For Samsung, it's just business. Hell, Samsung and Apple have been suing each other for years, yet Apple keeps going back to them for CPUs and memory.
For Samsung, it's just business. Hell, Samsung and Apple have been suing each other for years, yet Apple keeps going back to them for CPUs and memory.
Gikaseixas
Platinum Member
- Jul 1, 2004
- 2,836
- 218
- 106
Frozen I never said their 14nm is the best thing since sliced bread, but it sure must be better than 28nm AMD is stuck right now.
Also, why think that AMD would be the sole benefactor? AMD is great in the graphics dept and own a bunch of IP's. Let that sink in
Last edited:
DrMrLordX
Lifer
- Apr 27, 2000
- 23,222
- 13,300
- 136
No. But then I can't quote you many logic parts that Samsung has manufactured in that past, period, so that doesn't mean much now does it?
Hmm. Well we'll find out within a 3-month period (or sooner) of any money changing hands then.
China's stake is in IP. People don't just hand out money for nothing. They're looking for designs to rip off, that'd be my guess. They're tired of ripping off MIPS and the money would be a way to prevent AMD from suing them over infringement issues in the future.
When it comes to most of AMD's processor tech, the technical merits are lacking in some (read: many) departments, at least compared to Intel offerings. When it comes to dGPUs and iGPUs, AMD's tech stacks up quite well against anyone's offerings, even Nvidia's. Market be damned, the GPUs they make work and work nicely. It makes total sense that someone would want to license their graphics technology. AMD's marketing department may stink but let's not slag the engineers and their work on that account.
And I'll say it again: Steamroller's market failure is due in large part to GF's inability to provide a suitable process for performance CPUs. IBM rolled out 22nm SOI for POWER8 (albeit too late for the Steamroller launch window) and GF just sort of stuck its thumb up its arse and whistled Dixie.
You can argue markets all day long, but in the end, this is a technical forum. Can we at least focus on one or the other instead of trying to claim that a design is a "failure" just because it isn't selling enough units? Hell the P4 outsold K8, what do you think about that?
Now how do you know that? IBM "sold" some of their fabs and sent some of their engineers to GF, but the timing would have been prickly; POWER8 didn't even paper launch until, what, mid 2014? And GF didn't pick up those fabs until this past October.
Regardless, one thing that must be kept in mind is that "rabbit hole" would have required AMD to wait until mid 2014 at the earliest to launch their Steamroller products, and that's assuming GF would roll out working parts as quickly as IBM was able to produce the first POWER8 chips.
I have no idea if HDL would have worked with the 22nm SOI process; I wasn't going in that direction. I suspect IBM's 22nm SOI has superior transistor density to 28nm planar as used with Kaveri. What AMD *could* have done on a process better than 28nm planar is continue to improve the 3M and 4M product lineup on AM3+ (or a successor socket) and G34, which up until now was the bread-and-butter of their flagging CPU lineup. 28nm planar really did not leave them many options when it came to producing viable successors to FX or Opteron.
GF didn't deliver 28nm SOI/PDSOI/FDSOI/anythingSOI in an acceptable timeframe or at an acceptable price, and that really screwed AMD. If there is anything AMD did wrong, it was producing a design like Piledriver in the first place, which more-or-less made them dependent on future high-performance nodes that GF was obviously not equipped to deliver. Now that they've finally got things going pretty good on a per-module basis with Steamroller and (in a few months) Excavator, they don't even have a suitable node for production of FX or Opteron chips using those designs, which is absurd.
Nope, a relevant fact like this would be announced independently of the quarterly EC.
The only victory that matters is the commercial victory, what's the point of designing a technologically superior product if you cannot pay the engineer's bills for the next round?
That AMD management is very, very, very incompetent.
IBM 22nm time frame explains Steamroller on 28nm but not Excavator: If node is *the* problem of the product then why isn't AMD moving Carrizo to IBM 22nm SOI, or developing an Excavator Opteron with IBM 22nm SOI? GLF has access to that node because of the Common Alliance, it didn't need to wait to acquire IBM fabs to do that. All AMD had to do was to ask GLF to implement the node and commit itself to a minimum wafer quota.
The fact that they chose not to proceed on that fashion tells us that the main problem isn't the manufacturing node, there's much more problems with the basic concept of the Bulldozer family. Remember that AMD itself published a slide complaining about design complexity of the current designs of the time and promising to develop simpler chips in the future.
DrMrLordX
Lifer
- Apr 27, 2000
- 23,222
- 13,300
- 136
Hence "or sooner". But whateva.
Whether or not there's any point to having a technologically superior (or in this case, technologically pretty good) product is . . . not the point! The tech is good, especially in the case of AMD's GPUs, which is why someone outside of AMD might want to license it. You called their designs "subpar" which is a direct shot at the engineers and their work, which is just not factually accurate.
That, and Intel was good at bribing OEMs not to use AMD products back in the P4 days . . .
Cost and failure-of-confidence. It was cheaper to keep flogging 28nm (with HDL + AVFS) for Carrizo while narrowing its market focus. SOI = $$$. You would have to be able to produce something with a pretty good sum total profit margin per CPU (such as an Opteron product that people would actually want to buy) to justify the expense involved. And, as you mentioned, there's no way for us to know if the 22nm SOI process was suitable for GCN cores so that might have shot it in the foot for Carrizo anyway.
AMD short-circuited their entire FX/Opteron ecosystem. I don't think they had the money to keep up the design work that would have been necessary to produce a competent Opteron using Excavator after their server revenue went into the toilet. Had Steamroller Opteron on 28nm somekindofSOI actually happened, the ecosystem would be more alive than it is today and moving Excavator to 22nm SOI would make perfect sense (and the money would be there to pay for the design work . . . well, maybe).
Here we see the failure-of-confidence. After Bulldozer and Piledriver, I don't think AMD management was willing or able to go in front of shareholders and tout any derivitive design as having merit. So they've relegated the Construction cores to APU duty. Anyone that has bothered to use Steamroller (or just pay attention to benchmarks where xOP and/or AVX are in use) can see that Steamroller has a pretty big jump in fp performance per module at the very least over a Vishera product. It's still a pittance compared to Haswell + AVX2, but vs older Sandy products it's actually not that bad.
The bitter irony here is that AMD's engineers snuck that performance into what has become an almost-throwaway consumer processor running software that rarely supports the latest and greatest SIMD instruction sets. In the server ecosystem, Steamroller would have been better-supported. Yeah Haswell-E would still have pwned it, but not by as much as it pwns Piledriver.
Once some significant benches and power numbers are leaked it will be a shock for the eventual urban legends believer...
AMD s soon to come 28nm Carrizo has better perf/Watt than its core M counterpart, be it in CPU or in GPU, this make it the CPU of choice for any efficency centered APU such as Nitendo s who are targeting lower cost/lower power items than the XBox or the PS4.
Certainly not. There are many, many ways management can take to screw up the most stellar engineering team you can think of. Did managers provide enough engineers as requested by the team managers? Did management provide the tools as requested by the team managers? But more important, did the set of design premises provided by AMD management allowed for a good product?
The kind of problem AMD faces today, screwing up a R&D pipeline, should not be tracked to a single person or even to a single department. It's a more fundamental, structural problem with the company itself, and to make things worse, it starts and end with the BoD, the same BoD that is in charge of "fix" the company.
I think the reasons behind AMD cancelling the Opteron line is its inherent inefficiency, and even if AMD could match Intel core scaling they wouldn't be able to match the efficiency gains Intel has been delivering with Xeon. AMD had roughly parity in terms of modules with Intel cores with Piledriver/Sandy Bridge. Fast foward to 2014 and Intel is fielding 18C monsters, and unless AMD was able to match core increase with IBM node, their competitive position would be as ruinous as it is now, regardless of the money they would have to throw to develop the chip. AMD has a dog in terms of design in their hands, there's no node on earth that would make this chip shine.
And think about it, despite AMD shunning Bulldozer complexity they were not shy to try to develop an ARM server chip. There was disposition at AMD BoD to take risks and they took a big one when they forked their engineering teams by developing ARM and x86 cores, but they chose to not follow the 22nm SOI + Steamroller/Excavator route. Why would they do that if Steamroller/Excavator problem on servers and HEDT was "just" the manufacturing node? They could have a node if they wanted to, but instead decided to pursue an even riskier venture with their ARM product.
Your past assumption that Carrizo was a failure without even checking the numbers was what you could call speculation, and from the wildest barrel...
As for me it is not speculation but opinion after checking AMD s technical infos delivered at ISSCC 2015, as well as the reviews of the current competition counterparts.
Last edited:
Its still speculation until 3rd party reviews are presented.
I don't recall saying Carrizio was a failure. I said that I expect it to underdeliver like past products compared to the marketing literature.
It is fine to say that (you) expect something to be a failure. It is not fine to definitively state that is was a failure unless there is evidence.
ISSCC is not a marketing event, it s an engineers dedicated event, indeed i would have appreciated to see Intel publishing their core M frequency/power curve like AMD did recently for Carrizo..
If they went as far as publishing such technical material then it s a sign that they are 100% sure about the qualities of their design.
Here the slides in good definition, check page 9 about the curve, i give you a few hints that will allow you to extract the CPU caracteristics that interest us.
http://www.goldfries.com/downloadables/AMD_Carrizo.pdf
First the power is related to loading by Prime 95, with Cinebench you can increase the blue curve values by 10% up to the value of 8-10W/module.
Second the uncore use about 35-40% of the SoC power at low levels below 5W, and about 30-35% above 6-7W, with thoses numbers you can compute at wich frequencies it will be clocked in function of the total TDP.
It wouldnt be a surprise that Nitendo use this APU as basis given its expected efficency that should be better than Beema s.
Please do not argue with ABWX about performance for watt on any intel or amd cpu, it will go on forever and nothing will change both sides will have the same opinion as the beginning.
Lol, a big ad hominem...
Are thoses ISSCC numbers that annoying for you that you re advising people to ignore them.?.
Indeed your sayings are typical of one who know that he has no valuable argument technicaly speaking, hence the ad nauseam reliance on ad hominems, hey attack the messenger if you cant attack his message..
Last edited:
I'm not getting into this except to say wait and see.
Also for the record these are marketing slides and specifically state
ie) could be completely wrong
Also for the record these are marketing slides and specifically state
ie) could be completely wrong
This disclaimer is at the end of any corporate documents, i dont remember that you ressorted to such tengential arguments when some other brands published datas that were much less verifiable, it was all comments about this and that without waiting for reviews that indeed ended by negating the claims.
In this case though we have much more precise datas that were extracted from real silicon and wich are not hopes or simulations, as said ISSCC is an enginers event and datas published are up to this audience, it s not a pure marketing event a la Computex or CES.
Anyway keep on burying your head in the sand, i guess that even when reviews will be published we ll still read here and there that theses are not actual numbers if ever they dont suit some hopes...
When can we reasonably expect to see a Carrizo performance review put up on Anandtech? Just wondering about the timeline.
http://www.anandtech.com/show/8995/amd-at-isscc-2015-carrizo-and-excavator-details
That s three months or so at most, but actualy my point wasnt about a laptop dedicated Carrizo but about the circulating rumour that Nitendo could use the Excavator core in their next console, thing is that the published datas point to this core being more efficient than the one used in the XBX and PS4 while being cheap process wise, so this rumour is backed by technical informations and could well be right.
If you don't think that is marketing slide deck... LOL!
ISSCC has plenty of marketing masquerading around. That presentation is majority marketing, with very very little actual data presented, and even then the data presented is lacking actual values. Sorry but 10 years ago, that presentation wouldn't have made it through.
Where is the marketing since it s the kind of data that is verifiable by the exellence.?.
They published a frequency/power curve, they know that thoses datas will be checked once the product is launched and that thoses numbers can be cheked contrary to infos like parasitic capacitances reduced by 60% or transconductance improved that we saw on some slides, heck it require m$ gear to extract the transistors Spice (electrical) parameters from real silicon with 5% accuracy while checking power comsumption of a chip can be done by any amateur that has very basic knowledge and equipement.
Personaly i wouldnt take the risk to truncate a F/P curve for marketing purpose, moreover so close of the product launch, would you.?..
DrMrLordX
Lifer
- Apr 27, 2000
- 23,222
- 13,300
- 136
Despite all that, they did manage to produce a few good designs that others might want to license/poach (Mediatek, etc).
Well, here's the thing: AMD still sells Opterons, albeit in a small niche. They could have sold more with a better product (more accurately, their server market share would have undergone slower atrophy under the influence of a reduced perf/watt delta with Intel). They'd still be relegated to a minority of the market, but they probably could have maybe 5-10% of market share now instead of the maybe 2.5% they have in reality.
I've already stated my case in other threads that Steamroller actually has respectable fp performance. Not sure how much integer performance improved, since it's kind of hard to find integer workload benchmarks on the desktop (and Kaveri has specialized features for x264 so it isn't a good indicator).
Personally I think they were insane to take on that task. Seattle is still MIA. My guess is they thought they needed a microserver product that they could not deliver with Construction cores. Looking at what they achieved with 28nm planar + HDL + AVFS, I think they made the wrong call.
From what I can tell, their ARM products are (or were supposed to be) based on Cortex A57. It wasn't like they were designing something entirely new from scratch, or updating a complex and buggy mess like Bulldozer/Piledriver. Most ARM designs are supposed to be simple and cheap enough that they can be used in low-power applications where the cost-of-manufacture and the cost of deployment is low enough that it makes financial sense to sell the chips. The entire ARM ecosystem is cheap from top to bottom.
My guess is that it was actually cheaper to take a whack at an ARM design than it was to try and design 4M, 6M and MCM Excavator designs. AMD still has a lot of work to do fixing the many L3 cache issues from Bulldozer/Piledriver which is something they avoided entirely with Kaveri.
Besides, Excavator on 22nm SOI would have produced a "high performance" x86 server part facing Haswell-E/EP/EX in the server space, so, again, failure-of-confidence. Maybe they thought the ARMy was going to chip away market share from Intel with lots of slow cores and they wanted in. It might have appeared, at the time, to be a lower-risk venture than continuing Opteron development. They didn't expect stuff like Xeon-D.
Yeah I'd like to see one of those. If Carrizo is as efficient as AMD says it's going to be, why not leak some laptop models to the press? So far the only numbers I've seen have been a few 3DMark leaks.
The thing is, the first products were supposed to be A57 based, but the follow up to it is not. The follow up is K12, which not only is a custom ARM core, it also provides the IP blocks for its x86 counterparts. So AMD is indeed taking the risk of developing a new product and is tying the last cash flow positive part of their business (small core x86 chips) to it. If this isn't a risky strategy I don't know what it is, and yet AMD preferred to go down this path instead of a 22nm SOI Excavator.
This is a testament on how hard it would be hard to bring the CMT family up to competitive standards, if it could be fixed at all. The entire family really deserves the moniker of unmitigated failure.
Probably because Carrizo isn't as efficient as AMD says it is.
DrMrLordX
Lifer
- Apr 27, 2000
- 23,222
- 13,300
- 136
Now you're dealing with a whole 'nother ball o wax. K12 isn't due out until we-don't-know-when and its x86 counterpart, Zen, isn't due out until Q32016. Zen is supposed to be using Samsung's 14nm process, or some variant thereof. It wouldn't make much sense to use IBM's 22nm process by that point with a (hopefully) superior node being available by that time. Technically it's available now, never mind that nobody seems to be using it! Hmm Samsung, do tell!
22nm SOI is already losing relevance as Intel marches onwards to 10nm, so speaking of it within a 2016/2017 timeframe is not exactly what I had in mind. When keeping that time frame in mind, AMD basically has the choice of 14nm Excavator or 14nm Zen, and they picked 14nm Zen. Should be interesting to see how that works out for them.
Oh boo, now you're just being mean.
NTMBK
Lifer
- Nov 14, 2011
- 10,523
- 6,047
- 136
If AMD are making chips for Nintendo, I would expect something ARM based. Nintendo have stated that they want better compatibility between their handhelds and main consoles. My wild guess would be handheld with A53 cores and 2-4 GCN CUs, and a home console with A72/K12 cores and 12-18 GCN CUs.
I was talking about 2015 time frame. AMD could have launched a 2015 22nm Excavator around that time if they were willing to pay GLF for the IBM node, but instead they decided to not launch anything, dump the CMT family for good and go for its x86 ARM derivative.
That would put them in collision course with the rest of the industry, which have gathered around x86 as its CPU foi choice. Not that they didn't do this before, but this is a moment that all players are going to cost cuts measures.
NTMBK
Lifer
- Nov 14, 2011
- 10,523
- 6,047
- 136
Take a look at the Mark Cerny talk on design choices for the PS4- the reason ARM got rejected was lack of 64 bit. And until 2 years ago they were gathered around PPC. 
I don't think AMD will have an x86 core low power enough to scale down to a fanless, 3DS size device with a decent GPU. It just makes more sense to license an ARM core and focus on GPU tech.
Interestingly, the 3DS almost had a Tegra 2 chip but the deal fell apart: http://www.eurogamer.net/articles/digitalfoundry-3ds-techspec-exploration-blog-entry
I don't think AMD will have an x86 core low power enough to scale down to a fanless, 3DS size device with a decent GPU. It just makes more sense to license an ARM core and focus on GPU tech.
Interestingly, the 3DS almost had a Tegra 2 chip but the deal fell apart: http://www.eurogamer.net/articles/digitalfoundry-3ds-techspec-exploration-blog-entry
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter