I think the release schedule is too distant specially when we expect Desktops & Mobile parts should be released simultaneously or at least a month or two apart
See my response below. Not only that, but the desktop / enterprise is the high volume market. If anything a larger laptop SoC is more complex to make for a lower selling price, so of course they'll go after desktop and enterprise first.
I also would like to add not only do they make less per chip but Intel's 10nm is what they are competing against. I think it makes perfect sense to concede the laptop for now.
The 3000 mobile platform based on Zen+ is still not competitive against Intel - AnandTech's review of Surface laptops clearly shows it. And it isn't by a sizeable margin. I was curious to understand why the technology roadmap on mobile is basically 1 year late compared to desktop and servers. Is it a bandwidth issue? Or TSMC cannot provide more wafers anyway at this point?
Most likely that. TSMC has not capasity to produce new mobile soc, amd chiplets, amd GPUs etc... So some market areas Are gonna left behind, untill those that Are most important at this moment move to the next stage.
The Zen 3000 series is still on Zen+ and is using 12nm from GF, not 7nm from TSMC. Intel 14nm is way better than GF's 12nm, not counting the fact that most of the 3rd gen gains from changes like how threads are kept in same complex etc. are there, which contributed hugely to the increase in single/some multi-threaded performance.
Zen 4000 mobile/APU series should be Zen 2, we'll see quite a jump then, before that don't bother with AMD on mobile.
Don't bother with AMD on mobile? What a stupid thing to say. If the price is right, they've got some excellent mobile offerings. I'm very happy with my X360 with 2500U, which I got cheaper than the Intel version with a vastly weaker GPU. So much cheaper in fact, that I bought 200Wh of AC power banks and now I have more performance AND 2X the battery life.
All the questions in this thread can be answered with this: AMD is focused first and foremost on servers. Going for the big money. The rest is secondary. It's the right choice to prioritize on.
Five to six months late, not a year. Zen2 based desktop chips came out in July of 2019, Zen2 based laptop chips will be released in January(CES being the announcement that most expect). Zen2 based APUs have also not shown up yet. There is also a fair amount of work that needs to go into linking the CPU to the GPU in an APU. In a laptop chip that needs to be tiny, a chiplet design may not be a good choice, though it would make it easier.
If you think about it, 7nm was new for the release of Zen2 based desktop products. As such, not only were there capacity concerns for fabrication, but there are other things on the architectural side. For an APU, you need to link CPU to GPU within the chip, and that is going to be a lot more complicated than you might think. You can't miss that AMD has not released desktop APUs either, so it's not just a mobile chip issue.
Clock speeds, power draw, Vega vs. Navi for the graphics aspects and trying to get things working. It's really not a simple thing. Now, from what I have heard, AMD is not moving to a chiplet design for APUs, at least, not at this point. If AMD had the design for an APU that used an I/O die, then AMD would be in a better position to just throw a CPU die in there, a GPU die, and with the I/O die to tie them together, it's going to be simpler than directly linking the CPU to the GPU.
If AMD were to release a laptop CPU without graphics, that might allow for them to go out sooner, but the vast majority of laptops don't have discrete graphics.
That is a good question. I would guess that AMD looks at the market and what they can offer compared to the competition and determines on which fronts they can perform best.
The server market is a very important market as the margins are better there than with desktop and laptop and AMD has a multiple advantage over Intel as they can cram more cores and more PCIe IO on one chip. Also it takes more time to get you foot in the door in the server market but if your in you wont get ditched as fast as on the desktop/laptop. But I would say the server market is a very important market for AMD.
For the desktop and laptop markets I think 2 or 3 important things made them choose desktop over laptop. They desktop chips are closer to the server ones than the laptop ones, most PC enthousiasts use desktop, with more cores AMD could make a bigger difference on desktop workloads. And most importantly Intel has its most advanced core design only on laptop and Intel has been very good at power management. So AMD will have a much harder time competing with Intel on laptop.
Lastly for continued success market share is import so software makers will optimize for the AMD platform. This is very important and its one of the last reasons while on some workloads Intel still outperforms AMD as so many software has been optimized for Intel. And to change that AMD will need market share and time.
Mobile is difficult...and it is not as high revenue as server parts. They are a small company so they probably focused on getting the server parts ready first which then convert pretty easily to desktop products. Mobile parts, if you recall are SoCs, meaning everything on a single chip. That means more time to design, verify, test and lets not forget about the hardest part, making an efficient product, because in mobile efficiency is paramount.
It is possible Zen 2 mobile might have been delayed so long because of 7nm yields. TSMC experienced a chemical spill early on which corrupted a large batch of wafers... that could have contributed to the problem. Adding to that, companies such as Apple gobbled up A LOT of 7nm availability.
There's that of course.. and there's the possibility AMD didn't have Zen 2 mobile ready earlier (which would be a downright shame as I'd prefer the release schedule to be the same as their desktop counterparts).
Its possible we may see release of future mobile parts (such as Zen 3 and Zen 4) follow a schedule more in line with desktop releases.
This looks wrong. On die size basis or consume of silicon, a server cpu and a performant mobile cpu have the same ROI. It was AMD choice to lower too much the prices in server. A small and fast Mobile cpu is a 400$ thing today, pretty in line with the money advantage of a big 64 cores SKU that is in the 1000mm2 range of silicon size and it have an huge packaging cost.
Because the laptop market is not driven by DIY performance but rather it's dependent on power, perfromance and features. This makes it difficult to compete with intel on 12nm Zen+ or even 7nm if intel is on 10nm. Intel provides a fuller stack reference design making it easier for laptop manufacturers to implement. However with 7nm APU's with 7nm I/O and 7nm Vega AMD will be able to outperform Intel now all of a sudden this looks like an excellent option. The icing on the cake is that 7nm+ Zen3 APUs will come next year further pushing the envelop. We will see AMD laptops in huge numbers in 2020 and beyond imo. This is why Surface is vetting out the supply-chain with their AMD surface-edition laptop.
Why are the release schedule of Zen 2 desktop parts & Zen 2 Mobile parts Sooo distant from each other why doesn't AMD releases them together or at least at one month apart
Actually, mobile designs aren't that different vs their desktop counterparts. They do have an iGP which changes things around, but the overall CPU part is still the same. And also, if Zen 2 design was completed long before its release, then AMD would also have a mobile lineup designed at the same time. All it comes down to really is production at this point.
But its possible that supply issues with 7nm have caused delays... plus there was a chemical spill early on which contaminated large amount of wafer batches, and huge orders from other companies such as Apple.
Its possible things might change a bit with Zen 3 which uses 7nm+... and most notably after that Zen 4.
I'm glad you're telling me the mobile designs aren't that different. I wouldn't have known otherwise! I mean, I only cover AMD's product line in-depth. /s
One of the big issues is going to be around power on the chip, because managing power between CPU and GPU on the same die is a drastically more complicated issue than just CPU alone - there are various ways to do with with in-die LDOs, on-die regulators, on-package regulators, and therein all the sensor requirements to deal with current density hotspots generated from the design that have different ramp up functions from just a CPU-alone model. Then there's the fact you have to have masks created for the design, which in of itself is a big cost. But then we should also talk about any shared cache design, or the security aspects of the PSP dealing with different cores and different stages of memory access...
I could go on and on and on
It's not as simple as you make out. There is a reason that hundreds if not thousands of people work on every little detail, because an error in silicon isn't always solvable.
FWIW, Andrei of AT mentioned on Twitter that his Ryzen chip appeared to have about 17W power draw at idle: https://twitter.com/andreif7/status/11554204423599... -- this sort of lines up with things AT has previously mentioned about uncore/interconnect eating nontrivial power under even light load. (Ian kind of alludes to it in some questions about IF.)
Not a huge deal in a high-performance desktop, but a dealbreaker in mobile. And so another thing for the list of changes to put a core into a mobile chip.
1. The Mobile Variant is a Single Pieces of Silicon with GPU and Memory Controller, so it is entirely different to the Desktop and Server.
2. The Mobile Variant is also on a different node, even though they are the same 7nm, one is HP another is LP. Where LP stands for low power.
3. Lower Margin on most Laptop CPU, ( comparatively speaking ), AMD is basically making one Die, for Desktop and Server Market, both are higher margin market and just so happens AMD does them with lower cost and giving them another competitive advantage.
4. High Margin Laptop CPU is basically dominated by one vendor, Apple. AMD is not in the Apple ecosystem yet.
5. AMD is small and has limited resources. So I dont blame them for taking the path which leads to more profits first.
Mobile is different - not only CPU alone has fewer cores and smaller power envelope but you also need to include the cost of the whole platform. You need chipset with less power demand than X570, you need to include NIC, WiFi, etc. Intel had Centrino and now it has Athena - AMD has nothing like that. It means that for mobile AMD needs to transform their solution - maybe make it more SOC-like like in the server space. Servers with AMD CPUs do not have chipset -> everything is driven from IO Die. Server CPUs though have bigger prices and as everyone here agrees in mobile space margins are thin (and Intel really can dough money to OEMs to not to be too creative with AMD). I really would like to see AMD changing the mobile market the way they changed servers and desktops. It's not that easy though and it's better people realise that. BTW @Ian is it possible for you to ask whether the IPC uplift for the next gen will be the same for server and for desktop CPUs ? I see some obvious possibilities to increase the performance but they are more one the server side.
That is correct. However, in retrospect, it turned out to be the right move and in no uncertain terms is one reason why AMD has leapfrogged Intel yet again. Meanwhile, Intel is struggling with its fabs in a years'-long detrimental situation. Fabs are double-edged swords, not to doubt. They can be a boon but also an albatross 'round a company's neck. Also, "tic-toc" is what Intel did as essentially a monopoly supplier on higher-end x86. That sort of lazy, relaxed, utterly predictable production scheduling is something Intel will have to abandon if it expects to keep up with this new AMD. Great interview--glad to see that AMD intends to keep the R&D pedal to the metal from now on--as AMD seems to have learned its lesson about waiting on Intel to catch up with it...;)
I think it is a benefit to AMD to have a third party fab for multiple reasons. These new nodes are becoming extremely complex and take billions of R&D. Having a large fab that can spend those billions is a huge plus to stay cutting edge. Also the fab is able to spread those costs out across all their clients. If one day AMD finds that it is around the size of Intel it may make sense but until then I think they are better off with a third party.
Yup. TSMC gets to amortize their R&D costs across not just AMD's chips but some of NVIDIA's and most smartphone SoCs (from Apple, Qualcomm, MediaTek, Huawei...).
And if TSMC someday has their own version of Intel's 10nm disaster AMD could theoretically go see what's new from Samsung at the time. A switch could be super rough in therms of schedule and cost, but not "stuck on one process for five years" rough.
Plus, the cash AMD got out of the deal helped pay for all the new chip designs and so on. They even seemed to make a positive out of being locked into buying GloFo silicon for a while by using the older/(I'm guessing) cheaper node for I/O on desktop/server chips.
Can always second-guess but going fabless seems to have worked out well here.
AMD is not buying IO/12nm from GF .. just because they had prior agreement .... a part of the reason is also that IO performs better at higher node size and is a lot harder to shrink without loosing the performance/speed,etc .. so for forseeable future we might see IO die at a higher sized node than compute dies ( advantage because of chiplet arch/design )
AMD always talked about RDNA being scalable I think you forgot them to ask how scalable is it can it also work in sub 5 watts environments? & What about AMD's ambition in entering Smartphone market
'How scalable is it' is an open-ended question. It doesn't really matter how scalable it is when it depends on enabling it for the right product in the right market at the right time.
Regarding mobile, there's a specific question in here regarding mobile, and an answer to your question.
I loved your questions But you missed some questions on NVIDIA & Why is AMD lagging behind it soo much Has AMD settled for second spot in that market? is AMD always gonna fight on the basis of price ? Maybe in future you can do a twitter or an article where you take questions from readers
If you were AMD's CTO, how would you answer those quite blunt questions? Part of being an interviewer is developing a rapport to the point where you get unique answers out of your interviewee. If you go in like a rusty rake, the answers you get are only going to reflect that.
To your last point, I often ask my twitter followers for questions when I know about interview opportunities in advance. Sometimes these things can be organised ad-hoc at an event. But you have to follow my personal twitter for when the potential arises.
Don't know if you have been keeping up with the latest marketshare estimates, but nVidia has been leaking marketshare to AMD like a sieve, lately...;) But you have asked a loaded question, you realize, because it contains assumptions not everyone agrees with. For instance, some people think nVidia is far behind AMD on a number of fronts--is your assumption based on the fact that nVidia sells a ~$1200+ discrete GPU that less than 1% of the global gaming market has purchased? Or is it based on things like nVidia's faux "real-time ray tracing" features supported by D3d that possibly a half-dozen games support at present to any degree at all? Do you know the difference between real ray-tracing and rasterization? Etc. When making statements like yours it's generally a good idea to expound on the substance you think supports your position, imo.
I don't see how mark could answer nvidia related questions. No company wants to settle for second spot, they just try to launch products as fast as possible but because of budget, limited no of employees, maybe issues in the lab, they just can't compete with nvidia which is a much bigger company.
Intel seems to be betting on heterogeneous hardware. Even though AMD looks like to have the opportunity with regard to the GPUs (not the software side yet) but they don't look like to be working on Storage, Memory, IO, and etcetera. Don't they think this can be risky? In HPC at least, memory hierarchies are looking to be the norm in the future.
Don't forget that Intel makes more 3x profit in a quarter than AMD makes revenue in a whole year. Intel employs 10x more people and makes 60x (?) more money. AMD has to place bets 3-5 years in advance, and they're going to go after the big opportunities.
Ah yes I was going to say it one way then the other and failed on both. But the point still stands, the companies are vastly different in size and financial prowess.
However, as AMD has demonstrated more than once, it certainly takes more than financial prowess + employee count to stay out front in the x86 CPU marketplace--it remains to be seen if "all Intel's horses and all Intel's men can put Intel back together again," doesn't it?...;) Intel also likely has far and away more liability and operating costs than AMD, especially including its FABs at the moment. And last, of course, do we really want to see an AMD as big, clumsy, and ungainly as Intel--ever? IMO, that's the only reason they've been able to best Intel several times now--AMD is a lean, mean, engineering machine. Intel is rife with internal management turf wars and all of the bad, bad stuff that comes when companies get too large. All assumptions historically made about AMD's inability to compete and best Intel have been based on assumptions made as to relative sizes and profits, and all of them have been wrong to date...;)
AMD is focusing on it's core products of compute and graphics right now. They are MUCH smaller shop than Intel and simply don't have the bandwidth or capacity to work on a million things at once.
What I gathered from the interview is Zen3 for Desktop will arrive 4th quarter 2020 and will have an IPC 10.7% greater than Zen2. RDNA2 ? as nothing was said.
EMIB is a mini-interposer and a packaging technology. Not something AMD would be developing at this time and would work with a partner like TSMC for solutions like this.
One takeaway from interview with respect to the Moore’s law bit is that CPU frequency scaling has run into a wall around the 5 GHz barrier. Years in the past(90s early 2000s), new cpu generations brought higher clock speeds and thus more performance. But we’ve hit a wall, and so to squeeze more performance out in 2020+, the manufacturer has to consider other levers such as cache size/latency etc.
So a few questions. Why did we run in to this wall? Is it because intel/AMD wanted to maintain 95 Watt TDPs vs say a 300 W TDP? Also, as manufacturing processes shrink to 5 nm and below, do we expect the 5 ghz frequency barrier to remain, and if so why? Can manufacturers even use another material besides silicon to resume increases in clock speeds? I’m a noob just trying to learn...
Simplified: one reason is that thermal issues have become too large to easily overcome. At 7nm and smaller there is too much heat per unit of volume of material to be easily conducted away. Silicon, heatspreaders and coolers are at their limit there. Frequency is slightly sloping downwards even, what was 4.4GHz wil be 4.3GHz in the next generation etc. Before, on larger nodes, thermal issues did not present a wall so quickly with higher frequencies, now it is the case.
As SSTANIC kind of stated, the frequency is a marketing gimmick, it's more important what the performance of that chip is. E.g. AMD is on average ~10-15% more efficient per clock than Intel's 14nm, with e.g. Intel 7nm being more efficient than the 14nm but not being able to clock high enough to be more performant, so this is a game of total efficiency or efficiency per watt etc. all dependent on circumstances.
Intel focused on reducing power consumption over increasing speeds as higher clocks increased thermals while still having the same (or even smaller) chips which limited the max amount of heat, performance per watt is one of the most important aspects for servers and was an aspect showing a lot of opportunities for easy growth.
Sometime soonish we'll see desktop/laptop CPU and server CPU designs starting to split, I think it will probably be accelerated by ARM's entrance as server CPU being more specialized.
I see the comment ofteh that Intel "not able to clock high enough ...", but I also see on wikichip that Intel has available low power vs high performance cells at 10nm. I'm just going to assume that Intel will use their high performance cells for 10nm and, further, that AMD has already done so with the similar tsmc high performance cells. If that is the case, then AMD is also likely moving to low power cells to meet TDP constraints of laptop chips and will similarly clock at the lower base rates required to compete on battery life.
The one thing I take away from the article is basically AMD did not expect the demand were strong, and possibly did not expect Intel to be completely messed up. And has nothing to do with TSMC capacity problem.
The Enterprise answer might well be correct for HyperScaler like AWS, there are many other small cloud hosting services that simply could not get any volume from EPYC2.
I guess they were conservative, as they could not afford any lost at this point.
And I questioned whether the limitation is actually on GF's side. Since every chip will require an IOD.
Hopefully EPYC 3 they will be much more aggressive with capacity planning.
If there's something that I really wonder about AMD and their plans - is the Workstation market with regards to major OEM's (Dell, HPE, Lenovo, Supermicro etc...) - not a single one of them offers any WS machine based on Threadripper. Dell has something but with "gaming" in the machine's name, which makes it really tough to recommend to enterprise customers.
You said: " We now have a situation where your main competitor has increased its L1 cache size by 50%, in exchange for a 25% increase in latency right at the heart of the core, which was a big surprise to us."
Actually, Skylake cores had 4 cycle minimum latency in some select scenarios, but most of the time it was 5 cycles. Icelake just moved to 5 cycles for everything, so the increase isn't 25% across the board.
Seems pretty obvious AMD targeted the server markets from the get go for the superior ROI and the same design was leveraged down to desktop with the Ryzen chips, I think the next gen of mobile will have a much better answer to Intels solutions but frankly am all for the two to become closer in performance as its only good for the buyer ie me. I would be more interested in the GFX side and would love to see AMD stick it to Nvidia in the same way they did Intel not seen them really be on top since the 9800xt 256
By that metric, I'd say AMD were ahead with the 4870 and 5870 - the best competition Nvidia could offer at the time had absurd power requirements (and cost) just to do a little better on performance. Fermi in particular arrived late, broken and extremely hot.
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
68 Comments
Back to Article
Kishoreshack - Monday, December 30, 2019 - link
Why are mobile Zen 2 architecture processors coming soo late?Why are you neglecting the Laptop market?
Ian Cutress - Monday, December 30, 2019 - link
CES is next week. You can bet they'll say something about the 2020 roadmap then.Kishoreshack - Monday, December 30, 2019 - link
I think the release schedule is too distantspecially when we expect Desktops & Mobile parts should be released simultaneously or at least a month or two apart
Ian Cutress - Monday, December 30, 2019 - link
See my response below. Not only that, but the desktop / enterprise is the high volume market. If anything a larger laptop SoC is more complex to make for a lower selling price, so of course they'll go after desktop and enterprise first.FreckledTrout - Monday, December 30, 2019 - link
I also would like to add not only do they make less per chip but Intel's 10nm is what they are competing against. I think it makes perfect sense to concede the laptop for now.yankeeDDL - Monday, December 30, 2019 - link
The 3000 mobile platform based on Zen+ is still not competitive against Intel - AnandTech's review of Surface laptops clearly shows it. And it isn't by a sizeable margin. I was curious to understand why the technology roadmap on mobile is basically 1 year late compared to desktop and servers. Is it a bandwidth issue? Or TSMC cannot provide more wafers anyway at this point?haukionkannel - Monday, December 30, 2019 - link
Most likely that. TSMC has not capasity to produce new mobile soc, amd chiplets, amd GPUs etc... So some market areas Are gonna left behind, untill those that Are most important at this moment move to the next stage.RSAUser - Monday, December 30, 2019 - link
I think it's more likely margins weren't good enough yet then and OEM need to certify. You'll see Zen 2 with the 4000 series.RSAUser - Monday, December 30, 2019 - link
The Zen 3000 series is still on Zen+ and is using 12nm from GF, not 7nm from TSMC.Intel 14nm is way better than GF's 12nm, not counting the fact that most of the 3rd gen gains from changes like how threads are kept in same complex etc. are there, which contributed hugely to the increase in single/some multi-threaded performance.
Zen 4000 mobile/APU series should be Zen 2, we'll see quite a jump then, before that don't bother with AMD on mobile.
AshlayW - Tuesday, December 31, 2019 - link
Don't bother with AMD on mobile? What a stupid thing to say. If the price is right, they've got some excellent mobile offerings. I'm very happy with my X360 with 2500U, which I got cheaper than the Intel version with a vastly weaker GPU. So much cheaper in fact, that I bought 200Wh of AC power banks and now I have more performance AND 2X the battery life.Flying Aardvark - Tuesday, December 31, 2019 - link
All the questions in this thread can be answered with this: AMD is focused first and foremost on servers. Going for the big money. The rest is secondary. It's the right choice to prioritize on.Targon - Tuesday, December 31, 2019 - link
Five to six months late, not a year. Zen2 based desktop chips came out in July of 2019, Zen2 based laptop chips will be released in January(CES being the announcement that most expect). Zen2 based APUs have also not shown up yet. There is also a fair amount of work that needs to go into linking the CPU to the GPU in an APU. In a laptop chip that needs to be tiny, a chiplet design may not be a good choice, though it would make it easier.Gondalf - Wednesday, January 1, 2020 - link
Not released but announced in January. Likely we will see something in summer.Targon - Tuesday, December 31, 2019 - link
If you think about it, 7nm was new for the release of Zen2 based desktop products. As such, not only were there capacity concerns for fabrication, but there are other things on the architectural side. For an APU, you need to link CPU to GPU within the chip, and that is going to be a lot more complicated than you might think. You can't miss that AMD has not released desktop APUs either, so it's not just a mobile chip issue.Clock speeds, power draw, Vega vs. Navi for the graphics aspects and trying to get things working. It's really not a simple thing. Now, from what I have heard, AMD is not moving to a chiplet design for APUs, at least, not at this point. If AMD had the design for an APU that used an I/O die, then AMD would be in a better position to just throw a CPU die in there, a GPU die, and with the I/O die to tie them together, it's going to be simpler than directly linking the CPU to the GPU.
If AMD were to release a laptop CPU without graphics, that might allow for them to go out sooner, but the vast majority of laptops don't have discrete graphics.
AeroWB - Monday, December 30, 2019 - link
That is a good question. I would guess that AMD looks at the market and what they can offer compared to the competition and determines on which fronts they can perform best.The server market is a very important market as the margins are better there than with desktop and laptop and AMD has a multiple advantage over Intel as they can cram more cores and more PCIe IO on one chip. Also it takes more time to get you foot in the door in the server market but if your in you wont get ditched as fast as on the desktop/laptop. But I would say the server market is a very important market for AMD.
For the desktop and laptop markets I think 2 or 3 important things made them choose desktop over laptop. They desktop chips are closer to the server ones than the laptop ones, most PC enthousiasts use desktop, with more cores AMD could make a bigger difference on desktop workloads. And most importantly Intel has its most advanced core design only on laptop and Intel has been very good at power management. So AMD will have a much harder time competing with Intel on laptop.
Lastly for continued success market share is import so software makers will optimize for the AMD platform. This is very important and its one of the last reasons while on some workloads Intel still outperforms AMD as so many software has been optimized for Intel. And to change that AMD will need market share and time.
yeeeeman - Monday, December 30, 2019 - link
Mobile is difficult...and it is not as high revenue as server parts. They are a small company so they probably focused on getting the server parts ready first which then convert pretty easily to desktop products. Mobile parts, if you recall are SoCs, meaning everything on a single chip. That means more time to design, verify, test and lets not forget about the hardest part, making an efficient product, because in mobile efficiency is paramount.deksman2 - Monday, December 30, 2019 - link
It is possible Zen 2 mobile might have been delayed so long because of 7nm yields.TSMC experienced a chemical spill early on which corrupted a large batch of wafers... that could have contributed to the problem.
Adding to that, companies such as Apple gobbled up A LOT of 7nm availability.
There's that of course.. and there's the possibility AMD didn't have Zen 2 mobile ready earlier (which would be a downright shame as I'd prefer the release schedule to be the same as their desktop counterparts).
Its possible we may see release of future mobile parts (such as Zen 3 and Zen 4) follow a schedule more in line with desktop releases.
Freeb!rd - Monday, December 30, 2019 - link
Because chiplets for servers are a much better ROI than an APU will ever be... and sub-par server chiplets work just fine as Desktop parts.Gondalf - Wednesday, January 1, 2020 - link
This looks wrong. On die size basis or consume of silicon, a server cpu and a performant mobile cpu have the same ROI. It was AMD choice to lower too much the prices in server. A small and fast Mobile cpu is a 400$ thing today, pretty in line with the money advantage of a big 64 cores SKU that is in the 1000mm2 range of silicon size and it have an huge packaging cost.zangheiv - Monday, December 30, 2019 - link
Because the laptop market is not driven by DIY performance but rather it's dependent on power, perfromance and features. This makes it difficult to compete with intel on 12nm Zen+ or even 7nm if intel is on 10nm. Intel provides a fuller stack reference design making it easier for laptop manufacturers to implement. However with 7nm APU's with 7nm I/O and 7nm Vega AMD will be able to outperform Intel now all of a sudden this looks like an excellent option. The icing on the cake is that 7nm+ Zen3 APUs will come next year further pushing the envelop. We will see AMD laptops in huge numbers in 2020 and beyond imo. This is why Surface is vetting out the supply-chain with their AMD surface-edition laptop.Kishoreshack - Monday, December 30, 2019 - link
Why are the release schedule of Zen 2 desktop parts& Zen 2 Mobile parts Sooo distant from each other
why doesn't AMD releases them together
or at least at one month apart
Ian Cutress - Monday, December 30, 2019 - link
They are two completely different designs on silicon. You can't just magic them up overnight.deksman2 - Monday, December 30, 2019 - link
Actually, mobile designs aren't that different vs their desktop counterparts.They do have an iGP which changes things around, but the overall CPU part is still the same.
And also, if Zen 2 design was completed long before its release, then AMD would also have a mobile lineup designed at the same time.
All it comes down to really is production at this point.
But its possible that supply issues with 7nm have caused delays... plus there was a chemical spill early on which contaminated large amount of wafer batches, and huge orders from other companies such as Apple.
Its possible things might change a bit with Zen 3 which uses 7nm+... and most notably after that Zen 4.
Ian Cutress - Monday, December 30, 2019 - link
I'm glad you're telling me the mobile designs aren't that different. I wouldn't have known otherwise! I mean, I only cover AMD's product line in-depth. /sOne of the big issues is going to be around power on the chip, because managing power between CPU and GPU on the same die is a drastically more complicated issue than just CPU alone - there are various ways to do with with in-die LDOs, on-die regulators, on-package regulators, and therein all the sensor requirements to deal with current density hotspots generated from the design that have different ramp up functions from just a CPU-alone model. Then there's the fact you have to have masks created for the design, which in of itself is a big cost. But then we should also talk about any shared cache design, or the security aspects of the PSP dealing with different cores and different stages of memory access...
I could go on
and on
and on
It's not as simple as you make out. There is a reason that hundreds if not thousands of people work on every little detail, because an error in silicon isn't always solvable.
twotwotwo - Monday, December 30, 2019 - link
FWIW, Andrei of AT mentioned on Twitter that his Ryzen chip appeared to have about 17W power draw at idle: https://twitter.com/andreif7/status/11554204423599... -- this sort of lines up with things AT has previously mentioned about uncore/interconnect eating nontrivial power under even light load. (Ian kind of alludes to it in some questions about IF.)Not a huge deal in a high-performance desktop, but a dealbreaker in mobile. And so another thing for the list of changes to put a core into a mobile chip.
ksec - Monday, December 30, 2019 - link
1. The Mobile Variant is a Single Pieces of Silicon with GPU and Memory Controller, so it is entirely different to the Desktop and Server.2. The Mobile Variant is also on a different node, even though they are the same 7nm, one is HP another is LP. Where LP stands for low power.
3. Lower Margin on most Laptop CPU, ( comparatively speaking ), AMD is basically making one Die, for Desktop and Server Market, both are higher margin market and just so happens AMD does them with lower cost and giving them another competitive advantage.
4. High Margin Laptop CPU is basically dominated by one vendor, Apple. AMD is not in the Apple ecosystem yet.
5. AMD is small and has limited resources. So I dont blame them for taking the path which leads to more profits first.
Zibi - Monday, December 30, 2019 - link
Mobile is different - not only CPU alone has fewer cores and smaller power envelope but you also need to include the cost of the whole platform. You need chipset with less power demand than X570, you need to include NIC, WiFi, etc. Intel had Centrino and now it has Athena - AMD has nothing like that. It means that for mobile AMD needs to transform their solution - maybe make it more SOC-like like in the server space. Servers with AMD CPUs do not have chipset -> everything is driven from IO Die. Server CPUs though have bigger prices and as everyone here agrees in mobile space margins are thin (and Intel really can dough money to OEMs to not to be too creative with AMD). I really would like to see AMD changing the mobile market the way they changed servers and desktops. It's not that easy though and it's better people realise that.BTW @Ian is it possible for you to ask whether the IPC uplift for the next gen will be the same for server and for desktop CPUs ? I see some obvious possibilities to increase the performance but they are more one the server side.
AshlayW - Tuesday, December 31, 2019 - link
We've got an armchair Silicon engineer here.Kishoreshack - Monday, December 30, 2019 - link
Shouldn't AMD also work on their own FAB ?Too much dependency on TSMC is causing it trouble?
Ian Cutress - Monday, December 30, 2019 - link
AMD used to have its own fabs. Due to financial issues they were spun out and are now called Global Foundries.WaltC - Tuesday, December 31, 2019 - link
That is correct. However, in retrospect, it turned out to be the right move and in no uncertain terms is one reason why AMD has leapfrogged Intel yet again. Meanwhile, Intel is struggling with its fabs in a years'-long detrimental situation. Fabs are double-edged swords, not to doubt. They can be a boon but also an albatross 'round a company's neck. Also, "tic-toc" is what Intel did as essentially a monopoly supplier on higher-end x86. That sort of lazy, relaxed, utterly predictable production scheduling is something Intel will have to abandon if it expects to keep up with this new AMD. Great interview--glad to see that AMD intends to keep the R&D pedal to the metal from now on--as AMD seems to have learned its lesson about waiting on Intel to catch up with it...;)FreckledTrout - Monday, December 30, 2019 - link
I think it is a benefit to AMD to have a third party fab for multiple reasons. These new nodes are becoming extremely complex and take billions of R&D. Having a large fab that can spend those billions is a huge plus to stay cutting edge. Also the fab is able to spread those costs out across all their clients. If one day AMD finds that it is around the size of Intel it may make sense but until then I think they are better off with a third party.twotwotwo - Monday, December 30, 2019 - link
Yup. TSMC gets to amortize their R&D costs across not just AMD's chips but some of NVIDIA's and most smartphone SoCs (from Apple, Qualcomm, MediaTek, Huawei...).And if TSMC someday has their own version of Intel's 10nm disaster AMD could theoretically go see what's new from Samsung at the time. A switch could be super rough in therms of schedule and cost, but not "stuck on one process for five years" rough.
Plus, the cash AMD got out of the deal helped pay for all the new chip designs and so on. They even seemed to make a positive out of being locked into buying GloFo silicon for a while by using the older/(I'm guessing) cheaper node for I/O on desktop/server chips.
Can always second-guess but going fabless seems to have worked out well here.
wolfesteinabhi - Tuesday, December 31, 2019 - link
AMD is not buying IO/12nm from GF .. just because they had prior agreement .... a part of the reason is also that IO performs better at higher node size and is a lot harder to shrink without loosing the performance/speed,etc .. so for forseeable future we might see IO die at a higher sized node than compute dies ( advantage because of chiplet arch/design )Kishoreshack - Monday, December 30, 2019 - link
AMD always talked about RDNA being scalableI think you forgot them to ask how scalable is it
can it also work in sub 5 watts environments?
&
What about AMD's ambition in entering Smartphone market
Ian Cutress - Monday, December 30, 2019 - link
'How scalable is it' is an open-ended question. It doesn't really matter how scalable it is when it depends on enabling it for the right product in the right market at the right time.Regarding mobile, there's a specific question in here regarding mobile, and an answer to your question.
Kishoreshack - Monday, December 30, 2019 - link
I loved your questionsBut you missed some questions on NVIDIA &
Why is AMD lagging behind it soo much
Has AMD settled for second spot in that market?
is AMD always gonna fight on the basis of price ?
Maybe in future you can do a twitter or an article where you take questions from readers
ANANDTECH readers insight is always valuable
Ian Cutress - Monday, December 30, 2019 - link
If you were AMD's CTO, how would you answer those quite blunt questions? Part of being an interviewer is developing a rapport to the point where you get unique answers out of your interviewee. If you go in like a rusty rake, the answers you get are only going to reflect that.To your last point, I often ask my twitter followers for questions when I know about interview opportunities in advance. Sometimes these things can be organised ad-hoc at an event. But you have to follow my personal twitter for when the potential arises.
Kishoreshack - Monday, December 30, 2019 - link
I think you didn't understand my questionthe question was what are they doing to fight Nvidia
what is their roadmap & how far have they reached
lmcd - Monday, December 30, 2019 - link
You literally cannot ask that question, unless you aim to impersonate a bull in a china shop.WaltC - Wednesday, January 1, 2020 - link
Don't know if you have been keeping up with the latest marketshare estimates, but nVidia has been leaking marketshare to AMD like a sieve, lately...;) But you have asked a loaded question, you realize, because it contains assumptions not everyone agrees with. For instance, some people think nVidia is far behind AMD on a number of fronts--is your assumption based on the fact that nVidia sells a ~$1200+ discrete GPU that less than 1% of the global gaming market has purchased? Or is it based on things like nVidia's faux "real-time ray tracing" features supported by D3d that possibly a half-dozen games support at present to any degree at all? Do you know the difference between real ray-tracing and rasterization? Etc. When making statements like yours it's generally a good idea to expound on the substance you think supports your position, imo.yeeeeman - Monday, December 30, 2019 - link
I don't see how mark could answer nvidia related questions. No company wants to settle for second spot, they just try to launch products as fast as possible but because of budget, limited no of employees, maybe issues in the lab, they just can't compete with nvidia which is a much bigger company.mmrezaie - Monday, December 30, 2019 - link
Intel seems to be betting on heterogeneous hardware. Even though AMD looks like to have the opportunity with regard to the GPUs (not the software side yet) but they don't look like to be working on Storage, Memory, IO, and etcetera. Don't they think this can be risky? In HPC at least, memory hierarchies are looking to be the norm in the future.Ian Cutress - Monday, December 30, 2019 - link
Don't forget that Intel makes more 3x profit in a quarter than AMD makes revenue in a whole year. Intel employs 10x more people and makes 60x (?) more money. AMD has to place bets 3-5 years in advance, and they're going to go after the big opportunities.ilt24 - Monday, December 30, 2019 - link
@Ian ... "Intel makes more 3x profit in a quarter than AMD makes revenue in a whole year"Did you mean per year for both ??? Just asking because last year AMD generated $6.5B in revenue while Intel generated $19.3B in profits.
Ian Cutress - Monday, December 30, 2019 - link
Ah yes I was going to say it one way then the other and failed on both. But the point still stands, the companies are vastly different in size and financial prowess.WaltC - Wednesday, January 1, 2020 - link
However, as AMD has demonstrated more than once, it certainly takes more than financial prowess + employee count to stay out front in the x86 CPU marketplace--it remains to be seen if "all Intel's horses and all Intel's men can put Intel back together again," doesn't it?...;) Intel also likely has far and away more liability and operating costs than AMD, especially including its FABs at the moment. And last, of course, do we really want to see an AMD as big, clumsy, and ungainly as Intel--ever? IMO, that's the only reason they've been able to best Intel several times now--AMD is a lean, mean, engineering machine. Intel is rife with internal management turf wars and all of the bad, bad stuff that comes when companies get too large. All assumptions historically made about AMD's inability to compete and best Intel have been based on assumptions made as to relative sizes and profits, and all of them have been wrong to date...;)extide - Monday, December 30, 2019 - link
AMD is focusing on it's core products of compute and graphics right now. They are MUCH smaller shop than Intel and simply don't have the bandwidth or capacity to work on a million things at once.HardwareDufus - Monday, December 30, 2019 - link
What I gathered from the interview is Zen3 for Desktop will arrive 4th quarter 2020 and will have an IPC 10.7% greater than Zen2.RDNA2 ? as nothing was said.
HStewart - Monday, December 30, 2019 - link
It looks like AMD's Semicustom design is AND;s answer for EMiB like in in last years Dell XPS 15 2im1 - not sure if they have answer to Fovoresextide - Monday, December 30, 2019 - link
EMIB is a mini-interposer and a packaging technology. Not something AMD would be developing at this time and would work with a partner like TSMC for solutions like this.Exotica - Monday, December 30, 2019 - link
One takeaway from interview with respect to the Moore’s law bit is that CPU frequency scaling has run into a wall around the 5 GHz barrier. Years in the past(90s early 2000s), new cpu generations brought higher clock speeds and thus more performance. But we’ve hit a wall, and so to squeeze more performance out in 2020+, the manufacturer has to consider other levers such as cache size/latency etc.So a few questions. Why did we run in to this wall? Is it because intel/AMD wanted to maintain 95 Watt TDPs vs say a 300 W TDP? Also, as manufacturing processes shrink to 5 nm and below, do we expect the 5 ghz frequency barrier to remain, and if so why? Can manufacturers even use another material besides silicon to resume increases in clock speeds? I’m a noob just trying to learn...
SSTANIC - Monday, December 30, 2019 - link
Simplified: one reason is that thermal issues have become too large to easily overcome. At 7nm and smaller there is too much heat per unit of volume of material to be easily conducted away. Silicon, heatspreaders and coolers are at their limit there. Frequency is slightly sloping downwards even, what was 4.4GHz wil be 4.3GHz in the next generation etc. Before, on larger nodes, thermal issues did not present a wall so quickly with higher frequencies, now it is the case.RSAUser - Monday, December 30, 2019 - link
As SSTANIC kind of stated, the frequency is a marketing gimmick, it's more important what the performance of that chip is. E.g. AMD is on average ~10-15% more efficient per clock than Intel's 14nm, with e.g. Intel 7nm being more efficient than the 14nm but not being able to clock high enough to be more performant, so this is a game of total efficiency or efficiency per watt etc. all dependent on circumstances.Intel focused on reducing power consumption over increasing speeds as higher clocks increased thermals while still having the same (or even smaller) chips which limited the max amount of heat, performance per watt is one of the most important aspects for servers and was an aspect showing a lot of opportunities for easy growth.
Sometime soonish we'll see desktop/laptop CPU and server CPU designs starting to split, I think it will probably be accelerated by ARM's entrance as server CPU being more specialized.
JayNor - Wednesday, January 1, 2020 - link
I see the comment ofteh that Intel "not able to clock high enough ...", but I also see on wikichip that Intel has available low power vs high performance cells at 10nm. I'm just going to assume that Intel will use their high performance cells for 10nm and, further, that AMD has already done so with the similar tsmc high performance cells. If that is the case, then AMD is also likely moving to low power cells to meet TDP constraints of laptop chips and will similarly clock at the lower base rates required to compete on battery life.ksec - Monday, December 30, 2019 - link
The one thing I take away from the article is basically AMD did not expect the demand were strong, and possibly did not expect Intel to be completely messed up. And has nothing to do with TSMC capacity problem.The Enterprise answer might well be correct for HyperScaler like AWS, there are many other small cloud hosting services that simply could not get any volume from EPYC2.
I guess they were conservative, as they could not afford any lost at this point.
And I questioned whether the limitation is actually on GF's side. Since every chip will require an IOD.
Hopefully EPYC 3 they will be much more aggressive with capacity planning.
RSAUser - Monday, December 30, 2019 - link
AMD got burnt on the graphics side with demand, I think it will take a long while for them to ever oversupply again.hetzbh - Monday, December 30, 2019 - link
If there's something that I really wonder about AMD and their plans - is the Workstation market with regards to major OEM's (Dell, HPE, Lenovo, Supermicro etc...) - not a single one of them offers any WS machine based on Threadripper. Dell has something but with "gaming" in the machine's name, which makes it really tough to recommend to enterprise customers.Kishoreshack - Monday, December 30, 2019 - link
I think on the launch they had OEM'S like Dell HP & Amazon were very eager to use their server partsThey showed the same enthu in threadripper parts
I think you have been misinformed
IntelUser2000 - Monday, December 30, 2019 - link
You said:" We now have a situation where your main competitor has increased its L1 cache size by 50%, in exchange for a 25% increase in latency right at the heart of the core, which was a big surprise to us."
Actually, Skylake cores had 4 cycle minimum latency in some select scenarios, but most of the time it was 5 cycles. Icelake just moved to 5 cycles for everything, so the increase isn't 25% across the board.
Adonisds - Monday, December 30, 2019 - link
That's excellent Ianalufan - Tuesday, December 31, 2019 - link
Seems pretty obvious AMD targeted the server markets from the get go for the superior ROI and the same design was leveraged down to desktop with the Ryzen chips, I think the next gen of mobile will have a much better answer to Intels solutions but frankly am all for the two to become closer in performance as its only good for the buyer ie me.I would be more interested in the GFX side and would love to see AMD stick it to Nvidia in the same way they did Intel not seen them really be on top since the 9800xt 256
qwertymac93 - Tuesday, December 31, 2019 - link
They got very close with the 290x.alufan - Thursday, January 2, 2020 - link
yes but that thing needed a power supply all of its own !Spunjji - Monday, January 6, 2020 - link
By that metric, I'd say AMD were ahead with the 4870 and 5870 - the best competition Nvidia could offer at the time had absurd power requirements (and cost) just to do a little better on performance. Fermi in particular arrived late, broken and extremely hot.Fataliity - Tuesday, December 31, 2019 - link
@Ian Cuttresshttps://news.synopsys.com/2019-09-11-Synopsys-Deli...
CXL specs on designware IP (most likely AMD's solution, they use Designware)
https://news.synopsys.com/2019-10-29-Synopsys-Acce...
New Infinity Fabric. Both are already finished.
Fataliity - Tuesday, December 31, 2019 - link
maybe you can make more out of the numbers, I dont know enough about the IF and PCI-E standards to really judge what the changes mean.JayNor - Wednesday, January 1, 2020 - link
No mention of 3D chiplet stacking. Is there a rift between Norrod and Papermaster on this?