1. The Fastest CPU on the Planet – HOLY $H!T – EPYC Milan 7763 64 Core
  2. More Cores, More Better: AMD, Arm, and Intel in 2022-2023
  3. AMD EPYC 7003 Milan Performance, Features, and Intel Ice Lake/ Cooper Lake Competition
  4. Intel Xeon Ds Go-FAST Feature
  5. AMD EPYC Milan Is Better Than We Thought
  6. Crazy! AMDs Milan-X Delivers 1.5GB of L3 Cache to EPYC Servers
  7. AMD EPYC 9004 Genoa Gaps Intel Xeon with 96 Cores WOW

The Fastest CPU on the Planet – HOLY $H!T – EPYC Milan 7763 64 Core

this is the fastest cpu on the planet it,costs eight,thousand dollars has 64 cores and 128,threads 128 lanes of pci express gen 4,and runs on amds latest seven nanometer,zen 3 architecture thats the same one,powering their top of the line 5000,series ryzen gaming cpus its code name,is epic milan and,we have two of them thanks to our buddy,wendell hows it going were at level,one text lets take them for a spin,cause ill wait are they ever fast you,know what else is fast the way i segue,to my sponsor honey honey is the free to,use shopping tool that helps search for,some of the best promo codes on loads of,your favorite sites get it today at,joinhoney.com,ltt,hey this is the top of the line ls and,gs the epic 7763 that im holding in my,head yeah,it may only run at a mere 2.45 gigahertz,base 3.5 gigahertz boost but dont let,that fool you this aint no civic okay,this is the hummer with the cummins swap,all right,these things are huge i cant get over,it,why do they have to make them so big,its not a measuring contest just never,gets old you know i think the only way,to get a bigger one is to buy one on,ltteststore.com oh yeah by the way were,gonna have epic threadripper shaped ones,coming soon probably in a few months now,of course,yoda would happily remind all of us that,size,not important performance is what,matters and when it comes to performance,this is pretty much the finest display,of kicking someone while theyre down,that i think i have ever seen and i,dont even feel bad,heres the thing intels latest and,greatest 10 nanometer ice lake xeons,theyve got some platform advantages,particularly when theyre paired with,intels optane memory but for purely,cpu-bound applications i mean theyre,not even competitive with amds last,generation epic roam chips price for,price let alone the new ones because,heres the thing amd comes in with more,cores,more gigahertz more cash and literally,double the pci express lanes per socket,yet amd still prices their top dog,cheaper than intels top dog and you,might think well at eight thousand,dollars a chip 200 bucks thats a thats,a rounding error right but if youre,buying a thousand of them at a time like,intel thinks you are,sorry intel not good enough for you i,guess that works out to a quarter,million dollars not to mention that if,youre an enterprise or a data center,thats upgrading your systems from say,last gen on intel youre going to need a,whole new motherboard which for servers,generally means an entirely new system,complete with financing and warranty,about the warranty like if something,goes wrong either with the hardware or,software,amd by contrast allows you to simply,swap out the cpu update your bios,preferably in the other order and youre,good to go theyve actually been using,the exact same socket for their server,platforms ever since the original epic,launch back in 2017 and this,intercompatibility is fantastic you know,how i know because this,is the exact same motherboard that we,used to build our pc mac pro killer with,our last gen epic something to note,before we go any farther is that server,components are generally designed to be,run in an extremely high airflow,environment so,this cooling,if that doesnt look suitable for a,nearly 300 watt cpu,thats because its not we need to get,to get a fan on there lets just throw,our m.2 in here this is by the way the,greatest thing ever this little,tool-less m.2 retention mechanism,absolutely love it they should all have,that this supreme x 3090 is so hilarious,its so big weve never even tried to,integrate it into an actual build,its test bench use only now you might,notice that im installing it in one of,the bottom pci express slots normally i,would say,dont do that but epic has so many pci,express lanes that every one of these,16x slots all the m.2s all of it is just,running full fat pcie gen 4 with all the,lanes so it just doesnt matter okay,okay i shouldnt quite say it doesnt,matter there are situations where the,lanes might need to be reallocated and,theyre the different chiplets and where,theyre attached to the io die and all,that but,mostly its fine,oh there it goes,128 threads guys this is cinebench r23,this isnt the like weak,easy one,now,the reason that it was so easy tm for,amd to maintain compatibility across,generations is that under the hood the,layout of their milan cpus is nearly,identical to their older rom,predecessors but thats not to say that,amd didnt change anything like other,zen3 cpus each of the eight chiplets,here which each contains eight cores,gets a full 32 megs of cash that is,completely accessible to all eight of,the cores rather than being split into,16 megs and 16 megs that means more,efficient use of resources and lower,latency theyve also increased the speed,of the infinity fabric links within the,cpu and theyve made a ton of other,under the hood architectural changes,that have allowed amd to hit a,surprising 19,average ipc uplift that means that clock,for clock this thing is almost 20,faster oh and they upped the maximum tdp,for many of their chips including the,7763 to allow up to 280 watts per socket,i mean that is how you get stuff like,this happening look where it sits,relative to anything else in the,database,61 000,what so while the cpu is running a heavy,benchmark like this its self reporting,a peak of around 277,278 watts of power what that means is,that according to wendell from level 1,techs this tie-in board we have here is,actually a bit on the conservative side,for power delivery the daytona reference,chassis that weve got back here can,actually do closer to 285,200 290 watts per socket with some power,tweaks,which actually sounds a lot like amd has,some overclocking headroom here you can,even tweak some i o settings like,disabling pci express gen 4 or lowering,your memory speeds to give a bit more of,the cpus power budget to the cores,themselves potentially improving,performance but then again,slowing down your pci express lanes or,lowering your memory bandwidth can also,hurt performance in other ways so its,going to depend on your application,either way i cannot freaking wait to see,what zen 3 based threadripper is going,to be like especially with the core,clocks and the power limits unlocked as,we assume theyll be i suspect were,going to need some new motherboards with,new power delivery for those,now lets have some more fun last time,around with our roam based 64 core epic,blender was only able to use half of the,cores per instance so we ran it twice,with both of them running at the same,time finishing in around a minute and 14,seconds,this time around we are having no,trouble utilizing the entire cpu and you,can see i mean just look at this thing,it is absolutely,crushing this benchmark and the craziest,part is check out these boost clocks we,are seeing in excess of 2.8 gigahertz,even as high as 3.1 gigahertz on some of,these cores as its running through this,blender render,and whats interesting is our temps are,so well under control the maximum weve,seen so far after about 10 minutes of,cinebench running continuously and,blender is around 70 degrees,super reasonable,the fans arent even ramping up its,almost like having a gigantic cpu with,the cores all spread out helps to get,rid of the heat,almost if you cant wait for the next,threadripper and you decided youre,crazy enough to use this as a,workstation check this out im running,this at,half quality in the preview here this is,an 8k timeline with multiple 8k clips,obviously theres a bunch of lumetri,color being applied to it,look how smooth that is its butter,smooth can we make it all the way to,full man that scrubbing is still,extremely responsive,naturally our next order of business is,to play video games on it we got uh apex,legends running at a cool,144 fps average that sounds not too bad,maxed out at 4k i dont have any ammo,thats okay its not about being good at,the game its about having a good,computer to run the game right,so we got f1 2020 going here

More Cores, More Better: AMD, Arm, and Intel in 2022-2023

hey guys this is Patrick from sth and,today we are going to cover all that you,need to know about the major design,philosophies and ideologies of the big,chip makers for their server CPUs in,late 2022 to early 2023 or all the way,through 2023. you may have heard a lot,about arm server CPUs and even client,CPUs recently were going to talk about,what they are doing and what the vendors,that are building arm-based CPUs are,doing were also going to talk about,what Intel is doing because it is,completely different in a lot of ways,its actually the opposite of what arm,is doing or like what arm was trying to,do 10 plus years ago and so thats kind,of crazy and then were also going to,talk about where AMD falls on that,spectrum and then not only are we going,to talk about what the design,philosophies are and how those are going,to be implemented in different chips but,were also going to discuss why that,matters for even if youre buying kind,of the low run systems maybe youre not,even buying the top end you know maximum,core count systems but this new ideology,and these design philosophies are going,to impact you in a way that is very,different and from any of the server,like chip Generations that weve been,seeing for years if you grew up in the,Xeon E5 generation perhaps then this is,going to be completely different than,what youre used to because Intel is,going to be doing something completely,different and there are other players in,the market now long time sdh readers and,those of you on YouTube will know that a,lot of times during the summer things,slow down a little bit at sth because,there tend not to be a ton of product,launches over the summer but I do like,to take this break in our busy schedule,to do things like catch up on our,networking our tiny mini micro and Mini,PC series and then also I really do like,to go and do these kind of just thought,pieces this is definitely a discussion,piece so if you think of something feel,free to go put those in the comments,down below also feel free to just share,this piece with some of your co-workers,and talk about it over a zoom call over,coffee lunch whatever it is because the,last pieces that weve done like last,year as an example we did the gigabyte,era of CPUs and look where we are now we,have basically Milan X which is just shy,of that gigabyte of cash but were gonna,see with uh the Next Generation Genoa X,that will be in the gigabyte class and,then also Sapphire Rapids hbm we have,pictures on the sth main site with me,and that chip I mean that will be a,multi-gigabyte chip so you know that has,definitely come to fruition like the,video that we did on that we also did a,piece on the dpu ipu thing really early,and we talked about you know what the,differences are between them and all,that kind of stuff and now weve,actually started to test them and the,dpu IP Market is absolutely crazy hot,right now and so that has definitely,come true also cxl is just about to,start taking off and so thats about to,happen and so you know these are really,videos that I just like to get to do,theyre not really like super Hands-On,Hardware but on the other hand they are,Ultra important because frankly you need,to know about this stuff and before we,get too far on this I just want to let,you know that we dont really have a lot,of b-roll or anything like that for this,so this one is kind of more like a,podcast or something that you can just,throw in the background and if you want,to just leave it in the background and,listen to it thats kind of more of the,idea of this video its not necessarily,were not going to like show you any new,hardware that we havent shown you,before so with that I want to introduce,my framework for the next 18 to 24,months of servers CPUs between arm Intel,and AMD and Im going to make this not,proper English but we are going to say,that the arm is basically more cores,more better Intel is more accelerators,more better and AMD is kind of more,medium is more better okay so the first,one that I want to talk about is arm arm,has a lot of momentum right now and,youre going to see things that it used,to be the like you know the Big Arm chip,was really like the Huawei campaign 920,and that chip has been very hard to get,in the US I can now tell you that um Im,re-recording this video and yes we do,have some chips now so that is an a,system to actually run those so we will,have a review of that uh I think would,probably be the first review in the US,because uh its really it was took me,like two weeks to import the thing but,after a while wait the really big one is,lets face it is ampere there are two,big chips I mean really right now its,youre either ampere or your AWS,graviton 3. now we did see that the new,Alibaba Chip is going to be coming out,and that I think will be available in,the Alibaba cloud in I think September,and that is that definitely is an,interesting chip and well talk about it,soon but right now I mean as the market,stands youre basically either going to,be on like graviton 3 or youre going to,be on an ampere based product now all of,the Armed vendors try doing the exact,same thing and it is absolutely well,understandable but its also,mind-numbing to me and that is they try,comparing their CPUs directly to x86,CPUs it is not what theyre trying to,sell it as its not why Cloud providers,are trying to buy these things I mean,the the idea of just comparing straight,away from an arm processor today to an,x86 processor is just frankly silly but,I guess thats just all we have to do,and people have to talk about something,so thats what happens but realistically,the vast majority of arm processors,today are really focused on a single,metric and that is providing the best,and most consistent vcpu performance,that they can and for vcpu were just,going to say whatever kind of cloud,vendor you have theres some kind of,compute metric like you know CPU compute,metric and thats really the metric that,folks are trying to hit and what,basically happens is that the cloud,providers say hey we need you know on,this Benchmark Suite or like whatever,the set of tests we need this level of,performance per vcpu and we want it to,be as consistent as possible so that way,the more folks and the higher,utilization we get from our Hardware,because were Cloud providers those are,basically you know thats thats the,performance profile we want to maintain,so everybody has the same experience so,when it comes down to it if you know the,difference between the Intel P core and,e-core the performance versus the,efficient core a lot of the arm cores,today are actually probably closer to,the Intel e-cores right the efficient,course and that makes a lot of sense,right the efficient cores dont have,things like smt so they dont have hyper,threading and that means that they are,one core one thread and thats exactly,what a lot of the arm vendors are,focused on these days they also dont,necessarily have all the same,acceleration and you know Provisions for,a lot of the high-end acceleration that,some of the performance course has and,theyre not trying to hit like five,gigahertz or something and so with the,arm vendors try to do is they basically,say okay we are going to go use the best,tsmc process usually tsmc but the best,process we possibly can to go and make,our chips which means that they are,smaller theyre more power efficient and,we can make them denser we can put more,cores on a chip again more cores more,better and so the current generation,ampere Ultra Mac well that chip is,actually up to,128 cores remember this is at a time,when the top end ice Lake CPUs those,things are only at 40 cores the AMD epic,are at 64 cores the big difference of,course and if those numbers seem way off,is just because the those chips are,really looking at like you know using,smt and so they are really kind of,thinking like you should be comparing,threads to cores and then you get a lot,closer now the Ampere ultramax has now,been adopted by the Micros

More: the voyeurs movie review

AMD EPYC 7003 Milan Performance, Features, and Intel Ice Lake/ Cooper Lake Competition

hey guys this is pat from sdh and today,were finally,able to share the brand new amd server,processors codename milan,and those are the amd epic 7000 3 series,these processors are amds newest,generation of processors that are going,to carry,the epic line through 2021 and all the,way up until we get to genoa in 2022,but frankly this video comes at a very,weird time for the industry,because well lets just take milan for,example milan has been shipping for,revenue to customers i think since like,q3 of 2020 we got our first chips in,november,of 2020 although amd wasnt really too,happy about that,but that happened on the next generation,intel xeon iceland,intels already shipping those for,revenue and they havent actually,formally launched yet but,you know we know the performance of them,we just cant talk about it and in the,market thats actually kind of a big,deal because what that practically means,is that the milan chips we can only,really compare them performance wise to,what,well i guess cascade lake which was a,derivative of the 2017,sky lake or we can go to cooper lake,which is also a derivative basically of,the 2017 sky lake,but those are like our two options we,dont have the real competitor,which is the ice league xeon that we can,talk about right now but i think i have,a way that we can at least have that,discussion without,having the actual discussion so basic,game plan for this video is that first,were going to look at the,amd fx 7003 series mulan you know skew,stack pricing like what what are the,actual chips that are launching today,were then going to go and talk about,some of the improvements that amd made,from the,epic 7002 series which is rome to the,7003 mulan series,were going to go into the performance,and power consumption and of course the,new version,is well faster and it uses more power,and im just going to show off,just a couple little platforms that,were going to be doing actually some,actually really big platforms that were,going to be doing features on,uh that we already have that were,reviewing and so im going to just show,off a couple of those just to kind of,give you an idea,of how the market is receiving milan,and then creating new server platforms,in segments that milan really,and rome i guess really didnt access,prior to this,launch then what i want to do is inform,you,and kind of give you some idea in terms,of how to think about milan,in the context of what intel has,publicly stated and shown us,about ice lake already now we went,through all of the public bits on the,next generation ice lake zeons,and i think we actually have enough that,we can frame,a good discussion or at least a good,discussion guide,that will let you talk about milan and,ice lake intelligently and what each are,going to offer,and then what were going to do is were,going to synthesize that into,kind of giving you in the audience a,high-level guide that you can use,to think about how you would plan your,infrastructure purchases and,where these chips are really going to be,good and where i think intels going to,be good and if that seems like its an,absolute lot well,hey look it is i know the sth team has,been working for weeks getting ready for,this,and i know that the pieces that we have,on the sth main site,typically are read by just tons of,people i mean those things are just,huge pieces theyre often you know i,dont know what this ones going to end,up being but,7 10 maybe even more than 10 000 words,they become big reference pieces,and so this is the first time that were,really doing a video accompanying that,if you want to get more detail i would,definitely suggest going to opening that,we have a link in the description,check it out theres gonna be a little,bit more detail because were just gonna,kind of skim through some of the high,level details,that i speak and so you can go and do,that deep dive if you want to on some of,the parts,as part of the amd epic 7003 milan,series amd is launching a total of,19 new public skus and of those,there are four p skus which are only for,single socket as well as,four high frequency so frequency,optimized skus,since we get this question every once in,a while you can use a two socket sku in,a single socket,server but you tend not to do that,because the single socket skus are much,less expensive on the frequency,optimized part those are really skews,that are designed,because there are a lot of software,packages and licenses out there that are,licensed on a per core basis instead of,on a per socket or,per server basis and when you have that,per core licensing,having higher performance course rather,than more course is often,more cost effective now when we say that,there are 19 skus thats,really only kind of part of the story,there are six skus,from the amd epic rome series that are,also going to continue,into this next generation mulan so,youre still going to purchase at,some price points and some i guess,capability points the roam series cpus,something that i noticed when i looked,at the six skus is that the vast,majority of them,are the four channel memory optimized,skus,that amd had in the rome generation we,actually did a video,and an article on what that is but,effectively what amd does,is they have a set of skus that have,only a certain number of ccds active and,the placement of the ccds are optimized,for only four of the eight channels of,memory,being populated and that really is a,response,to the folks that are coming from the,intel xeon e5 generation where you had,four channel memory,and if you didnt necessarily want to,put a whole bunch of memory into a,system and you just wanted a little core,count and low memory,amount that is actually a good way and,thats a way that amd has to,optimize their platforms for you know,those kind of lower memory accounts,where youre only going to put four,dimms in,now were going to flash up the skus but,i want to give a couple high points the,first one youre going to notice is that,in terms of list price and price per,core youre going to see that there is a,migration,upward its not necessarily a linear,migration on all of the skus,but at some point youre going to see,that all the skus at the same level are,going up in price,i think that kind of makes sense milan,is faster on a,per core basis and so i think amd is,starting to capture some of that,value the other side to it is from a,competitive standpoint,intel tends to have higher list price,parts and then realistically whats,happening is that theres a pretty,fierce discounting war thats happening,on deals and so by having a higher list,price,on the amd cpus it means that vendors,can then discount,more and make it look like theyre,giving a bigger percentage off to hit,their street prices for their servers,and just to give you some sense of that,discounting so that way you can all you,know,be realistic here its just if youre,like a cloud provider or something like,that,the 64 core skus that are on this price,list,basically those cloud providers are,paying less than a third of what youre,seeing there,i think the other side to it is just,that amd has been a little bit supply,constrained in terms of server,processors and so i think this is really,just amd saying well,you know we cant make enough of these,things anyway so,why are we selling them for a bigger,discount than we need to so maybe we,should get a couple extra dollars out of,it i think that totally makes sense from,a business perspective and thats kind,of what youd expect,two other points that i think are very,related are the fact that tdps are,generally up in this generation so we,actually see a formal launch 280 watt,tdp part and that is something that we,did see in the previous generation we,did a review of the amd fx7 h12 we did a,video of that as well which we can link,and thats now a formal sku so with,those higher tdps amds able to go and,raise,their clock speeds up a little bit and,thats one of the big benefits of milan,is that you get two things you get the,ip

More: bose quietcomfort earbuds review

Intel Xeon Ds Go-FAST Feature

hey guys this is Patrick from sdh and,today were going to take a look at,Intels secret weapon in the data center,specifically we are going to look at,integrated quick Assist Technology that,is really for crypto and compression,offload and that saves an absolutely,huge amount of CPU resources if you,actually take advantage of them for,really common tasks crypto and,compression right now Intel has had,quick Assist Technology weve been,covering it I mean weve been covering,it since like 2013. we did some articles,on like how tos and stuff in like 2016,2017 but this year weve really shown,some of the performance benefits you get,from both using adding cards like this,but also just using the qat engine in,our last video we actually showed what,happens when you take a quick assist,accelerator and put it into a dual,socket Intel Xeon server now in a big,server thats maybe not a big deal right,you have a lot of pcie slots but when,youre in the embedded space where you,have limited space and also just fewer,pcie slots having to add an admin card,to get faster crypto and compression,thats something that a lot of people,just dont want to go do and that is why,were not just going to take a look at,this in the isolate D series because,well that is definitely a cool platform,what is I guess more impactful is just,the Next Generation Sapphire Rapids and,these are going to have you know theyre,going to sell a gajillion of those CPUs,and when they do that the vast majority,of them are going to have quick assist,built in so youre going to get quick,assist without having to add an add-in,card and then its just a matter of,actually using it and by the end of this,youre gonna see why not adding a card,and having it built into the CPU is,going to give you absolutely massive,gains and frankly that is Intels secret,weapon in the next generation of CPUs in,the data center so with that lets get,over to the new set and take a look at,some numbers okay so for our test setup,what were going to be using is an Intel,Xeon D,2776 NT now this is a 16 core 32 thread,part which is really a pretty high-end,part its also part of the Xeon D 2700,series thats important because its the,UND 2700 series and 1700 series actually,have different versions of qat,acceleration the way the the qa2 works,on them is is actually different this is,definitely the higher you know like kind,of higher performance version of it but,because we were looking at the qat,accelerator cards and also we knew that,were going to be doing Sapphire Rapids,um you know I kind of want to get like,the high-end Xeon d uh acceleration with,qat so thats why were using that part,on the AMD side it felt kind of weird,and Ill tell you exactly why we were,using the AMD epic 3451 which is a 16,core 32 thread processor also its also,in the exact same embedded space when,they compete directly with the Xeon D,that we were using but on the other hand,its uh its frankly an older part I,think the first time that we reviewed,that or a platform with that was in like,2020 and just to be sure that thats,what we should be using I actually sent,a note to uh you know like my AMD rapper,I said like hey is should we be using,this and hes like yeah thats thats,basically what you got theres nothing,that you know was going to be releasing,between when we did that testing back in,the end of June and uh and now so uh I,kind of knew I was like okay well thats,thats uh thats what were gonna use,and I even said like hey if youre gonna,release something let me know and Ill,hold this off and and uh and they didnt,so uh you know this is kind of what AMD,has and its kind of a bummer theres,another difference with the AMD platform,on the znd side we could actually to use,the onboard NYX because those are you,know kind of like a little bit higher,end mix the version that the AMD folks,have um theyre like kind of good if you,just kind of need absolutely basic like,Bare Bones 10 gig networking but for the,kind of stuff that we were doing uh you,know especially at higher speeds they,just kind of theyre not fast enough so,we actually had an ad in a uh Intel 800,series Nick to be able to actually you,know handle the level of performance,that were hoping to get so that is a,difference in the test setup because we,are using an added Nick for the AMD,platform but we just kind of had to do,it to get better performance thats the,reason and hey before we get too far on,this I do want to point out that were,going to say this is a sponsored because,we did do that flight up to uh you know,to Intel and the other thing that youre,going to see later on this video is that,we also have Sapphire Rapids and so you,know Intel had to give us permission to,actually be able to publish those,Sapphire Rapids numbers because the,launches not for a while and so uh we,actually got permission to be able to,actually show that so we can kind of,talk about the story that really goes,from like you know embedded in the qat,accelerator in the current generation,isolate D part but then all the way,through what will happen in the next,Generation where we have Sapphire Rapids,like build,with the qat accelerator so I kind of,want to show that entire story but of,course like everything that we do on sth,Intel didnt get to like see this video,they get to see any of the website stuff,they can see the numbers any of that,kind of stuff before we actually,published and so you know this is being,done completely editorally independently,just want to make sure thats clear,because I know there are a lot of places,on the internet that are not like that,with that lets get to the numbers okay,the first workload that were gonna look,at is compression and by the way these,are going to be the same workloads were,going to test like the same things that,we tested in the previous generation we,actually had like an add-in card,accelerator for qat so were gonna be,doing the exact same thing but were,just going to be using the built-in,acceleration on the Xeon D 2776 NT so,here were going to be using the Calgary,Corpus as kind of our you know dated to,go and actually compress were going to,be using qatzip which is something that,is on GitHub Im not going to like read,to you guys the GitHub links and stuff,like that so again if you want to go see,like where all the stuff is on GitHub,you can actually just go look at the,main side article in in the section for,these well have all the links there so,you can just go find that but the big,things that were going to test here are,were going to test like a base case,were going to look at the isil,performance and and I said well well,talk about a little about what we saw,the differences not only on the the,Intel side but also on the AMD epic 7003,series versus the 3000 Series that were,testing here and then uh were gonna,look at the qat hardware acceleration as,well so base case is our you know brute,force that were doing both on Intel and,AMD we then have ice L which actually,using some extensions that we have in,the CPUs to actually accelerate things,and we actually get a lot of performance,from that and then finally were gonna,be looking at the hardware acceleration,and thats basically the setup okay so,lets look at the base case and,something that should be immediately,noticeable to anybody that saw our,previous piece versus this one is that,when we have all 32 threads working the,Intel and AMD like in the previous,generation the AMD ones are actually a,little bit faster but because these are,the older like kind of Zen generation,cores uh in this chip and you also have,like the two ccds and stuff like that so,its a kind of a different setup the ice,like processor actually a lot faster in,this one when we kind of just open,things up even in the base case now the,same thing happened when we went to ISA,L and in isil again and when we looked,at the you know version where we had the,qat accelerator on the big big ice Lake,and Milan chips uh you know we actually,saw the Milan was

AMD EPYC Milan Is Better Than We Thought

modern processors are dedicating more,power to i o,and memory and modern amd processors are,getting in a little bit of a tears with,their performance with the latest,milan epics,whats your minimum specification,lenore makes managing cloud,infrastructure easy with same pricing a,full featured api and 100,human support whether its a personal,vpn game server,website or a big-ass gpu solution new,users controller node today with a free,100.60 day credit at leno.com,tech tech potato so lets go on a,journey,with a standard modern desktop cpu,we may have 105 watts 125 watts,and most of that power is dedicated for,the cores,the i o dual channel memory and some,pcie lanes doesnt take up that much,so ultimately most of the power that you,get is going to the cause,giving you higher frequency and higher,performance,now this becomes a problem when you move,into the sort of high-end desktop,and server space where youre not only,dealing with more cores,but more io in this world were dealing,with six channel memory,or eight channel memory and pcie lanes,are going up to 40,48 64 even 128 pcie lanes,so not all of the power is going to the,cause so youre not getting as much,performance as you might expect,this actually becomes doubly worse for,chiplet-based designs,because now you have to transfer your,data across,your bus across for example on amd,the infinity fabric offer intel the mesh,the downside of having these chiplet,architectures and when intel moves to,chiplets theyll come across the same,thing,is that you now have to move the data,from silicon through the package or,through an interposer,and then back into other silicon and,that has an inherent power cost,as your processor has a full power,budget,that power for to transfer the data has,to come out the overall power budget,reducing power for the cause,therefore reducing peak frequency,therefore reducing,peak performance this is usually a,trade-off however because there are lots,of workloads on these systems which,require,more memory bandwidth more pcie,bandwidth more o bandwidth,maybe youve got a workload which,involves a lot of crosstalk between,cores crosstalk between threads,youll need to go across that fabric and,talk to those cores,if the power wasnt there for those for,the fabric,then you would have a degradation,performance so it is very much a,balancing act,based on your workload for those of you,that follow consumer processors,such as the ryzen 9 3950x and the ryzen,9 5950x,youll know that we have two core,complex chiplets,with an i o die attached and on review,sites where theyve managed to,measure package power and core power,such as anantek,youll notice that these processors with,their extra i o die,consume more power at idle and as a,result,theres less power available for the,cause this becomes more of a problem,with amds rom and milan processors,where youve got,eight core chiplets and one big io die,with eight ddr4 channels and 128 pcie,lanes,and in initial testing of that rome i o,die we found that it was consuming,around the 50 watt mark,that means that out of the 225 watt 240,watt,for the whole processor you now have 191,to 175 watts,for the cores reducing what would,potentially be the peak performance if,you just looked at that tdp,power value when it comes to the latest,generation of milan,processors from amd there was a little,bit of a roadblock,so most reviews for milan day one,used amds daytona reference platform,this is amds reference platform from,the first generation naples epic that,theyve repurposed for rome second,generation rome,and now third generation milan most,reviewers if not all reviewers who did,launch day reviews,got access to a daytona server with the,newest cpus,what was interesting to find at the time,with andres numbers is that the i o die,instead of using 50 watts,roughly per cpu in rome was now using up,to 100 watts,in milan this is a base idle,power which is quite a lot so now youve,got a 280 watt cpu,only having 180 watts available for the,cause,for frequency for performance so what,exactly was going,wrong here at the time of that initial,milan review,pretty much everybody who reviewed the,processor lineup said great lots of,performance,its a winner and regardless of this you,know 100 watts per i o die,bass power consumption performance was,still,great especially with the lower core,count models,however when anantec got hold of a,third-party motherboard from gigabyte,things changed slightly,as andre points out in his recent,updated review,so this is andres uh updated review,milan review part two uh amd platform,versus gigabyte platform io,our own head per core uh basically it,says,in the original review we saw a high,idle package power exceeding 100 watts,as i mentioned um we are communicating,our findings and worries to amd prior to,the review publishing,but never root caused the issue and,were never able to confirm whether this,was intended behavior of the new third,gen milan chips,or not we theorized it was a side effect,of the new i o die,which has the infinity fabric running,slightly higher frequency,and then generation also runs one to one,with memory control clocks,so package idle power this graph here we,look at intels cascade lake and ice,lake,26 27 watts idle and then the ultra,uh 80 core arm processor 32 watts,then if we have a look at the whole,graph uh,focus on the bottom bit for a second,daytona so this was,the third generation you can tell,because it has a three at the end of the,epic name,running 100 watts package idle power,however with the new gigabyte,motherboard,these processors were reduced so lets,take the 7763 thats 64 core part,99 watts idle package power reduced down,to 65,with the new gigabyte motherboard and uh,yeah as andre says to our surprise when,setting up the new gigabyte system,we found out this behavior extremely,high or idle power was not being,exhibited on the new test platform,instead of 100 watt idle figures,it tests with the daytona system were,now seeing figures that are pretty much,in line with amds roam system at around,465.72,and then he goes into some performance,numbers showcasing you know,what the i o die you know difference is,and what performance uplift this means,but essentially across,um spec 2017 were seeing anywhere from,a one percent to a,13 uplift just by virtue of having,more power available to the cause,so what exactly is going on here now,as i said amds daytona reference,platform,that was used in those first reviews,dates back,to first generation naples that,that chassis that motherboard that,design was built,for naples and with knowledge of rome,with knowledge of 2nd gen,it just so happens that amd decided to,update it to support,third generation milan as well however,if you go look at motherboards,in the market that support uh naples,rome and milan youll find that they are,very distinctly into two different,categories,you have motherboards that support 1st,gen and 2nd gen,or motherboards that support 2nd gen and,3rd gen there really isnt any,commercial,solutions that support both first 2nd,and also 3rd gen,at the same time if anything those that,do,are built targeting third gen with,support for second and first gen,daytona was built for first gen with,support for second and third,now if you can follow where im going,here the fact that,amd you built daytona for those first,generation naples,it was built as a 240 watts per socket,platform,were now testing 280 watt um,third gen epic milan processors in that,platform,and somewhere along the way this is,probably a bios,update a microcode issue just monitoring,of the system,the i o power the i o die power on third,gen epic,in a daytona system was,practically double you know and thats,taking power away from cause taking,uh performance away from end users so,when andre tested with this new gigabyte,motherboard,hey easy 13 up to 13 uplifting,performance,and actually thats pretty amazing,because that means that amd is further,ahead,than anyone thought about in in t

Crazy! AMDs Milan-X Delivers 1.5GB of L3 Cache to EPYC Servers

hey guys this is patrick from sth and,today we are going to talk about these,things now these are the amd epic,7773 x chips also known as milan x and,these are not just any milan x chips,these are actually the 64 core versions,now you might say patrick i have heard,about 64 core cpus for years amd epic,milan was launched in 2021 and rome was,launched in 2019 so 64 core cpus they,have been out for a long time what makes,these special what makes these special,is that these things each,have three quarters of a gigabyte of,cash that means that in these two chips,that im holding in my hand right now i,have a total of 1.5,gigabytes of level 3 cash which is,absolutely insane just for some you know,frame of reference here if you had the,intel,xeon platinum,8380 which is their you know 40 core cpu,you would only have 60 megabytes of,level 3 cache on one of those or 120,megabytes of level 3 cache on two of,them which basically means that these,have well over 10 times the cash of,their intel counterparts so today were,going to talk about how these things are,made were going to talk about the,performance and for that we have a,special guest hey guys this is wendell,not from sth and today were going to,take a look at the milan launch and if,you dont know who that is that is,wendell from level one techs we were,both out in austin why live here and he,flew out but we were at amd headquarters,just a few days ago and you know we,finished this entire thing we finished,our dinner and said hey we should go,film something so we went back to his,hotel room and we did a quick little,segment and so were going to have a,portion of that later in this video and,were going to talk about that when we,get to it there he is by the end of this,video what i really want to do is talk,about not just our testing and some of,the other testing thats been in the,market but i also want to talk about,well should you go buy this and like how,do you even think about going and buying,these things and like you know what,whats amd doing with this line anyway,because this is not a full replacement,for the amd epic mulan series this isnt,like oh we had milan and now we have,mulan x mullen x is like a,segment or a family of processors within,the broader milan series so if youre,still going to go buy milan processors,and probably the vast majority like i,dont know the exact number who knows,the exact numbers on these things but my,guess is that something probably like 90,percent of the processors that are going,to be sold over the next couple months,are still going to be the standard milan,not milan x,but at the same time i think that these,things are really important and there,are some people that are just going to,and some organizations that are going to,see these things and be like this is,exactly what i want so with that lets,start getting to how the heck these,things are made and then lets start,getting into some details okay so the,quick background on milan x and im sure,that there are people that are going to,go into like the nitty gritty like all,the bump pitches and like all that kind,of stuff but lets just kind of talk,about what the heck amd is even doing,here and to understand whats going on,here lets just talk about why cache is,important well cache is important,because when you have data if you can go,store that data on the package rather,than having to go off package to main,memory to retrieve that data that,basically means that you have an order,of magnitude or more of better,performance because you have lower,latency you have more bandwidth and also,the amount of power it takes to move a,you know bit of data from you know in,the chip to the other side of the chip,is relatively small compared to having,to go follow a trace all the way out to,main memory and all the way back and so,its more power efficient and its also,you know just faster so having more,capacity means that you dont have to,necessarily make trips to maine memories,often and that means you have more data,thats close to the cpu and because you,have data thats closer to the cpu the,cpu cores are like hey im constantly,fed with data rather than saying like,hey when are you going to get back from,main memory and like you know when am i,going to get this data that i need to go,do compute so thats basically the,entire reason that a lot of people,really love big caches and we actually,did a piece about a year ago talking,about how we were going to start,entering the gigabyte era in terms of,cpus these chips are probably the first,ones that we would really say that like,you know we have something like three,quarters of a gigabyte so were starting,to express the cache in these in terms,of gigabytes rather than just megabytes,and we cannot get into this without,talking about manufacturing at least a,little bit and if you dont care about,manufacturing i would just say skip,ahead but heres basically the challenge,that amd faced so if you look at like,logic scaling and the density and stuff,when you start talking about analog,circuits those things really are not,pushing at the same rate that youre,seeing like the compute circuits that,you would have like for cores and things,like that thats called like logic and,somewhere in the middle of those you,actually have sram which would be like,your cache and so if you create a dye,that basically has your logic and your,sram on the same die well that basically,means that you have part of what youre,manufacturing scales well and the other,part of what youre manufacturing is not,scaling as well and so that kind of,creates challenges and thats exactly,why we see things like amd has a giant,io die in the middle of their package,because well you can go do your stuff,there for io and you dont necessarily,have to go do that signaling on the,expensive die that youd use for your,logic so thats why theyre actually on,two different process notes and amd is,thinking about this and saying hey maybe,theres another opportunity and one of,the big challenges is that if they,wanted to expand the cache size and they,were still trying to use that seven,nanometer or something like that what,you would basically see is that the,overall die size of each of those,compute dives would start to get just,huge because well you have a lot of sram,you have a lot of cash plus you have,your course but its not necessarily an,efficient use because sram doesnt scale,as well as logic so basically what amd,did is this thing called vcash and what,they did with 3dv cache is they,basically created sram so they basically,created just little cache dies and they,said okay well we cant we dont want to,create one die with all of our cash and,our logic so our course what were going,to do is were just going to go take our,cash die and were going to go add,another cash die on top of that its,called 3d stacking and it literally,looks like you know like legos or,something like that almost where you,just go and you take you know you have,your base die and then you go put a a,cache die on top and then you have to,have other little bits that are really,just there to make sure that,everythings level so you can actually,put like heat spreader and stuff like,that on it and if youve ever played,with legos youre probably like yeah,thats super easy you just go and you,just take it and you just plop the,things on i know a three-year-old that,can do that no problem why is this such,a big deal on chip design and the reason,is a big deal is because well theres a,whole lot that has to go on because you,all of a sudden need to have wires and,things going up out through the top of,one die and then you have to have wires,going to the other die and then you have,to go figure out how to bond those two,dies together so they dont move you can,actually have them be reliable and by,the way youre doing this at super super,small scale and you have to also scale,out your manufacturing to be able to,create these things cheaply and reliably,so theres a whole bunch o

AMD EPYC 9004 Genoa Gaps Intel Xeon with 96 Cores WOW

hey guys this is Patrick from sth and,this is the new AMD epic 9004 codename,Genoa and this processor is absolutely,walloping Intel right now let me give,you a couple key specs of the AMD epic,9654 that we have right here this,processor has 96 cores,192 threads it also has 12 channels of,memory thats ddr5 in this generation,and it also has pcie Gen 5 that can do,cxl and all kinds of other stuff now,youre gonna see in front of me that I,also have a wide selection of other,processors that are on the market today,we have things like the AMD epic ROM,system 2002 series 7003 this is actually,the Milan X right here over here we have,the Ampere Ultra m128 30 which is 128,arm cores but this is definitely a,different class of processor over at the,very end over here we actually have the,Intel Core just desktop thing just you,can kind of see the send to scale here,and then on this side youre going to,notice that we have two processors from,Intel the processor on the far end is,the ice Lake generation processor and,you probably have seen on the sdh main,site and I think we actually did a video,where we violated a little bit of this,but we are starting to test Intel,Sapphire Rapids and so youll see that,here as well we have a couple other,General skus that were going to talk,about in a little bit but just overall,heres the deal this thing is number one,absolutely Giant and number two it is by,far the fastest processor out here and,not only is it the fastest processor out,here but compared to a lot of these its,actually more like maybe a 2X processor,versus what youre going to see on the,table so in this video were going to,talk about just about everything so,definitely use the chapter markers if,you want to skip ahead or move back,were going to do a ton here but this is,only going to be a subset of what we go,into on the sth main site those articles,are absolutely Giant and they become,industry references so we put a lot of,effort on the main site article and,were going to try to compress things,for YouTube but that also means that,were gonna have a lot here so feel free,to skip around okay so lets just talk,about the context of what were seeing,in the market right now the previous,generation AMD epic Milan and Milan X,went up to a total of 64 cores and 128,threats the molonics also had a lot of,extra cash but were going to talk about,that in a bit but of course Intel is,still the largest player in the market,and right now their chip in the market,they actually have two by the way the,first one is this one over here which is,the isolate processor we also did the,Cooper Lake which is the other third,generation Intel Xeon scalable that a,lot of folks forget about but we,actually did that in four socket,configuration a couple years ago the,next generation will be launching on,January 10th if all goes according to,plan about two months after youre gonna,see this video Intel will have up to 60,cores a number of accelerators that,weve already started to show you the,performance on on the sth Mainstay a,little bit on YouTube as well and Intel,will be joining the PCA Gen 5 and also,the ddr5 and cxl generation as well,however I cant talk about exact,Sapphire Rapids numbers because Im just,not allowed to yet because uh we have to,wait for two more months now when you,look at these processors I mean,absolutely they are awesome right you,see the lsdp youre going to see tons of,cores we do each top youre gonna see,lots of cores and we do the things like,LS toppo youre gonna see the topology,and youre gonna see all these little,clusters and then there are little notes,that like oh theres 12 of these now,having more performance of course is a,big deal but lets just kind of be clear,when we get to a platform perspective,the fact that we get more cores means,that we can do at least a three to two,consolidation just over the previous,generation Milan chips if you have Intel,isolate processors youre probably,talking about a two to one if not a,little bit more like two and a half to,three to one server consolidation ratio,from AMD epic today to what Intel has in,the market right now but core count is,clearly not everything and I want to be,very clear in terms of the overall,context because we talked about Sapphire,Rapids but this is an amds only trick,either were going to see a Genoa X,version of this chip which will have,more V cash or 3D V cache and the 3D V,cache will mean that we have well over,one gigabyte of cash per socket so to be,clear this is kind of more of a,mainstream processor this has,384mb bytes of level 3 cash per socket,goes down of course with other other,ones well talk a little bit about how,that works in a bit in our architecture,section but the key thing is that there,is going to be a general X it is going,to have a ton of cash and that is coming,so this isnt really the CPU thats,designed really for that high,performance Computing mode because that,will be General X the other big one,though is the scale out one which is,going to be Bergamo which will come out,in the first half so guessing kind of,like Q2 of next year that will use the,same socket as Genoa but programmable,will have up to 128 cores in a single,socket so thatll be more of like the,cloud provider scale out core solution,not really for those high performance,Computing workloads more for like Cloud,scale out Sienna will be the smaller CPU,because these things are clearly,becoming too big to just service the,entire market so AMD needs a smaller CPU,that will be Sienna the other one that,folks forget about is the Mi 300 which,is the Apu the giant Apu that AMD is,targeting for the high performance,Computing really super Computing Market,that will combine both the x86 as well,as gpuip onto a single Apu and really,creates something thats new and very,different than what we have today and,the reason Im going into all that is,because the market for this is this is,the mainstream just kind of like old,school Xeon replacement part there are,other parts that were gonna see from,AMD and they occupy other market,segments and so there are folks that are,doing different things in the market and,AMD clearly sees that and they are going,after specific applications with,different silicon in this generation so,this is the first one really of this,generation but its also probably the,most important because its the,mainstream one so what are the skus,anyway well first off Im going to give,you the decoder ring and the AMD folks,say that this is such a big generational,leap that theyre going to call this the,9000 Series they have a new set and if,you want to spend a couple seconds just,pausing this you can totally go look up,the SKU Im not going to explain it but,you can go look up the SKU thing and you,figure out what the decoder ring is for,these different parts next lets take a,look at the skew list just kind of,seeing what AMD is really putting out,there are a total of 14 dual socket skus,and then there are a total of four,single socket only peace use core counts,range from anywhere from 6 teen up to 96,cores the top end one which is the 96,core version that were looking over,here that can go up to 400 watt TDP,youre definitely using more power but,youre getting a lot more performance,for that were going to see that in a,bit four of these skus are F skus which,means that they are frequency optimized,and the reason those are important is,because if you have something like a,Windows server or VMware server or,something that is licensed on a per core,basis the F-series skus are designed to,give you a maximum amount of performance,for a given core that means even on the,16 core part we still get a 400 watt TDP,and we get 256 megabytes of level 3 cash,but frankly this SKU stack is really,simple compared to what weve seen from,the Intel Xeon side this is darn easy I,mean theres basically 14 skus plus,another four of the single socket ones,now one thing that you will notice is,that the prices have gone up in this,generat

Categorized in:

Tagged in:

, ,