around $4.1k each.
Those X14 motherboards from supermicro cost around $600 for the single socket and $1.1k for the dual node motherboards. That was the price for the X13motherboard.
The rest of the chassie components are around 1.5k usually.
+$ per ram stick
Around $7k -$8k for the system.
Q: who makes ECC server memory with RGB lighting ?
Can you mix a p core cpu und an e core cpu on a 2socket system?
I don't believe you can as that would cause some interesting task sharing situations, but it would be really cool to see done.
The n100 is such a cool homelab cpu. I’d love to see a 32 core e core only cpu on the deaktop. If it was only 2-3 watts per core you could cool it to only a single socket with a tower cooler.
One emerging problem with 2S systems in cloud is network bandwidth per vCore. CSPs like 2S systems because the $/core is better than 1S because you amortize more shared resources (case, PSU, boot drive, etc.) over more cores but the customer perf/core is better with 1S systems.
Doth mine eyes deceive me? Did they just put all the accelerators on all the chips? Where is Intel and what did you do to make it act like it wanted to provide a product their clients want and not simply what they should be satisfied with having the honor of paying Intel for?!!
Being deep in the red and on the highway to broke does wonderful things to both companies and people :D
is it normal to have such a narrow IO (height dimension but wider in X-AXIS) die on XEON 6th? On the surface comparison, it seems that AMD IO is a centralised piece of die with identical X and Y axis measurement?
I need a baby 32 core version of this, with 60+ lanes... I don't have a need for that level of compute, or power consumption.
Heh, that Task Manager shot of 288 CPU cores on a two socket machine - I remember when the biggest shared memory multiprocessor box you could buy was a 64 (unicore) processor Starfire E10000 (simultaneously, the largest x86 SMP box was a Unisys ES7000 with 16 processors, later upgraded to support 32).
One of the other potential uses for these processors is certain types of DB workloads (especially graphDB’s) where you get the query engine, on deep traversal, generating a lot of recursive queries with dependent data. In those scenarios, more cores across fewer NUMA nodes is advantageous. NUMA node affinitization and cross node performance issues are still a big optimization problem in large memory, high-concurrency workloads.
I’d really like to see them also do a KnightsLanding style “near” memory (possibly as a large L4) of 16-32GB in package, though I suppose a “son of Phi” is a lot to ask (as I was probably one of five people who actually had a useful workload for the later Xeon Phi processors).
Something akin to their Xeon Max lineup with HBM2 memory on package? I think it was 64GB.
@@morosis82 Yep, something along those lines.
It's nice to see competition in the server space again after Epyc had been embarrassing Xeon on pretty well every metric for over 5 years now. Annoyingly, the Dell Poweredge range was still skewed heavily towards Xeons during that period, despite being significantly worse value. It makes me think Intel were throwing large bags of money to Dell to keep their range 90% Xeons (I'd often see large discounts on Dell Xeons only to keep them competitive - Intel Development Fund money or something?).
Xeon 6 looks very promising, but let's wait for the Zen 5 Epycs before deciding who is the 2024 server winner - I'm just glad it's a close race again because customers win when there's strong competition. As Wendell said, it does make picking a "side" trickier depending on your workload. It might come down to pricing if the perf/power consumption is close between the two.
Sounds like Intel 6 will be an optimal platform for Evernode! Too bad I am running all of mine on the best 2016 had to offer (E5-2697V4). lol
The last 2 minutes gave me goosbumps! I can't wait for all the old servers to swamp the used market!!!
Microsoft Licensing (Windows & SQL Server) is based on CPU's core count. Intel do not have any powerfull Xeon 6 with something like 32 cores?
Do we know prices on these chips yet? and pricing on the motherboards that would be needed!
Yes, the prices are online, the 64 and 96 core cpus are around 4k each
Why not run Bergamo at a lower cTDP, and then make the comparisons?
@Level1Techs Do you heard something about the Intel HEDT/WS platform's refresh for LGA4677? In February there was news that new LGA4677 processors would be announced. Maybe you heared something at Now Computex 2024? Intel told about new Xeon 6 platform with new socket but it is servers segment not WS and mamy people are waiting for something new for LGA4677. Is there still a chance we'll see something new this summer for 4677?
I remember our programming teacher when the 1st quad core processor came out, that we will never need more than 4 cores... xD
That's a weird thing to say. Though multithreading well really was not common back then (and frankly, it's still lacking in some areas to this day, almost 20 years later).
And to give the teacher credit, if 4 core high-end variants would be out on the market, you can basically do all the basic and medium-sized needs on that. A 4 P-core at 6GHz from Intel or one at 5+ GHz with X3D cache from AMD should be able to reach very near top FPS in most games. For basic things like browsing and Office, that will be taken care no problem too. And programming can be done as well, though in some circumstances it will show its limitations. Heck, except Chromium, which takes ages, all other programs compile in decent (for me) times on my almost-8-year-old i7 6700HQ, running Gentoo. Last time Firefox compiled in exactly 1 hour. Not sure about video/audio/media work, but I suspect it's going to be good there too.
Today he would have said we wouldn't need more then 100 core CPU
And let's see how that goes
@@User9681e Yeah, true! But also, I think people today have taken a taste of what things like machine learning can do. And we also have the promise of quantum computing, so maybe we kinda know that no amount of computing "power" can ever be enough! We could probably simulate the entire human brain with enough computing power! And make true AI
@@-_Nuke_- there is already supercomputer with the same processing power but it's not just processing power it's how our mind processes information
About quantum you need liquid nitrogen so zero consumer computing
About ML I think that it uses different instructions from my understanding not every ML runs on the npu some require more complex operations
Btw why people call ML as AI lol
@@User9681e That's why I said that you get AI if you manage to simulate the human brain's complexity in the amount of neural connection it has. Right now we are nowhere near that. And yea, maybe not even quantum computing will be enough for true AI
Does anyone know if the Xeon 6 e core uses skymont and the p-core model uses Lioncove??
uses crestmont plus cores fyi meteorlake used normal crestmont cores
It uses Crestmont Plus, which is slightly more powerful than Crestmont, between Crestmont and Skymont in performance.
@@auritro3903 what about the xeon 6 p cores is that using redwood cove plus?
@@auritro3903 what about the xeon p-core variant does that use redcove plus?
The Xeon 6710E with 64 cores, 4 QAT and 4 DLB would make an insanely good VPN Server for only 205W
Very impressive. I wish there was more innovation in the SMB market.
I just want a small all e core CPU to use as a home lab virtual machine host. This gives me hope in a few years i can pick one up off flea bay.
G'day Wendell,
Watching MEGA Core Count video on my Athlon 200GE 😁👍
is "intel3" lithography actually 3nm or is this just PR?
I thought that it was actually being fabbed at TSMC 3nm, and packaged at Intel. Backpane from Intel too
@@deansmits006 intel spent tens of billions to be competitive with tsmc that's why im curious
Every node name is just pr since the 16nm days. Almost nothing in a 3nm chip is actually only 3nm across. Intel3 is at similar density to N3 or N4 depending on the libraries used. Intel's node renaming was to better match up with the competitive densities offered by other foundries, as they previously used their own different system.
@@deansmits006 These are on Intel's in house node. No TSMC here. This is fully Intel 3.
It's comparable to n3 from tsmc in performance atleast
They all are marketing speak anyway no actual hardware feature is at 3nm physically
Intel is cooking good again, i'm getting excited. might get a simple 6 P-core CPU for my workloads after i'm done with AM4 (might be a while), the E-core scheduling is good but just a layer of complexity i don't need
the pricing is just not competitive though, Ryzen 5 5600G vs Core 5 12400, the latter is 50% more expensive
7:45 I see the GPD Win Max 2.
Can't wait for GPD to get their hands on the rumored 40CU Strix Point APU, can GPD build a heat sink good enough for a 120w TDP? I hope they try.
They might have to make something the size of a regular laptop, but I'd bet they still try to build a controller into it somehow. Maybe deplayable joy-con style things that slot into the sides?
@@DigitalJedi Ideally it will just be thicker to accommodate bigger fans and more copper, and at the same time that would make room for more IO Ports!
Man people are going to hate trying to license Windows Server for these...
So what will the pricing of the platform be like
Hooray for competition; maybe this leads to prices getting a bit better for customers over the next few years instead of increasing.
Intel will over price this massively. There is a duopoloy between AMD and Intel, there will never be real competition and they will price fix the market and buy out competition indefinitely.
Give me P-cores clocked to the moon, no HT, and in-CPU NIC bypass(UDP preferably), and we're talking.
The noisy neighbor issue (one vCPU consuming more than it's fair share of memory bandwidth, or other shared resource) is a big issue in public cloud (Azure, AWS, GCP) so it looks like Intel's solution is a little superior in this regard.
If thier all e-cores then is it an e-core based CPU. if every one is sick then no one is normal. If nothing is more performative on die then it's just a more "efficient" Xeon version right???
Intel is back!
Does it crash out of box running all cores?
Let's hope not, Intel can't afford to create more broken CPUs
price?
Not all programs can be multithreaded. How is the IPC?
If you don't have many light workloads or one that parallelises good/easily, then you wouldn't buy one of these anyway.
I didn't got to see which type of E-core it is, but could it be Skymont ? The same that's in Lunar Lake ? That would make it somewhat on par with Raptor Cove IPC.
It seems to be all about energy efficieny and TOPS at the moment...
Will there be any good overclocking options on those CPUs?
It is pretty insane where we are with core counts and all the fabrics and interconnects and hell, throw in some secret sauce... so much complexity for customers to determine the best optimized systems for their workloads.
I wonder how much it costs to unlock the boost frequency after you bought the CPU
I want a 6701E with like 16 cores, a gpu, and 64-88 lanes. NO CXL needed, but sure. What i want is a N100 but with enough PCIE for a NVME NAS.
When Intel started this big little cores I was sceptical, but owning a i3 12100 and a N100 with the Ecore being over 50% power saving the N100 is a lot cheaper to run and it is no my always on PC/Home lad. Also E-core can run passively on small core counts too. The E-cores and more of them is the way to go with energy costs effecting earnings. I watching more out of interest of what might make it to the mini micro PC's and Hackers boards.
I have an e-core only CPU, an N100 with a whopping four of those bad boys at 3.4 Ghz base. It's in a BeeLink mini, and it's a very fun little beast.
144 cores? Oracle Licencing Department salivating.
There is one thing that intel is really great at. They are the industry leader at converting energy to heat in the processor market.
Is Sapphire Rappids ES worth it? Found 48core CPU for 150 dollars. Seemed like a real good deal for homelab (coming from Skylake Xeon).
Maybe, if you can get the board or server chassis cheap too? These new xeons will 20x to 50x the price of that cpu plus DDR5 is still double the price of DDR4
@@reubenmitchell5269 Thanks. Yeah, the price of mobo + RAM + custom heatsink/cooling will def. increase the price rapidly. Looking to upgrade my Skylake Xeon (Xeon W 2140). Need more cores, and would like a bit more futureproof as well. One option is buying a new Xeon W (up to 18 cores). But worried that it won't last long. Doing more and more VM's because of IT operations studies, and would like my server to be more performant.
Was thinking about Epyc Milan with 24/32 cores as well, but didn't find a good deal for ATX/E-ATX mobo + CPU so far. Could reuse the RAM (128GB) from my current setup then, but would need a new cooler.
Thanks Wendell!
how many teraflops, compared to gpu
Are you sure it's not a gpu? I wonder how it would handle sortware ray tracing
Highly parallel CPUs used to be popular for supercomputers as you just need to run x86 code. Matlab, for example uses x86 by default and it's a pain to code existing scripts to use Cuda.
It could run software ray tracing I guess. I think their advantage, compared to GPUs is that they can run massively parallel non numeric operations pretty quickly. GPUs are still slower with that.
This would be a beast for software ray tracing. Almost definitely not fast enough for gaming though. But Cinebench and Blender are basically just software ray tracing.
I will never forget the words of Intel CEO when he took the stage with Steve Jobs to introduce Intel Macs: “we are committed to the relentless pursuit of Moore’s Law.” That was peak Intel and it would have been hard to imagine that just a decade later Intel would be years behind the competition.
But can iy run crysis? (on software mode)
If it replaces at a 3:! ratio of performance per watt than you get 200% more rather than 66% more. Or to say it another way it is 300% of the performance of the baseline, whereas 66% more would be 166% of the performance of the baseline.
Reminds me of the Phi series of cpus... a LOT of cores but the cores themselves only as powerful as atom cpus. Interested to see how this plays out for Intel in the long run. Effeciency is pretty important in data centers so not sure how I feel about this.
This is probably the evolution of those.
Yeah! It was used in the Tianhe 2, the no. 1 fastest supercomputer in 2013.
Excited for this tech to move down into the professional space.
Haswell microcode fix when?
I ❤ Taiwan.
It still wonders me who this is for. Either you are a hyperscaler, someone who owns their own software stack or use opensource software to afford the potential licensing costs per core.
Wendel is the best
They have to bring hybrid Alder Lake style to xeons.
Bring back the cool background music.
I think CWF with chadmont plus is the one to watch. That will dominate for sure.
E-cores blow!!! IF you want to calculate serious data, you need P-cores assuming they match or beet previous real processors.
They should make a dual socket system where you can mix P-core CPUs and E-Core CPUs,. Now that would be dope, the best of both worlds
Does this mean ARM is dead Jim?
Nah but basically Intel is making x86 into arm