Its even worse for me in India. I was supposed to go to sleep about an hour back as I have to be in field at 530hrs(IST). Its going go be hard waking up or staying awake tomorrow.
We can guess quite a lot from this, for Zen 2, at least, if we assume that the same architecture will be used for the desktop (and mobile), which seems to me like a reasonable assumption. So, 8 cores per chiplet, probably 1-2 chiplets per desktop CPU, 2x floating point performance (and the other architectural enhancements discussed), PCIe 4.
The 2x floating point performance sounds like the double the size of their SSE engine. We are shown a 9.3% less time / 14.3% performance increase with the same number of cores in the Cray benchmark, but without clock frequencies for Rome (did I miss them?) it is hard to say how much is due to frequency advantage, the 33% memory bandwidth advantage, and the unknown IPC improvement. ---Speculation Begin--- The 7nm node gives them 25% faster transistor switching, but not all of that will translate into cpu frequency. They also said Rome wasn't final frequency, but that doesn't mean it didn't have a frequency advantage over the 8180M. Not that I ever expected a massive uplift like Zen over Excavator, but this should put to rest any rumors of a massive IPC uplift for Zen2 over Zen. We can probably take 14.3% as a rough ceiling with the likely IPC improvement being closer to the 5% - 10% range, though I'd love to be proved wrong.
That said, IPCs are no longer AMD's biggest issue anyways (frequency is more lacking at the moment). The 25% faster transistor switching helps, but it is also possible that AMD optimized critical paths to increase frequencies even further than just node improvements would allow. I doubt they've got something to match Intel's frequencies as they would probably be advertising such an improvement. I expect something in the neighborhood of 4.5GHz, but it would be nice to see something closer to 4.7GHz.
I hope they get some new graphics cards out soon. Rtx was a step backwards in performance/price. I have high hopes for navi in late 2019 but something for the meantime would be nice. And the 9700k looks competitive as a value buy from intel so some new cpus would be nice too.
Nvidia added a lot of hardware that lacks software support so they ended up with a very big die for the current perf. Once games start to support the 2 new types of cores, the perf will improve but for now, Nvidia can not do good value with this generation. For example the 2070 is 445mm2 so rather huge vs 314mm2 for 1080 and 1070. That's why prices are high and there isn't much that can be done about it in the short term.
Good point, no idea how much power those use and in the end, it's about perf per W not total power . We could wonder if the cooling solution is sufficient I suppose but they likely did provision for that.
Power limit will probably stay the same with the RT and Tensor cores active. But that means it will cut into the power budget available for boost.
A TUxxx GPU with RT, Tensor and Cuda cores all lit up at the same time has twice as much active silicon as on current games that only use the Cuda cores, so I'm expecting performance to take a hit once everything is enabled and utilized at the same time.
"And the 9700k looks competitive as a value buy from intel" If it ever reaches MSRP. 9th Gen CPUs in Germany are: 300€ for 6/6 9600k 480€ (tray) for 8/8 9700k / 500€ for boxed (guarantee) 630€ (tray) for 8/16 9900k / 700€ for boxed (guarantee)
AMD competition is 160€ for 2600 / 200€ for 2600x which has double the threads of the 9600k but loses on ST performance. And 260€ / 290€ for 2700 / 2700x.
Unless you absolutely have to OC them towards 5GHz or need AVX512 support, currently there is little room for Intel CPUs in the consumer space, from a price/perf. perspective. 8th Gen CPUs are a bit of an exception, but still below Ryzen2 imo. But that has never stopped anyone from buying Intel bevor. :D
AVX512 is only on skylake-x, which is a completely different class. Current intel mainstream CPUs can do full 256-bit AVX operations at once, while Ryzen has support for it, but needs 2 128bit operations, which is slower.
Hey guys when is this starting???????? Hey, I was wondering we if could took take a vote in honor of election day, on what we think the maximum price of epyc Rome will be????
zen 2 promises a lot. i don't even care if they come out and say "7nm can't be finished within our time window." as long as 10nm and architect improvements happen i think most of the regular community is happy. but we do need to see 4.5 ghz being a thing from amd and an ipc upgrade of at least 11 to see them really sqeeze intel. i hope, but you never can know. also, dont hope for 16 cores in main stream x470 - probably not happening.
I wonder how high off the ground they are. If they're reaching the horizon of their previous starting point every two years we could work out their ground speed, assuming they are on earth.
Will this event be live streamed anywhere or just Live Blogs? AFAIK, it's an "Investor Relations" event so they probably won't stream it on Twitch or YouTube but will it be somewhere on AMD's website just like their Earning calls etc?
Gamers are always super annoying when it comes to discussing AMD' priorities; the server market 's where the real money is. If AMD put thhe gaming market firstx they'd have never dug themselves out of their hole like they are now. Be patient, they will get to the gamer when the time is right, when it makes BUSINESS SENSE.
It's always better to go through business sector 1st, higher margin to cover lower yield from each wafer. Once the node is matured, they can start to do consumer grade parts cheap, with respectable margin. By the time it reach consumer, most bug and issue should be fixed. Gamers should know they aren't the biggest slice of the same pie.
Anyone know where it will be streamed? AMD's PR has reached a new level of weirdness with the ir.amd.com announcement page empty of details. I'd think they'd want to say where to find the event.
That's the current Vega 14nm based MI25's FP16 performance. The MI60 that was announced is a little better at 29.4, but one should still look at the FP32 performance instead.
Lisa Su is the one of the three people that makes me feel they truly believe in making great products, the other two being Steve Jobs and Pat Gelsinger.
I just want to see what the clock speeds for Zen 2 are and if/when they're releasing new high performance GPUs. I don't need to hear the fluff, AMD! Get to the point, LOL.
I hope that means they're making it faster and fixing the high latency issue associated with that particular interconnect. Closer to Intel Ring Bus and/or Mesh and AMD will really be making Intel sweat. Also, if price remains the same. I saw sweat because Infinity Fabric has really been that AMD-syndrome on this new lineup. In Piledriver, the AMD-syndrome was shared FPU units. But, if AMD is really, really saying what I mean they're saying by focusing energy into IF, then, I would bet that that syndrome would be gone and Ryzen would truly shine, IMO.
No they are not kidding you. Process improvements and design all go to the perf bottom line. High performance CPUs are designed for a particular process. AMD can't just magically shrink K8 to 10nm; it simply won't work out of the box. A tick is a pretty significant amount of work. So it is "holistic". It's not unique to AMD, but they're not kidding you.
Ok. So, CPU design has turned to lingo of Whole Foods product retailers. Not saying being "holistic" is not a good thing. But, why borrow that very Whole Foods-y lingo? And, not that I necessarily don't think Whole Foods is bad, either. But... ok, nevermind. This is the same company that came up with "Ryzen," "Threadripper" so I shouldn't get so uptight with the "H" word.
Just because a word has been appropriated by a "movement", does not mean we shouldn't use it where it's correct. holistic: relating to or concerned with wholes or with complete systems rather than with the analysis of, treatment of, or dissection into parts
It's more to do with lots of interfaces require more physical space so it's natural to use a larger process node for that. And have to vast cost benefits of much higher yields for the cores.
I always thought it was because AMD were betting the farm on 7nm, so in order to minimize the risk of the process having high defect rates initially they kept the things that don't scale down to 7nm as well on 14nm (memory controllers) to keep the chiplets small. If they're small then they'll yield well even if the defect rate is fairly high.
MI60 is still a vega, compares to previous node it is considerably smaller. As for Rome, each EPYC only need 1 IO die, while it need 4 chiplet, having chiplet on 7nm is a good move, almost the same real estate of a single io die makes each wafer match each set. Not a bad move at all. Man, AMD really wants to get the most out of each wafer, which is good in terms of cost and efficiency for manufacturing
They are betting hard that 7nm is faster and lower power than 14nm. They aren't betting it is cheaper (per transistor). They also aren't making any big bets on yield (which is one of the reasons it won't be cheaper), and thus are producing smaller chips (so process errors wipe out only one small chip at a time). Rumor has it that yield problems are the biggest problem with Intel's 10nm (performance might be the other big killer, existing 10nm chips aren't impressive).
There's also the issue of buying wafers from GF. It isn't clear if they will be on the hook to buy that many wafers, but they probably have to use GF fabs wherever possible.
It looks like a clever engineering solution to dealing with potential issues with a cutting edge process, and dealing with lingering political issues from the AMD-GF split.
It's also funny that they show a giant slide saying "Delivering as promised." And, perhaps, the roadmap shows that they did release Ryzen and Ryzen+ as promised. And, that they improved a lot from their prior CPU because their prior CPU was released in 2011. So, it is expected to be as much. Is that the promise? If, so, it isn't that great, IMO. And, let's not mention Vega. Was that "delivered as promise?"
My qualm is that the slide should have said "Delivered as promised, my ASS!"
AMD promised 40% IPC improvement over Excavator, and they achieved 52%. They actually delivered over what they promised. It's not about promising to launch Zen and Zen+, but to improve upon previous products... which they did.
Your complaints regarding Vega... the event wasn't about gaming, so we'll have to see what they say at CES.
They improved so much because buldozer was very slow to begin with. Let's not forget that Thuban which was a six core k8 based core was faster at 3.3ghz than 8150 at 4GHz. In any case Rome looks like good stuff
When I alluded to 2011 AMD CPU, and then Ryzen launch in 2017, I was thinking of Piledriver, of which I used to have as my main system FX8350. Of course, the numbers would be skewed to be a favorably high number when speaking of improvement because there is like a lotta years between 2011 and 2017. It's a duh moment which AMD has twisted into a Yay moment. That's fair. It's marketing doing their marketing stuff. But how hard was it to improve upon Piledriver when you have 6 years in between?
I also think that "12:54PM EST - Perf is holistic - the process technology and design all helps" is taking CPU design into the realm of vodoo-ism, shamanism, etc. And, I understand it matches the name of their architecture, Zen, like in Buddha. But, come on... are you kidding me?
No. Yeeeeman suggested/alluded to apologize the word an engineer might use. The phrase in question is"Holistic perf." I said, I doubt an engineer said those words. I could be wrong. But, whether an engineer said those words or a host of the event is not the point.
What are you referring to of what I said that is not of "substance?"
Because you are being a troll? Use a dictionary? Don't complain about the correct use of words? The entire conference was a succession of engineers (including Lisa Su being one)...
AMD couldn't very well launch a replacement for Bulldozer within three or four years of its original launch, they were stuck with it and needed to make it work for them. Additionally, Zen was a completely new design, and this design was ripped up and started again at least once.
Even if it seems easy to improve on Piledriver, it costs ridiculous amounts of money to keep pushing IPC up. Intel is TEN TIMES the size of AMD, yet they too have added more cores over time - a single core will only get you so far, and there hasn't been a noteworthy (in my honest opinion) uptick in performance with the Core architecture since Haswell.
Excavator itself was a large step above Piledriver, but shockingly, AMD chose to halve its L2 cache. I also believe that, had Piledriver done better (not that it actually really, truly bombed), and AMD had access to a better process than the 32nm SOI garbage they were on, we could've seen Steamroller and Excavator-based FX CPUs with L3 cache and much higher clock speeds than we got from 28nm bulk. They wouldn't have knocked Intel off the top spot, but who cares as long as they did well in specific workloads?
As yeeeeman mentioned, Thuban was a very impressive product for its time, but AMD realised that the venerable K8 core could only be pushed so far. Interestingly, I read that K8/10 has nine execution units per core, but can't retire ops fast enough to take advantage of this, so perhaps K10 could've been pushed further, but not without a substantial redesign to support newer instructions and higher clock speeds. Bulldozer made sense... in principle, at least.
Why does it always end up talking about "money?" For one, we don't know enough to know. But, these kind of conversations always ends up, as in defense of or something about money. Sorry if I am being too broad. But, convos regarding AMD... there is always someone talking about "budget," "money," "people were moved to Navi from the Vega GPU group," etc... Seems very, very, very insider... I mean, you have to really know to know this stuff. And, I doubt, anyone of us do. I am talking about budget, money stuff. Like, who cares.
For example, does a hypthetical AMD Ryzen 4800xxx CPU run faster knowing that it cost xxxx amount compared to the competition that might have been xxxx amount?
Well, we know enough to know that AMD was financially on its proverbial arse. They also had two CPU designs at the time, Bulldozer and "Cat", three if you count Stars (which they bled money for due to apparently overstating demand). They've also been on the same GPU architecture for seven years (TeraScale was four to five years), for better or worse.
You're right, we don't know, but isn't it a fair assumption?
Excavator came out in 2014 or 2015. Their last new uarch was Bulldozer and that was 2011. By that standard, Intel is still milking Conroe (2006). Interesting that they had made delivery promises to your ASS.
What standard? I was talking about the span between Bulldozer and Ryzen, 2011-2017... and AMD used the numbers for FX CPU and Ryzen to show how much they improved. And, you're right, it's as if Intel used a Core 2 Quad CPU from 2009-ish era performance to a recent 2018 Intel CPU and say, This is how good we are because that number is high.
While, were at it, if we look at performance of an FX-8350, 4-core/8-thread CPU, to a Ryzen 1500X, first gen Ryzen in real world apps and games, it's not that big. Plus, Ryzen introduced latency that didn't exist in the prior architecture.
A cheap shot. Charlie Demerjian is very smart and works very hard. That much is obvious, and I'd guess that as a result he's right far more often than not. His writing is occasionally overblown but that's style, not substance, and he earns the right to indulge in it.
Will someone pleeeezzzeee explain to me why AMD is not announcing Vega 7m as a gaming card with 1.25x which would be enough in most cases to match nvida's high end while they wait to launch navi et all??
Actually there is not, Nvidia and AMD make much more money off gaming gpus vs gpgpus. Data center is relatively low volume. It would be nice to see Vega 7nm for gaming
Comments like that make me laugh... no, the datacenter is the elephant in the room.
AMD told everyone they were going after business first.
The reason? because there is no alternative on the table from Nvidia.
AMD is finally launching their Infinity Fabric strategy on GPU and the big thing is scaling. AMD just said that 4 cards will work like 4 cards in computing, not 2.7X or less like Nvidia SLI technology.
Between the cost of the HMB and the overall design, it still won't compete with the 1080 series. Hopefully Navi will fix some of the architectural issues holding it back.
That's assuming "gaming card"=="desktop discrete GPU". I'd expect 7nm Vega to make a great notebook GPU, and am fairly curious why they didn't bother with that. Probably too much trouble to deal with consumer-level drivers (although wouldn't vega=vega?) and they are currently pushing 2200G/2600G as their "answer to all notebook problems".
This would also work as while datacenters can include discrete graphics cards (probably won't, but the threat would come up in negotiations), adding notebook GPUs would be much harder.
My guess is that they looked at the market, looked at how thin engineering was already spread and decided "we don't need to do that". Companies with relatively unlimited (compared to AMD) resources like Intel and Nvidia would probably make the notebook GPUs without a second thought.
Yeah, at first glance, seeing the hugeness of it might seem impressive. Such as when we first saw those Threadripper CPU's. But, upon further thought, it's like, so what? In computers, does size matter? To compare this to sports, if a long jumper was given more runway, and jumped 10 feet. And, another long jumper with less of a runway jumped similarly at 10 feet, who is better? In this case, the size of this thing makes me think of a long jumper in the former example with a really long runway.
"This one" is huge because, among other things, it has 8 memory channels and 128 PCIe lanes on the same package. Intel uses an external PCH and has 2 fewer memory channels. Skylake SP is physically larger than Haswell E5s because of 2 extra memory channels. Skylake SP is not small either. Check STH's photos. Epyc is maybe 25% larger.
Why is TR so huge? Most likely because AMD could just reuse the sunk cost of making the Epyc platform instead of designing a new one from scratch.
Weird, not even a word on demos provided at end of next horizon event (2 skylakes platimum 8180 @26k$ beaten by one aircooled prototype rome CPU on c-ray benchmark). Biaised or TBC ?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
104 Comments
Back to Article
Galdom - Tuesday, November 6, 2018 - link
And so it begins, the great battle of our time.nathanddrews - Tuesday, November 6, 2018 - link
When in Rome, do as the Romans.Meteor2 - Saturday, November 10, 2018 - link
If those benchmark are even half true that’s still astonishing. Zen 2 seems to blaze past whatever Intel’s latest Core uarch is.FreckledTrout - Tuesday, November 6, 2018 - link
This is not going to be a productive work day for me.The Chill Blueberry - Tuesday, November 6, 2018 - link
Haha, me neither.Chaitanya - Tuesday, November 6, 2018 - link
Its even worse for me in India. I was supposed to go to sleep about an hour back as I have to be in field at 530hrs(IST). Its going go be hard waking up or staying awake tomorrow.ToTTenTranz - Tuesday, November 6, 2018 - link
Ian is already giving us complete spoilers with those tags...Dadyal - Tuesday, November 6, 2018 - link
I hope we hear about next gen AMD high end gpus for gamers and Zen 2 cpus for desktopsSaturnusDK - Tuesday, November 6, 2018 - link
Unlikely. That will be discussed at the CES 2019 keynote.ET - Tuesday, November 6, 2018 - link
We can guess quite a lot from this, for Zen 2, at least, if we assume that the same architecture will be used for the desktop (and mobile), which seems to me like a reasonable assumption. So, 8 cores per chiplet, probably 1-2 chiplets per desktop CPU, 2x floating point performance (and the other architectural enhancements discussed), PCIe 4.BurntMyBacon - Thursday, November 8, 2018 - link
The 2x floating point performance sounds like the double the size of their SSE engine. We are shown a 9.3% less time / 14.3% performance increase with the same number of cores in the Cray benchmark, but without clock frequencies for Rome (did I miss them?) it is hard to say how much is due to frequency advantage, the 33% memory bandwidth advantage, and the unknown IPC improvement.---Speculation Begin---
The 7nm node gives them 25% faster transistor switching, but not all of that will translate into cpu frequency. They also said Rome wasn't final frequency, but that doesn't mean it didn't have a frequency advantage over the 8180M. Not that I ever expected a massive uplift like Zen over Excavator, but this should put to rest any rumors of a massive IPC uplift for Zen2 over Zen. We can probably take 14.3% as a rough ceiling with the likely IPC improvement being closer to the 5% - 10% range, though I'd love to be proved wrong.
That said, IPCs are no longer AMD's biggest issue anyways (frequency is more lacking at the moment). The 25% faster transistor switching helps, but it is also possible that AMD optimized critical paths to increase frequencies even further than just node improvements would allow. I doubt they've got something to match Intel's frequencies as they would probably be advertising such an improvement. I expect something in the neighborhood of 4.5GHz, but it would be nice to see something closer to 4.7GHz.
donwly - Thursday, November 29, 2018 - link
to watch lates movies or TV shows for free on your android devices then visit: https://tutuapp-apk.org/tutuapp-apk/ which free of cost.jjj - Tuesday, November 6, 2018 - link
Web Summit autoplaying video ad (with audio) on top of this page, not cool...Opencg - Tuesday, November 6, 2018 - link
I hope they get some new graphics cards out soon. Rtx was a step backwards in performance/price. I have high hopes for navi in late 2019 but something for the meantime would be nice. And the 9700k looks competitive as a value buy from intel so some new cpus would be nice too.imaheadcase - Tuesday, November 6, 2018 - link
Because RTX is about performance is why. Its the top of the line card, no one was going into the reviews and jumping to price point or power draw. loljjj - Tuesday, November 6, 2018 - link
Nvidia added a lot of hardware that lacks software support so they ended up with a very big die for the current perf. Once games start to support the 2 new types of cores, the perf will improve but for now, Nvidia can not do good value with this generation.For example the 2070 is 445mm2 so rather huge vs 314mm2 for 1080 and 1070. That's why prices are high and there isn't much that can be done about it in the short term.
ipkh - Tuesday, November 6, 2018 - link
If power draw is high now without the other cores being used/active, what happens to power limit once they do get used?imaheadcase - Tuesday, November 6, 2018 - link
Umm the cores are active now. Just not being used in the sense of the software is what he is saying.jjj - Tuesday, November 6, 2018 - link
Good point, no idea how much power those use and in the end, it's about perf per W not total power . We could wonder if the cooling solution is sufficient I suppose but they likely did provision for that.Voldenuit - Tuesday, November 6, 2018 - link
Power limit will probably stay the same with the RT and Tensor cores active. But that means it will cut into the power budget available for boost.A TUxxx GPU with RT, Tensor and Cuda cores all lit up at the same time has twice as much active silicon as on current games that only use the Cuda cores, so I'm expecting performance to take a hit once everything is enabled and utilized at the same time.
Death666Angel - Tuesday, November 6, 2018 - link
"And the 9700k looks competitive as a value buy from intel"If it ever reaches MSRP. 9th Gen CPUs in Germany are:
300€ for 6/6 9600k
480€ (tray) for 8/8 9700k / 500€ for boxed (guarantee)
630€ (tray) for 8/16 9900k / 700€ for boxed (guarantee)
AMD competition is 160€ for 2600 / 200€ for 2600x which has double the threads of the 9600k but loses on ST performance. And 260€ / 290€ for 2700 / 2700x.
Unless you absolutely have to OC them towards 5GHz or need AVX512 support, currently there is little room for Intel CPUs in the consumer space, from a price/perf. perspective. 8th Gen CPUs are a bit of an exception, but still below Ryzen2 imo. But that has never stopped anyone from buying Intel bevor. :D
AlexDaum - Wednesday, November 7, 2018 - link
AVX512 is only on skylake-x, which is a completely different class. Current intel mainstream CPUs can do full 256-bit AVX operations at once, while Ryzen has support for it, but needs 2 128bit operations, which is slower.danman81 - Tuesday, November 6, 2018 - link
Hey guys when is this starting???????? Hey, I was wondering we if could took take a vote in honor of election day, on what we think the maximum price of epyc Rome will be????The Benjamins - Tuesday, November 6, 2018 - link
Did you even read the title?Death666Angel - Tuesday, November 6, 2018 - link
Fewer questions marks, more reading of the headline please.danman81 - Tuesday, November 6, 2018 - link
It was a rehtoritcal question! I can't wait for it to start. What do you think prices will be like??SaturnusDK - Tuesday, November 6, 2018 - link
Around the $5000-6000 mark for the top Vega20 and EPYC2 parts, and decreasing from there.austinsguitar - Tuesday, November 6, 2018 - link
zen 2 promises a lot. i don't even care if they come out and say "7nm can't be finished within our time window." as long as 10nm and architect improvements happen i think most of the regular community is happy. but we do need to see 4.5 ghz being a thing from amd and an ipc upgrade of at least 11 to see them really sqeeze intel. i hope, but you never can know. also, dont hope for 16 cores in main stream x470 - probably not happening.The Benjamins - Tuesday, November 6, 2018 - link
AMD has not made any crazy promises, most of it is from the media.willis936 - Tuesday, November 6, 2018 - link
I wonder how high off the ground they are. If they're reaching the horizon of their previous starting point every two years we could work out their ground speed, assuming they are on earth.StayFrost04 - Tuesday, November 6, 2018 - link
Will this event be live streamed anywhere or just Live Blogs? AFAIK, it's an "Investor Relations" event so they probably won't stream it on Twitch or YouTube but will it be somewhere on AMD's website just like their Earning calls etc?tommo1982 - Tuesday, November 6, 2018 - link
I hope there is a live stream. Nothing beats seeing it with own eyes.amdrocking - Tuesday, November 6, 2018 - link
Yes. Go here:https://www.youtube.com/watch?v=XmciZH0FG-k
RoLukeSky - Tuesday, November 6, 2018 - link
So, how many hours left to begin?roadapathy - Tuesday, November 6, 2018 - link
Will AMD be here to announce the news or on their own news and announcement page?Dr.Neale - Tuesday, November 6, 2018 - link
It starts at noon Eastern Time.B18221 - Tuesday, November 6, 2018 - link
Gamers are always super annoying when it comes to discussing AMD' priorities; the server market 's where the real money is. If AMD put thhe gaming market firstx they'd have never dug themselves out of their hole like they are now. Be patient, they will get to the gamer when the time is right, when it makes BUSINESS SENSE.fastcx - Tuesday, November 6, 2018 - link
It's always better to go through business sector 1st, higher margin to cover lower yield from each wafer. Once the node is matured, they can start to do consumer grade parts cheap, with respectable margin. By the time it reach consumer, most bug and issue should be fixed. Gamers should know they aren't the biggest slice of the same pie.hoohoo - Tuesday, November 6, 2018 - link
Anyone know where it will be streamed? AMD's PR has reached a new level of weirdness with the ir.amd.com announcement page empty of details. I'd think they'd want to say where to find the event.The Chill Blueberry - Tuesday, November 6, 2018 - link
@IanCutress: Did you spot Linus yet?T1beriu - Tuesday, November 6, 2018 - link
The Youtube stream link is fake.Despoiler - Tuesday, November 6, 2018 - link
Valid live stream. https://www.youtube.com/watch?v=GwX13bo0RDQeva02langley - Tuesday, November 6, 2018 - link
25 TFLOPS?!!! OMG!PhilipJ - Wednesday, November 7, 2018 - link
That's the current Vega 14nm based MI25's FP16 performance. The MI60 that was announced is a little better at 29.4, but one should still look at the FP32 performance instead.iwod - Tuesday, November 6, 2018 - link
Lisa Su is the one of the three people that makes me feel they truly believe in making great products, the other two being Steve Jobs and Pat Gelsinger.wingless - Tuesday, November 6, 2018 - link
I just want to see what the clock speeds for Zen 2 are and if/when they're releasing new high performance GPUs. I don't need to hear the fluff, AMD! Get to the point, LOL.visualzero - Tuesday, November 6, 2018 - link
12:24PM EST - Focused a lot of energy into Infinity FabricBased on this https://www.anandtech.com/show/13124/the-amd-threa... this is more than correct ;)
wurizen - Tuesday, November 6, 2018 - link
I hope that means they're making it faster and fixing the high latency issue associated with that particular interconnect. Closer to Intel Ring Bus and/or Mesh and AMD will really be making Intel sweat. Also, if price remains the same. I saw sweat because Infinity Fabric has really been that AMD-syndrome on this new lineup. In Piledriver, the AMD-syndrome was shared FPU units. But, if AMD is really, really saying what I mean they're saying by focusing energy into IF, then, I would bet that that syndrome would be gone and Ryzen would truly shine, IMO.duploxxx - Wednesday, November 7, 2018 - link
for Epyc2 IF has its own dedicated speedKepe - Tuesday, November 6, 2018 - link
No live stream and now Anandtech is broken -_-Kepe - Tuesday, November 6, 2018 - link
Seems to be working again. And there's even a live stream linked by Despoiler.AlexonTech - Tuesday, November 6, 2018 - link
You can watch a live video stream of the event herehttps://www.youtube.com/watch?v=GwX13bo0RDQ
eva02langley - Tuesday, November 6, 2018 - link
Multi-GPU is not dead! Could it be coming to console next year?ishould - Tuesday, November 6, 2018 - link
I hope we get an IPC % increase number for Zen 2. Put an end to all the rumorswurizen - Tuesday, November 6, 2018 - link
@ "12:54PM EST - Perf is holistic - the process technology and design all helps"Omg! No way. Are you kidding me? BS!
wurizen - Tuesday, November 6, 2018 - link
At 12:54, it was stated that "performance is holistic."Are you kidding me?
deltaFx2 - Wednesday, November 7, 2018 - link
No they are not kidding you. Process improvements and design all go to the perf bottom line. High performance CPUs are designed for a particular process. AMD can't just magically shrink K8 to 10nm; it simply won't work out of the box. A tick is a pretty significant amount of work. So it is "holistic". It's not unique to AMD, but they're not kidding you.wurizen - Thursday, November 8, 2018 - link
Ok. So, CPU design has turned to lingo of Whole Foods product retailers. Not saying being "holistic" is not a good thing. But, why borrow that very Whole Foods-y lingo? And, not that I necessarily don't think Whole Foods is bad, either. But... ok, nevermind. This is the same company that came up with "Ryzen," "Threadripper" so I shouldn't get so uptight with the "H" word.Zoolook - Thursday, November 8, 2018 - link
Just because a word has been appropriated by a "movement", does not mean we shouldn't use it where it's correct.holistic: relating to or concerned with wholes or with complete systems rather than with the analysis of, treatment of, or dissection into parts
edzieba - Tuesday, November 6, 2018 - link
Uncore still on 14nm? No magical solutions to get big 7nm dies then, same issues Intel are having with prices scaling.edzieba - Tuesday, November 6, 2018 - link
s/prices/process.SaturnusDK - Tuesday, November 6, 2018 - link
It's more to do with lots of interfaces require more physical space so it's natural to use a larger process node for that. And have to vast cost benefits of much higher yields for the cores.ptmnc1 - Tuesday, November 6, 2018 - link
I always thought it was because AMD were betting the farm on 7nm, so in order to minimize the risk of the process having high defect rates initially they kept the things that don't scale down to 7nm as well on 14nm (memory controllers) to keep the chiplets small. If they're small then they'll yield well even if the defect rate is fairly high.That MI60 GPU looks pretty big though.
fastcx - Tuesday, November 6, 2018 - link
MI60 is still a vega, compares to previous node it is considerably smaller. As for Rome, each EPYC only need 1 IO die, while it need 4 chiplet, having chiplet on 7nm is a good move, almost the same real estate of a single io die makes each wafer match each set. Not a bad move at all. Man, AMD really wants to get the most out of each wafer, which is good in terms of cost and efficiency for manufacturingwumpus - Wednesday, November 7, 2018 - link
They are betting hard that 7nm is faster and lower power than 14nm. They aren't betting it is cheaper (per transistor). They also aren't making any big bets on yield (which is one of the reasons it won't be cheaper), and thus are producing smaller chips (so process errors wipe out only one small chip at a time). Rumor has it that yield problems are the biggest problem with Intel's 10nm (performance might be the other big killer, existing 10nm chips aren't impressive).There's also the issue of buying wafers from GF. It isn't clear if they will be on the hook to buy that many wafers, but they probably have to use GF fabs wherever possible.
It looks like a clever engineering solution to dealing with potential issues with a cutting edge process, and dealing with lingering political issues from the AMD-GF split.
wurizen - Tuesday, November 6, 2018 - link
It's also funny that they show a giant slide saying "Delivering as promised." And, perhaps, the roadmap shows that they did release Ryzen and Ryzen+ as promised. And, that they improved a lot from their prior CPU because their prior CPU was released in 2011. So, it is expected to be as much. Is that the promise? If, so, it isn't that great, IMO. And, let's not mention Vega. Was that "delivered as promise?"My qualm is that the slide should have said "Delivered as promised, my ASS!"
silverblue - Tuesday, November 6, 2018 - link
AMD promised 40% IPC improvement over Excavator, and they achieved 52%. They actually delivered over what they promised. It's not about promising to launch Zen and Zen+, but to improve upon previous products... which they did.Your complaints regarding Vega... the event wasn't about gaming, so we'll have to see what they say at CES.
yeeeeman - Tuesday, November 6, 2018 - link
They improved so much because buldozer was very slow to begin with. Let's not forget that Thuban which was a six core k8 based core was faster at 3.3ghz than 8150 at 4GHz. In any case Rome looks like good stuffwurizen - Tuesday, November 6, 2018 - link
When I alluded to 2011 AMD CPU, and then Ryzen launch in 2017, I was thinking of Piledriver, of which I used to have as my main system FX8350. Of course, the numbers would be skewed to be a favorably high number when speaking of improvement because there is like a lotta years between 2011 and 2017. It's a duh moment which AMD has twisted into a Yay moment. That's fair. It's marketing doing their marketing stuff. But how hard was it to improve upon Piledriver when you have 6 years in between?I also think that "12:54PM EST - Perf is holistic - the process technology and design all helps" is taking CPU design into the realm of vodoo-ism, shamanism, etc. And, I understand it matches the name of their architecture, Zen, like in Buddha. But, come on... are you kidding me?
yeeeeman - Wednesday, November 7, 2018 - link
Engineers are not very good with words so it doesn't matter what approach they used. All it matters is that the product is successfulwurizen - Wednesday, November 7, 2018 - link
I doubt an engineer was speaking at that event.Alistair - Wednesday, November 7, 2018 - link
man you really enjoy farting around saying nothing of substance don't you....wurizen - Wednesday, November 7, 2018 - link
No. Yeeeeman suggested/alluded to apologize the word an engineer might use. The phrase in question is"Holistic perf." I said, I doubt an engineer said those words. I could be wrong. But, whether an engineer said those words or a host of the event is not the point.What are you referring to of what I said that is not of "substance?"
Alistair - Thursday, November 8, 2018 - link
Because you are being a troll? Use a dictionary? Don't complain about the correct use of words? The entire conference was a succession of engineers (including Lisa Su being one)...Alistair - Thursday, November 8, 2018 - link
You should be embarrassed about your incomprehensible comments. AMD-syndrome? Long jumper? Holistic? ASS? Wurizen the ASS?silverblue - Wednesday, November 7, 2018 - link
AMD couldn't very well launch a replacement for Bulldozer within three or four years of its original launch, they were stuck with it and needed to make it work for them. Additionally, Zen was a completely new design, and this design was ripped up and started again at least once.Even if it seems easy to improve on Piledriver, it costs ridiculous amounts of money to keep pushing IPC up. Intel is TEN TIMES the size of AMD, yet they too have added more cores over time - a single core will only get you so far, and there hasn't been a noteworthy (in my honest opinion) uptick in performance with the Core architecture since Haswell.
Excavator itself was a large step above Piledriver, but shockingly, AMD chose to halve its L2 cache. I also believe that, had Piledriver done better (not that it actually really, truly bombed), and AMD had access to a better process than the 32nm SOI garbage they were on, we could've seen Steamroller and Excavator-based FX CPUs with L3 cache and much higher clock speeds than we got from 28nm bulk. They wouldn't have knocked Intel off the top spot, but who cares as long as they did well in specific workloads?
As yeeeeman mentioned, Thuban was a very impressive product for its time, but AMD realised that the venerable K8 core could only be pushed so far. Interestingly, I read that K8/10 has nine execution units per core, but can't retire ops fast enough to take advantage of this, so perhaps K10 could've been pushed further, but not without a substantial redesign to support newer instructions and higher clock speeds. Bulldozer made sense... in principle, at least.
wurizen - Wednesday, November 7, 2018 - link
Why does it always end up talking about "money?" For one, we don't know enough to know. But, these kind of conversations always ends up, as in defense of or something about money. Sorry if I am being too broad. But, convos regarding AMD... there is always someone talking about "budget," "money," "people were moved to Navi from the Vega GPU group," etc... Seems very, very, very insider... I mean, you have to really know to know this stuff. And, I doubt, anyone of us do. I am talking about budget, money stuff. Like, who cares.For example, does a hypthetical AMD Ryzen 4800xxx CPU run faster knowing that it cost xxxx amount compared to the competition that might have been xxxx amount?
silverblue - Thursday, November 8, 2018 - link
Well, we know enough to know that AMD was financially on its proverbial arse. They also had two CPU designs at the time, Bulldozer and "Cat", three if you count Stars (which they bled money for due to apparently overstating demand). They've also been on the same GPU architecture for seven years (TeraScale was four to five years), for better or worse.You're right, we don't know, but isn't it a fair assumption?
(assuming you care, of course.)
deltaFx2 - Wednesday, November 7, 2018 - link
Excavator came out in 2014 or 2015. Their last new uarch was Bulldozer and that was 2011. By that standard, Intel is still milking Conroe (2006). Interesting that they had made delivery promises to your ASS.wurizen - Thursday, November 8, 2018 - link
What standard? I was talking about the span between Bulldozer and Ryzen, 2011-2017... and AMD used the numbers for FX CPU and Ryzen to show how much they improved. And, you're right, it's as if Intel used a Core 2 Quad CPU from 2009-ish era performance to a recent 2018 Intel CPU and say, This is how good we are because that number is high.While, were at it, if we look at performance of an FX-8350, 4-core/8-thread CPU, to a Ryzen 1500X, first gen Ryzen in real world apps and games, it's not that big. Plus, Ryzen introduced latency that didn't exist in the prior architecture.
So, yeah, Delivering as promise... my ass!!!
silverblue - Thursday, November 8, 2018 - link
I'm not sure you understand what a promise is.eva02langley - Tuesday, November 6, 2018 - link
64 cores... nice try Intel. Keep going with your benchs by disabling SMT and comparing 32 cores with 96 cores... and claiming 3.4X performances.DigitalFreak - Tuesday, November 6, 2018 - link
You sat next Charlie? I'm feel sorry for you. Hopefully the NeverAccurate doesn't rub off.Arbie - Tuesday, November 6, 2018 - link
A cheap shot. Charlie Demerjian is very smart and works very hard. That much is obvious, and I'd guess that as a result he's right far more often than not. His writing is occasionally overblown but that's style, not substance, and he earns the right to indulge in it.WinterCharm - Tuesday, November 6, 2018 - link
And Intel is Carthage.dan_ger - Tuesday, November 6, 2018 - link
Will someone pleeeezzzeee explain to me why AMD is not announcing Vega 7m as a gaming card with 1.25x which would be enough in most cases to match nvida's high end while they wait to launch navi et all??Dug - Tuesday, November 6, 2018 - link
Because there is more money to be made in datacenters than selling gpu'sdan_ger - Wednesday, November 7, 2018 - link
Actually there is not, Nvidia and AMD make much more money off gaming gpus vs gpgpus. Data center is relatively low volume. It would be nice to see Vega 7nm for gamingeva02langley - Thursday, November 8, 2018 - link
Comments like that make me laugh... no, the datacenter is the elephant in the room.AMD told everyone they were going after business first.
The reason? because there is no alternative on the table from Nvidia.
AMD is finally launching their Infinity Fabric strategy on GPU and the big thing is scaling. AMD just said that 4 cards will work like 4 cards in computing, not 2.7X or less like Nvidia SLI technology.
People dont get it right now, but this is huge.
wumpus - Wednesday, November 7, 2018 - link
Between the cost of the HMB and the overall design, it still won't compete with the 1080 series. Hopefully Navi will fix some of the architectural issues holding it back.That's assuming "gaming card"=="desktop discrete GPU". I'd expect 7nm Vega to make a great notebook GPU, and am fairly curious why they didn't bother with that. Probably too much trouble to deal with consumer-level drivers (although wouldn't vega=vega?) and they are currently pushing 2200G/2600G as their "answer to all notebook problems".
This would also work as while datacenters can include discrete graphics cards (probably won't, but the threat would come up in negotiations), adding notebook GPUs would be much harder.
My guess is that they looked at the market, looked at how thin engineering was already spread and decided "we don't need to do that". Companies with relatively unlimited (compared to AMD) resources like Intel and Nvidia would probably make the notebook GPUs without a second thought.
haukionkannel - Thursday, November 8, 2018 - link
Too expensive for Gaming card?pogostick - Tuesday, November 6, 2018 - link
The die shot looks like an Atari 2600 race car.pogostick - Tuesday, November 6, 2018 - link
Package shot, ratherabufrejoval - Wednesday, November 7, 2018 - link
Eight socket server at the size of a credit card: Wow!wurizen - Wednesday, November 7, 2018 - link
Yeah, at first glance, seeing the hugeness of it might seem impressive. Such as when we first saw those Threadripper CPU's. But, upon further thought, it's like, so what? In computers, does size matter? To compare this to sports, if a long jumper was given more runway, and jumped 10 feet. And, another long jumper with less of a runway jumped similarly at 10 feet, who is better? In this case, the size of this thing makes me think of a long jumper in the former example with a really long runway.Is Infinity Fabric making that thing so huge?
wurizen - Wednesday, November 7, 2018 - link
I guess, my question is, why is Threadripper and this one so huge?deltaFx2 - Wednesday, November 7, 2018 - link
"This one" is huge because, among other things, it has 8 memory channels and 128 PCIe lanes on the same package. Intel uses an external PCH and has 2 fewer memory channels. Skylake SP is physically larger than Haswell E5s because of 2 extra memory channels. Skylake SP is not small either. Check STH's photos. Epyc is maybe 25% larger.Why is TR so huge? Most likely because AMD could just reuse the sunk cost of making the Epyc platform instead of designing a new one from scratch.
deltaFx2 - Wednesday, November 7, 2018 - link
Actually, I was mistaken. It's not 25%, more like 10% larger. Check these side-by-side photos: https://www.servethehome.com/intel-offers-enterpri...dynamis31 - Thursday, November 8, 2018 - link
Weird, not even a word on demos provided at end of next horizon event (2 skylakes platimum 8180 @26k$ beaten by one aircooled prototype rome CPU on c-ray benchmark). Biaised or TBC ?dynamis31 - Thursday, November 8, 2018 - link
nope, my fault, mention is done indeed (webpage faulty loading ?).Kujar - Tuesday, June 18, 2019 - link
Follow herer to watch online or download latest movies and TV shows free, https://showboxdownload.online/apk/Merjala - Sunday, June 30, 2019 - link
This is also a very good post which I really enjoyed reading. Download latest apk aps for android from AC market app store. https://www.acmarket-download.net/android/Megal - Tuesday, August 13, 2019 - link
Great Blog. Download latest MOD apk apps free for android from https://acmarket-apk.org/download/FredJ - Thursday, October 10, 2019 - link
Get latest Movies and TV shows free from Moviebox app.https://moviebox.info/