I think Navi is wrong because of GCN and it's got some GCN bits but our DNA really is fundamentally different and that was something AMD was keen to like underscore and point out like we've got new I say documentation coming out and we've got you know it's totally totally new Hardware implementation so yeah it's like a it's a weird sort of concept of the compute unit is different as well to be honest

I've given everything that they said I reckon that our DNA is sort of like the trial run oh sorry Navi is the trial run for our DNA yeah there's our DNA - you know it's like the same D with the CPU stuff Naples was the trial run Rome is where the cash money is it's where all the cash is anyway yeah well I think that they're gonna have a license to print money with Rome especially in the second half of this year because there are a lot of data centers that are hanging on to those like v2 v3 v4 Zeon's that are overdue for an upgrade and Rome in terms of both legacy consolidation and new projects is going to make a lot of sense for the enterprise

It's I guess we should mention that Xen also comes with enhanced QoS yeah all machines yeah so one of the things that Intel introduced back in Broadwell is that you can get your VMs or your containers and you can allocate specific amounts of l3 or memory bandwidth so the concept of noisy neighbor in a cloud environment is is real you don't want somebody who's in the VM next to you hogging up all your resources so Intel implemented that with Broadwell II I think this idea of being at the hypervisor level being to control how much cash and how much memory bandwidth and now AMD has that was then - and that's gonna be a big kicker

And that's gonna be a big kicker and we're in I'm looking forward to actually testing that they also mentioned the changes to like the TLB and that will have implications for some of like the game machine virtualization stuff that I'm doing yeah I really want to test that it it given that these features exist it makes me want to test different amounts of l3 yeah for so I mean I am denoting that double the l3 has good knock-on performance for gaming and the consumer stuff you know going from two mega pack or two for Megacorp but I kind of wanted to do a scathing article where it's just comparing how much that actually affects it so it'll be an interesting thing to test hmm but it's glad that these actually exist we've been told that you know they gave us three cash families of instructions that they've included in his m2 but we were told by Mike Clark the father of Zen father the Godfather I beat him at bowling last night but he said you know we actually have a lot more you know instructions included that are related to cash which would be given a hot chips the hot chips conference which I'll be at in August will you be a hot chips probably no that's a shame maybe I mean that August is one way way I wasn't planning on it but the hot chips is gonna be one of the one of the most exciting conferences this year if you're into the technical side yeah all these chips they're going to go through all the changes it's it's not only Zen but we're gonna see cascade Lake we're gonna see Intel's joint venture in China that side that's called gin ties there's way for scale machine learning so yeah so you have your machine learning chips that are you know tiny imagine doing it on the size of a wafer I mean it depends what size wafer you want but yeah that that's new technology that's coming through at the same conference last year we saw nanotube based DRAM and that's being deployed in Fujitsu systems now

The MIT people that are not heating just release public details of their carbon nanotube based SRAM circuit 16 kilobytes and it's SRAM and it seems to be oh yeah it's a wafer wafer scale I thought I saw David Shore from wiki chip he was posting photos of the chip EC said he must get hold of it and then he did a line us and dropped it they did anyway no so it's it's exciting there's a lot there's a lot to unpack here there's a lot that we can probably discuss on future like a future thing so if you like this you should probably update it or share it or something I don't know so it's good well thanks again for for joining me or wait well you're gonna post it and I'm gonna post it to you yeah it's fine chance Wendell it's ears yes I like to it's much less complicated and much less awkward so we'll see you guys in the forum and or on tech tech potato photonic

"WEBVTTKind: captionsLanguage: enhey I'm here with dr. Ian Tetris we get that ride hey with him I got it right yeah I I struggle because I don't it's bad it's all right you know the anyway it's uh I see things and pictures in my head and then I have to translate it into words and so I get it wrong all the time like nano seconds and milliseconds and I see it but I don't say it wind and wind ya know tomato tomahto it's not even it's more complicated than that it's actually kind of scary so we're at the end of the I don't know AMD e3 press event which is technically before e3 yeah but after Computex yeah they decided to fly a bunch of a bunch of us out here for a tech day and AMD tech days are pretty good they like to go and deep dive into their products and um these products this year there's a lot there's a it's in it's a lot with very little competition that's last year last year I remember having to review five different platforms in the first half of the year and then like you know another two or three by the end of the year this year there's only one or two but they're all MD now I remember clearly from the Intel press event that something something new high-end desktop processors are coming at the end of 2019 but it was like one line in the entire presentation oh yeah new new new core X yeah which is we assume cascade lake eggs maybe was that the xenon W that already came out with the Apple was that what they were talking about no idea Magic eight Ball says arch again tomorrow it was the first time that we saw so their Intel is gonna do the naw mean I've heard KS it's five gigahertz all core out of the box how much your life to you not one oh five gigahertz we become those people yeah yeah I don't know that's if that's the best they got that's I got some bad news for them they must not have heard about multi-core enhancement that all literally every board partner offers some variation of also so the difference here is that is in back-right intel intel has this inspect definition if you if you as long as you don't change you overclock a bit you'll still in spec and you can get warranty what's in spec depends on who you're talking to from Intel Oh what so I actually had an interview with the Intel guy who has helped redefine TDP and turbo for the new era and he clarified what in spec and out of spec is I need to transcribe that interview and I'll post it on an antic but he was very clear in defining what inspecting a spec was but yes know if you talk to different people Intel that would give you difference but I finally have a definitive answer from the guy who matters well from the peanut stand or the peanut gallery for I am it just looks like there's a there's a level of panic in in Intel's outward messaging that I have not seen since the f1 x2 days it's it's weird how you you can position yourself in different ways to see it from different angles from some angles it seems like panic from other angles it's well what else can they do right if you are in that position would you be doing the same thing and probably yes but you know they are extracting more out of you know 14 nanometer plus plus plus plus plus if I if I run entails position I would probably jettison X 299 move everything to 36 47 and then keep the 11 5 X socket variant for like 8 to 10 cores and basically call it a day yeah I mean that would help simplify their products line I mean also having the same socket the different chipsets support yeah yeah and and you know changing their product categories based on that that's that yeah it makes it more difficult for people to understand but well I'm just looking at like on the server side you know when AMD did their their press event with like the protein folding thing they were like no no no you need to compare to the Platinum 80 to 80 which is like 4 to 500 watts from what I understand on the OEM so this is a funny discussion because I acted as a saw like go-between between AMD and Intel on this no I'm the prettiest no I'm the prettiest so so so to clarify AMD showed an AMD protein folding with the new 64 Rome socket server versus Intel 82 80 26 core dual socket 28 core sorry and it showed what a 20 nanoseconds versus 10 now Intel said favoring AMD yeah favoring AMD Intel said oh you're not using the right optimizations twitch AMD said we're using the latest public optimizations this is it's in avail saying well if you are something better than tell it and then they said well you're not using the Platinum 9282 which is their sort of dual die 56 core BGA only 400 watt monster you're not using two of those so you have more core parity so dual 56 core versus dual 64 core and twitch AMD replied saying well we would have loved to have used them but we can't buy them and Intel said well they'll be ramping through the second half of the year - which aim DS responses I still can't buy them and then you know that that we didn't even get into the argument of well you know price difference what's the price of Intel's 400 watt CPU versus you know 64 core Rome and well it's it's it was it was an interesting dynamic in how each of these companies want to represent their product in the best light and again you have it's do you come from the point of well should they be fair or should they be doing the product in the best light or are they just doing what they have with what yeah I mean it's marketing trying to do what the best with what manufacturing is giving them right so the the 56 core benchmark even with all of that the striking thing about that for me was that the margins by which Intel took the performance crown there were thinner than the not seer is what 23 to 20 yeah I mean it was just a shockingly small number that probably you could eke out with more optimizations especially if they had non-public optimizations on the one side yeah also did that had the meltdown inspector mitigations intel specifically states that any benchmark number that they showcase may not have all security updates enabled meet so that that's not to say they might not have but they might not have it's what I mean it's AMD does the same thing right I mean we saw this week at this event at the AMD tech day that they were showcasing in tell not with the latest security optimizations against AMD with the latest OS updates which is the best decibel connect case for both people except for Andy did not have the scheduler update right because they needed you it's in it's in the fast ring well they told me that it's in the fax ring and you don't need the chipset driver because that's why I got an anomalous result on one of my tests but for 19:03 you do need the chipset driver to enable the thing to make the scheduler aware of the compute complexes so that would put AMD at a slight disadvantage there it's it's yeah I'm just waiting to get the Hadron hand yeah I mean first first party benchmarks there what wait until the rest of us test it I think I think that this also like Andy has not done a lot of press events over the last 10 years I know a lot of benchmarking and things like that but it at least I have word appearances are that they are trying to be as forthright as possible while also of course putting the best foot forward there's no substitute for you know proper benchmarking I've noticed with both AMD and Intel and credit to both of them to be honest sir AMD hired Scott Wesson from the tech report and a lot of people know Robert how like he used to be in the press and Intel have hides you know Ryan shroud and you baby the no you okay yeah but it what we're seeing in these events is they're they're being more thorough with the data presentation because they're giving journalists who know how to present this data to present it to journalists right which to me is a great thing yeah I mean I I remember once razer launched the first razor smartphone and in their presentation they gave all the really intricate benchmarks that an antic does well that's because a former an antic employee helped them with their benchmark and yeah it's we're now seeing that in this sort of x86 base and as a result we're getting such a wider dynamic range of it is really interesting it's an it 20:19 is gonna be an interesting 2019 into 2020 it's gonna be an interesting time for processors not only consumers also Enterprise yeah well and you know at the Computex side of things now just the recent AMD event we're at the word to Beverly Hilton right now and on the eighth floor that the Hollywood sign is just over there yeah yeah we tried it did there's the exposure we don't have any lights it's fine but at Computex it was kind of electric for AMD it's on it was honestly kind of surprising because all of the board partners were super ramped what a suitors got 20 skews fancy 30 the gigabyte has like six or eight main skews yeah and as rocks got 10 and I think the official number was 56 altogether right yeah 56 X 570 so that's not even including the be 550 boards which we assume are coming in a couple of months Yeah right everything on the X 570 platform anywhere from say 120 bucks all the way up to $1,000 I mean as rock x5 70 aqua yeah with that full waterblock cover a lot of nice board nice-looking board a lot of effort has gone into these motherboards and you can tell because that's where the revenue for these companies is coming from for the rest of the year all right I hate to say it but it's true we're in it we're in a situation where AMD is now becoming a lot more competitive if you know equal with better power in some circumstances if you believe what AMD is telling us in combination with Intel having you know issues meeting demand on the server side and that's having issues that's having a knock-on issues with supply on the consumer side they keep saying that they've got the supply issues under control and it'll be ramping in next quarter but I've been hearing there for at least three quarters so I believe they said in the last earnings call that they will be ready to meet demand and supply by q4 is that because no one is going to be demanding them anymore if they keep so if they keep having security issues that knock-on performance people are gonna buy more processors I mean so one of the reasons why Intel is having such high demand or the service stuff is or overnight 30% of the performance decreased in your database workloads and stuff so how do you get three percent performance back you buy 30% more service kind of related to that I mean I hate I hate giving Intel a hard time because you know they are white like a forty billion dollar-plus organization on like the Rd and the fabrication yeah and all of that and we heard from AMD engineers at this event that you know they really didn't expect any kind of they actually expected a clock frequency regression with seven nanometers because the wires are so small the resistance increases it's a generally accepted engineering thing that your performance is going to regress and we see that with cascade we see that with the laptop parts I mean it's a 1 gigahertz regression in laptop parts I remember when the first 40 nanometer laptop parts came out the integrated graphics clocked to about 2 300 megahertz slower than what people were expecting Intel eventually got that back up over the next 2 3 generations but still yeah it's great you get a new new process node gives you more power budget to do more things to increase IPC but then you lose potential frequency at the high end yeah it's too much leakage or whatever whatever that whatever the electrical effect is that you get in heating when you should not because the resistance is higher and I think because it sounded to me like because that was an accepted engineering thing it's like well you know we're Intel we put so much into it this is gonna be a thing that everybody is going to suffer from and the champs at TSM see if you believe you know AMD like gosh darn it those plucky little guys figured it out and everything is amazing and and so they were able to claw back a lot of the a lot of the clock frequencies it's I mean this is a dynamic in the industry right when Intel has a new performance node a new process nodes they have they're the ones that are first to try it out because it's their note when TSMC has a new node like seven nanometer then all the smart phone guys jump on it first they're all doing the small source like 70 80 square millimeter dies and they get to iron out the process and an AMD comes along with an 80 square meter die saying hey we want to use high performance libraries at four point whatever gigahertz and tier same is here since he goes well okay great let's do that yeah it is a really interesting dynamic in a really interesting situation because even like if you look at like the liquid nitrogen overclocking like the super overclocking you know back on 28 nanometers like six and a half seven gigahertz okay and now we saw we saw a first hand at the same the event five point three seven five on the sixteen core all core and it's like five point four seven five that's not six or seven gigahertz yeah but it's also seven nanometers versus twenty eight yeah I mean well what's the world record now a eight point eight three eight gigahertz by Andrey yang and he did that on in fx-8150 on though so yeah this is yeah generations ago I'd rather have the the sixteen core 3950 X at four gigahertz then a completely stable FX you know 8150 or whatever at any frequency yeah yeah well yeah that's AMD's past AMD's future looks bright by comparison it really does what other interesting stuff we had let's see I'm trying to remember we did this once before and we lot there was a memory card in this happy enough time it because we had some other interesting things that we chatted about so one of the things we learned this week which we weren't aware of is the relationship between the i/o dies inside the CPU and the chipset oh yeah right so we've known about the server side of Zen too and the massive IO die that's built on globalfoundries 14 nanometer you know helping go through the wafers as part of AMD supply agreement in 512 yeah that's a whole other story but that chipset is 14 now the chipset in the Rison 3000 series is a cut-down version but that's actually built on 12 nanometer right so that's again part of the agreement and twelve nanometer means you know lower voltage the saving power and whatnot but the x5 70 chipset is actually the 14 nanometer version of that small io die so it's the same floor plan just built with slightly different design rules so you know it's voltage goes up slightly but on the chipset because it's exactly the same floor plan you don't need memory controllers so you can turn those off and there's all lots of other little chicken bits that you can turn off but so so what I tweeted this morning was that that chipset we've been told is you know 15 watts 11 watts depending on what configuration it's in that means that inside the matisse die is a 10 15 what io die inside the Matty's chip so that means that the cause have got 10 15 what's less than the TDP of the chip yeah to work with which means that 7 nanometre is more efficient than we thought yeah yeah yeah I mean that's that's that's a it's a very very astute observation and the you know the the PCI Express interface the other thing that I think that is an interesting takeaway is that the PCI Express 4 interface truly is just packet data switching no matter what sort of peripheral you've got plugged into it because SATA and USB and PCI Express now it kind of makes sense because we saw the diagrams where it's like oh if you want 8 PCI Express 4.0 lanes from the chipset fine or 8 and SATA interfaces fine or more USB fine and it's basically a one-to-one relationship the whole point of the chipset is really to just guess which ya put data on the fabric and oh yeah but that also means that you could theoretically run a PCI Express 3.0 by 8 device on a PCI Express 4.0 by 4 link with no loss of performance yeah assuming it does the conversion and I acts like a normal PC I switch yeah that I think it was it ASIS with their workstation board the head the 3 by 8 the x57 tws ace yeah yeah so that explains that board by 8 by 8 by 8 so the first 2 from the CPU and then the last one by the chipset yeah I remember seeing that board and thinking how are they doing that you can't get more than 4 by a chipset but it turns out based on the chipset diagram as we learn of yesterday that yes you can because that's the 15 what version of the chipset which we assume is going to go into all the thread Ripper boards as well yeah when thread Ripper Rison third gen Xen - seems like yeah I think that they probably would have done a PCI Express 4.0 by 8 interface and with a PCI Express 4.0 by 8 interface between the chipset and the CPU you can do a lot more you can run all of your envy me off of that plus all your other peripheral and have a lot more flexibility in terms of board design but it probably would cost more so the family they're using the same essentially the same floor plan for the IO die and the chipset could mean that in the future that becomes a configurable option for motherboard vendors so instead of using the by Ford link from City from CPU to chipset they can instead use the buy eight link and no sacrifice maybe the nvme lanes from the cpu in order to do so now that would be amazing yeah that really would and that opens up a wide range of possibilities that is essentially born from this sort of chip lit and reusable chip lat architecture yeah something which Intel doesn't there's still stuck on PCI Express 3.0 by four which is looking increasingly paltry by comparison I mean you can saturate that with just a couple of 10-gig USB devices and your envy me so the benefit that Intel has there is that their chipsets uz1 72 73 73 90 they are essentially big PCI switches I mean they have what's called HS i HS io lanes up to 24 I think tis 28 now and it'll say well these are all essentially PCIe but some of them can be used for in Ethernet some of them can be used for USB 3.1 gen2 or 3.2 Gen 2 and yeah you so you can mix and match with the controllers Intel definitely has an advantage there yeah but they've had time to work on it and put more money into it than AMD has well I even still see the same kind of an interface at the Xeon level which is a perhaps a bit disconcerting for being able to turn on a dime and do something a little more creative although I guess you've got the PCI interface and they've added 64 lanes at the very top end for the workstation for then using on W I think so that part interests me because it's the same silicon as you know the xeon cascade the cascade like sealed scalable but the cascade legacy on scalable have 48 PCIe lanes and these have 64 but it's the same silicon and yet and the reason why that is is because the design is actually does have 64 but ultimately originally 16 was meant to be dedicated for on a path like when they used to put on the path on the on the on the kids now they've been able to reroute those through pins that already existed so it makes you wonder how many of those pins very except on doing anything so is this an Apple driven innovation or is this something where AMD's come along and said well hey we've got a hundred and twenty eight they win that they need to compete speaking of all those PCIe lanes I'm not sure if you saw a complex but we saw an AMD system that was using the Intel ruler SSD form via in a two you have a hundred and eight drives that's the petabyte one you or two you server well it's it's just about it's AMD using all those PCIe lanes with Intel small factory no how do we make this start it's I mean ultimately it means that those drives are less hampered by the PCI bandwidth upstream yeah I mean ultimately you're not gonna be raised in the peak bandwidth software-defined storage yeah but that's how he goes you'd be surprised how many people are not thinking in terms of redundancy instead of a single server anymore a lot of the a lot of the moves that I saw a Computex four servers or that people really don't care anymore it's like I've got a full rack looked if I lose half the rack the software redundancy is gonna take care of it so I don't care about any intern machine redundancy at all as long as I've not got somebody that comes to this rack every day to deal with a hardware failure and Google's got that live migration feature for Google cloud and that's really amazing yeah yeah well I mean any software stack worth its salt is just swarms and swarms of relatively tiny VMs and that on an epic first-gen that was actually a better use case than large monolithic VMs is just a swarm of small teams because then you have to worry about can't coherency and cash across socket cache coherency in this kind of stuff yeah and then we get a situation where competing vendors aren't saying well you've only got an eight-core die what if you have a tank or VM sort of thing right which was which was part of the dialogue we had last year in the year before yeah I don't think that's gonna be a concern anymore no so speaking of server CPU is even though it's not related to the event why we're here in LA but back in Computex we saw a system with the new Chinese epic i GaN Zen one died licensed yeah that was Internet you got a closer look at that than I did you got to find out some of the the internals you were mentioning about the encryption system yeah so what happened is AMD and Phatak which is like a research arm of the Chinese government that's a very high level definition they formed a joint venture with you know for 51 49% ownership because if you want to work in China that's the sort of thing you have to do and they transferred the license of the Zen one core basically to this joint venture and the joint venture designed an epic like system now that joint venture can change parts of the core so it's not an architectural license it's like a core license with with approved modifications you can copy-paste and change a little bit yeah but those changes have to be approved by AMD which is something I learned recently so what they've done is they've taken out the American encryption engines and put in the Chinese ones makes sense you know it's you know they just have different algorithms for doing crypto and that's one change we've been told that there are other changes but AMD won't tell us what they are they said we're going to have to get a chip and to probe them and if we find any we can ask AMD and they may confirm them which is gonna be fun anybody have any of those in the audience yeah we kind of yeah we need we need the data it's do you follow in statics 64 on Twitter he's he's a guy that Damacy this does all the instruction Layton sees and throughputs oh no if every processor and yeah that when he gets hold of a new chip so when I did the 10 nanometer can the lake review last year last year earlier this year yeah so long yeah he worked with me I were I worked with him he worked with me on that to find out what you know some of the differences in the kennel microarchitecture and to be honest there are a few things that are in that which aren't in isolate that's surprising you know it's because Intel seems to life heated and then cut one of its poets traffic I don't know so I mean yeah I think that there was there was a shelling of talent and production at Intel and some things were lost or are in the process of migration that's probably a fair assessment it's it's you know I I had a conversation with Intel engineers a few years ago and I said are we at a single thread performance wall yeah which is kind of I knew the answer the answer's no but I just wanted to hear what their response was and they said yeah we'll give us a couple years and we'll give you a 20% IPC 18 percent but not bad no this was quite regression well so this was a few years before so we eventually got that that's now what five minutes ago literally but that's because sunny Cove on the new tenant 10m meter was delayed because it was tied to process yeah so Intel's new strategy and their new CTO dr. Marty render team Charla and said we're going to disaggregate microarchitecture to process so they're going to work on the best microarchitecture possible and then implement it on the best process available at the time which is what you should do when you're a company the size of Intel and that's the effect that complacency has had on Intel over the last 10 years if they were competing like for like every year for the last ten years this would have been implemented throughout the game these applets probably would have been implemented forever ago - I mean Chiclets are you have to do that in order to get the yields that you need to be profitable especially on a new process notes yeah and so speaking of chip lists where we're seeing more and more Chiclets being you know kind of showcased at events now Intel's future GPGPU which is coming on seven an ohmmeter they've kind of shown a it's not a block diagram it's an artistic representation of what they're doing dori behind that is really weird yes the artist did the thing before the thing and in it he'll was like ooh that's nice and then you do the thing with a little bit more knowledge so but that's that's using Chiclets interconnected with email+ overall stacking yeah so that's new packaging technologies with Chiclets in video just showcased her paper at the vsi conference in kyoto where that's their AI processor so trying to move you know AI necessary out of the general GPU and into its own dedicated up to 128 tops and that's that's using up to sixteen shifflet's as well that's a paper I really need to like fine-tooth comb through because that's really interesting cuz that's designed to be scalable from mobile all the way up to datacenter even this event from AMD which you know it was I don't I don't know if it was watered down but compared to like other press events it was much less watered down we learned about stuff that they were doing with their packaging technology like the organic substrate that's the carrier for their Chiclets and the manufacturing tolerances yeah I don't like that that went into it it's really it's getting kind of scary in terms of like manufacturing and repeatability and yeah what did they say only there are only two fabs in the world that have the capability in the technology to do micro bumps for seven nanometer the way they needed them yeah to in the world yeah and it sounded like they coached one of them a lot on being able to get usable substrates yeah so they've had to increase the the package substrate from up from 1012 layers to 14 layers because the PTI for signaling that sort of thing and the micro bump diameter is down to 150 micrometers from 180 it's 150 to 130 150 130 Wow and the way they're enabling that is instead of just using the solder the tin silver solder balls is by using epitaxial growth for copper pillars yes I'm putting my chemistry degree to good use they're literally growing copper yeah it's it you got to do what you got to do makes me really sound like yeah yeah mid 20th century communist how to do what are you gonna do it's true it's an exciting time in the market like most people I just want to play my games faster they don't understand the level of insanity that has done many stuff here and there's we met a lot of the engineering team I mean AMD's are relatively you know I think I think you joked is like not everybody should take the same return flight relatively small huge fortune 500 company yeah and it was really it was really interesting to see the heart of so many people like it's so easy to tell like Mike Clark and Scott Watson and and so many of the other guys that have poured so much of themselves into these products and I really you know regardless of how the market receives it I've really done a job for the relative you know David versus Goliath type situation we find ourselves in in terms of engineering and fabrication well this is one of the things I've learned over the 10 years I've been in you know journalism it's from the outside everything's a black box you just buy the part on the shelf you maybe you you know go into forums and speak about it but you just buy a product on the shelf I've been spending the last few years here trying to pull back the curtain on these black box companies you there are people there that are emotionally invested and technologically invested in what they're doing day to day now AMD has 10,000 employees right they have whether they say six corporate Fellows of which four of them were on the stage the presentation a couple days ago you know speaking about how they manage their teams and how ideas come to fruition and you know these fellows especially one of them is Joe Macri who's essentially the father of GDD our six but you know there's not only Intel not only AMD Intel's the same yeah Qualcomm is the same or even OEM partners you know super micro MSI Asus gigabyte it's not only do they have no the engineers who it's their job they have the engineers who is their passion yeah and it's that's part of the industry I didn't see from the other side yeah if my twitter is anything to be believed even after people leave Intel or are forced out of Intel I don't know which they're still super passionate about Intel and you see that I cross all of the industry but then you can't you can't help but think if there's some disconnect between what the company is doing and what the engineers want to do because like the tej branch predictor yeah that was a great story that corporate fellows had a story about how that was technically something I guess was in three or something yeah and so it was like no we think we can get the branch predictor ends in two it's like all right yeah let's work twice as much work and a half as much time yeah but hey if it works it works yeah it's it's gonna be interesting twenty nineteen one way or the other for AMD there anything else you can think of anything else that our podcast viewers lecture listeners would right so out of all the news we posted yesterday that came out of AMD the one they've got the most traction was the price of the sixteen core yeah right seven hundred and forty nine bucks now it's we thought we knew it was coming but they didn't tell us the price before the press event now I was thinking what what would be a good price for that and I was thinking you know seven nine nine people would be happy with and there'd be relatively no complaints if they did at eight nine nine it would probably still be alright but there would be complaints seven four nine you know and and the twelve core is what four nine nine yeah so you're paying an extra fifty percent for four more cores and it's got a higher frequency at yeah and it comes with the nice 125 watt raised by cooler yeah that's suddenly Intel's high-end desktop market you know it kind of does make it sound like the battery on Intel as we're talking listen I bought a 79 I heard X with my own money and the value that's a thousand dollar processor whenever it came out I bought it on launch day because it was like 10 cores on the high-end desktop socket this is gonna be great you mean you didn't buy this $1700 one no I didn't buy the $1700 one and I was really excited about that processor for the large cache in the new process the next tuning on and all this kind of stuff its value has tanked like usually Intel processors hold their value pretty good it's it's worth about $300 on a used market and it's just because it's not it's 10 cores it's just not that good yeah the eight core the nother K is faster out of the box one of the things AMD said and I've heard I've had to say say it before computers you know we built they are so weak they said we AMD said that Zent who was built to intersect a very competitive 10 nanometer desktop part from Intel right and Intel isn't there they've stalled AMD doesn't care they're still going to come up with the better with their best part they're still going to put their best foot forward regardless of what the competition is doing then are gonna play tit for tat when it comes to you know saying well well this you know is this feature bad or is this feature better or you know just slowly incrementing performance trying to milk the most out of users they just want to put the high-end one you know in as possible you know still validated and making business sense and you know they said you know you see amazing things on the 16 core you're driving that into the market wait for thread Ripper it's gonna blow your socks off I can't wait sir dripper is gonna be amazing I mean I'm already running the 2990 I can run a Windows of him with the 2990 under Linux as you ma and basically clawback the performance to law that's lost from applications that don't handle that correctly or as part of the operating system that doesn't handle that correctly and with the IO die that's legit not a problem just seeing the IO die in a quad channel configuration and they're reported numbers on latency are honestly very impressive like given that the chain the memory and the chip lid depending another hull yeah yeah you're so far removed and the signal propagation delay and I mean you know like a mesh or a ring type network for direct access to memory is going to be very very fast but the numbers that they had on the slides were very impressive yeah for this change and with the larger caches it may not really matter for 99% of workloads yeah yeah it's it's what what why interest is going to be is comparing the Rome's the epic server stuff to the thread Ripper equivalents right in terms of frequency about you are they gonna play for you maybe a high TDP on the workstation so I'd refer stuff if funnily enough AMD has a new workstation part to their website now and they're actually hiring for a new head of workstation but the products excellent which is thread Ripper so it very very much looks like thread Ripper is going to be you know previously it was kind of that kind of gaming content creation mix they're gonna pivot all the way to Conte creation that's the thing that's born my mind about the 2990 is pretty much since launch day I've had that on in a Linux machine just crunching numbers doing computation to have stuff running running integrated classes and things okay running stuff like that and I open it up because I wanted to see how like long-term stable it was and it's been running at about 325 watts to 350 watts on the overclock which is really just PBO with you know the restrictions removed and as long as the temperatures managed it's using like 325 to 350 watts so we're talking about 3.9 to 4 gigahertz on all the cores all the time and it has been rock-solid I grant I'm using air correcting memory but it has been a rock-solid staple I've only had like two or three corrected ECC events over the last almost a year yeah so that is how much doing a gig ok yeah so it's it's running 2666 to an average for you as well it's yeah if if AMD say come out with a thread Ripper with 48 cause okay so that's that's six on each CC X for each ghost and you can PBO to four gigahertz yeah that's gonna be really exciting for a lot of reasons it's like if the system is not super busy and I can get like 4.5 on say eight cores I'm gonna be so happy now so the flip side is we think is we know AMD is coming out with this right yeah it Intel's contra to that you know is they're starting doing new things with packaging with Lakefield with the phobos chip stacking technology and they've got email and even though we haven't seen emailing like a dual high-powered design we no doubt it's going to come yeah we've seen a couple of things with silicon photonics that might only be two generations away yeah yeah it Intel will have an answer to this of that there is no doubt right it's for a company that spends more on R&D in a quarter than AMD makes revenue in a year well the sun's out make hay yeah yeah AMD's gonna reap the benefits and the epic Roman stuff is gonna be at the core of that that's gonna really drive revenue there and that's really great for AMD as an opportunity to invest more money into 90 the cpu but also the GPU stuff which to be honest needs some investment in I know we haven't spoken necessary about Navi right now Navi now - we can talk about that yeah it's yeah I mean principles for Navi a solid yeah it's just a combination of execution process note and everybody's gonna get the details about Navi wrong because of GCN and it's got it's it's got some GCN bits but our DNA really is fundamentally different and that was something AMD was keen to like underscore and point out like we've got new I say documentation coming out and we've got you know it's totally totally new Hardware implementation so yeah it's like a it's a weird sort of concept of the compute unit is different as well to be honest I've given everything that they said I reckon that our DNA is sort of like the trial run oh sorry Navi is the trial run for our DNA yeah there's our DNA - you know it's like the same D with the CPU stuff Naples was the trial run Rome is where the cash money is it's where all the cash is anyway yeah well I think that they're gonna have a license to print money with Rome especially in the second half of this year because there are a lot of data centers that are hanging on to those like v2 v3 v4 Zeon's that are overdue for an upgrade and Rome in terms of both legacy consolidation and new projects is going to make a lot of sense for the enterprise it's I guess we should mention that Xen to also comes with enhanced QoS yeah all machines yeah so one of the things that Intel introduced back in Broadwell is that you can get your VMs or your containers and you can allocate specific amounts of l3 or memory bandwidth so the concept of noisy neighbor in a cloud environment is is real you don't want somebody who's in the VM next to you hogging up all your resources so Intel implemented that with Broadwell II I think this idea of be at the hypervisor level being to control how much cash and how much memory bandwidth and now AMD has that was then - and that's gonna be a big kicker and we're in I'm looking forward to actually testing that they also mentioned the changes to like the TLB and that will have implications for some of like the game machine virtualization stuff that I'm doing yeah I really want to test that it it given that these features exist it makes me want to test different amounts of l3 yeah for so I mean I am denoting that double the l3 has good knock-on performance for gaming and the consumer stuff you know going from two mega pack or two for Megacorp but I kind of wanted to do a scathing article where it's just comparing how much that actually affects it so it'll be an interesting thing to test hmm but it's glad that these actually exist we've been told that you know they gave us three cash families of instructions that they've included in his m2 but we were told by Mike Clark the father of Zen father the Godfather I beat him at bowling last night but he said you know we actually have a lot more you know instructions included that are related to cash which would be given a hot chips the hot chips conference which I'll be at in August will you be a hot chips probably no that's a shame maybe I mean that August is one way way I wasn't planning on it but the hot chips is gonna be one of the one of the most exciting conferences this year if you're into the technical side yeah all these chips they're going to go through all the changes it's it's not only Zen but we're gonna see cascade Lake we're gonna see Intel's joint venture in China that side that's called gin ties there's way for scale machine learning so yeah so you have your machine learning chips that are you know tiny imagine doing it on the size of a wafer I mean it depends what size wafer you want but yeah that that's new technology that's coming through at the same conference last year we saw nanotube based DRAM and that's being deployed in Fujitsu systems now the MIT people that are not heating just release public details of their carbon nanotube based SRAM circuit 16 kilobytes and it's SRAM and it seems to be oh yeah it's a wafer wafer scale I thought I saw David Shore from wiki chip he was posting photos of the chip EC said he must get hold of it and then he did a line us and dropped it they did anyway no so it's it's exciting there's a lot there's a lot to unpack here there's a lot that we can probably discuss on future like a future thing so if you like this you should probably update it or share it or something I don't know so it's good well thanks again for for joining me or wait well you're gonna post it and I'm gonna post it to you yeah it's fine chance Wendell it's ears yes I like to it's much less complicated and much less awkward so we'll see you guys in the forum and or on tech tech potato photonic\n"