Dinky Dino
Member
How much would these cpus even cost? I mean I always thought the i5 6600k wasn't THAT expensive at 230ish bucks
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
AMD's Zen CPUs to feature up to 32 cores and 8-channel DDR4
- Thread starter undu
- Start date
My point was that Scorpio can still have a faster CPU than xb1 and neo, while still using jaguar
K7 & K8's IPC advantage was mostly due to Intel making a misplaced bet on deep pipelines in NetBurst. Things reversed when Intel switched the the Core architecture, and have been that way for the last 10 years.
AMD has previewed the IPC improvement they are expecting with Zen, and unless they are sandbagging, it's not beyond Intel's current offerings.
Give me a MBP with an 8-core Zen CPU (suitably clocked down to fit within ~45W, of course) and I might finally have a reason to upgrade from my early 2011 model.
There is no "mobile Zen" versus "desktop Zen". AMD will use exactly the same core, and likely the same dies across both desktop and mobile, just like Intel does. The reason mobile CPUs consume so much less power is that (a) they're clocked lower, with lower voltages and are able to sustain maximum clock speeds for shorter periods of time and (b) they're binned dies, which means that only a small proportion of chips coming off the production line have to meet their power/thermal thresholds, with the rest going into desktop chips. In a console environment you'd get the benefit of lower clock speeds (although they may end up so low that you defeat the purpose of using high-performance cores), but you're at very much the opposite end of the spectrum when it come to binning.
Intel can run out a production line of dual-core Skylake dies where the best 5-10% become U-class Core i7 or Core m7 chips that sell for $400 a piece, and the rest sell for $100 a piece as desktop-class Core i3 chips. You can't design a semi-custom console chip on the assumption that every die coming off the production line is going to have that Core m7 class power efficiency. Instead, because margins are so slim and you don't have the opportunity to bin, you have to set your TDP thresholds such that 95+% of chips pass, which means you end up with desktop-class thermals.
Even aside from all of this, there's the question of the die size (i.e. cost) of using a octo-core Zen CPU in a console. While we don't have confirmed die sizes for Zen yet, if they're attempting to compete with Intel on performance then they can't be a whole lot smaller than Broadwell/Skylake cores, which would put an 8-core Zen CPU somewhere around the 200mm² range of 14nm. Attach a 6 Tflop GPU on there and you have an absolutely monstrous die, probably dwarfing even the GP102 used in the $1200 Pascal Titan X. There's just no way you can squeeze that into a console without taking massive, crippling losses on each unit sold.
Well I don't think any of the FX quad-cores were more than $140. A quad-core Zen CPU with the performance of a ~Skylake quadcore at that kind of price would be great.
zen isn't matching skylake performance at the same core counts and clockspeeds
How about Haswell then?
How do you know that?
anothertech
Member
So.. Scorpio/Neo 8 core Jag+ for sure at this point?
Link?
I probably know why you would say that...if you do the math with the 40% IPC improvement you would be roughly in the same ballpark as Skylake but not quite there yet. Dont state this as a fact though since it is not, we have to wait for some reliable data first.
chaosblade
Unconfirmed Member
I think AMD's target performance puts them close to Haswell in terms single threaded performance. Which isn't terrible by any means, especially if the price is right. And these could definitely end up being fairly cheap since they aren't going to have an IGP which takes up most of the die space these days. That greatly reduces cost.
In theory, an unlocked 4C/8T Zen with ~Haswell single threaded performance for $150 sounds pretty nice and would probably be pretty frequently recommended.
an educated guess and the way that amd has been talking about the core
normally i would say something like "i would expect zen to be a lower-performing core than skylake" but feel extremely confident that this will be the case
chaosblade
Unconfirmed Member
All indications are that Zen uses simultaneous multithreading (i.e., Hyperthreading), so yeah.
Rumors backed by AMD's history also point to them only having one consumer Zen chip used for all their configurations with cores disabled for the lower models, so feature-wise they should be the same.
AMD will definitely price Zen aggressively, but nowhere near that aggressively. You're basically talking about them undercutting their competition (i7 K chips) by almost 60%. Even if their performance does lag behind Kaby Lake by 10-15% there's no reason for them to go crazy on price. My guess is $200-$250 for the 4C/8T chip (i.e. Core i5 price point) if they're competitive with Haswell/Broadwell and can clock reasonably well.
THE:MILKMAN
Member
I didn't know Zen would be one chip fits all. Is that not different to how AMD have done things so far? Of course when I say desktop vs mobile I did mean in terms of TDP.
If you're estimate of 8-core Zen die size is anywhere near the mark then there is no way Scorpio will have even a low clock/volt Zen. Microsoft have stated Scorpio will have 8-cores and it would be massive overkill to use ~200mm2 die area for the CPU and it would be 16 threads too.
Given the info so far I think it highly likely.
I really shouldn't have used 'bus' because it can refer to a lot of things, but in reality 'wider' CPU architectures are always more advantageous per clock cycle. Long instructions are faster, data transfers are faster, so on and so on.
And to elaborate on a previous point, Keller's architecture designs directly led to x64 and Hypertransport, amongst other things.
The main disadvantage is that 'wide' architectures were traditionally more 'expensive' -- ie. silicon and transistors -- in the past, but that's a solvable problem in the 21st century.
That's a very good point. AMD's advantage during the K8 era was a confluence of several factors, two of which being AMD's innovation and Intel's architectural missteps.
I would have to agree that the chance of AMD surpassing Intel's offerings is extremely minimal. I would fully expect Zen to be slightly below or almost parity with Intel performance -- which is good enough for a shifting PC market.
edit:
That reminds me of a question I had on my mind the other day: Is there any confirmation that AMD will be bringing APUs to Zen? I suppose they can duct-tape a GPU and a Zen CPU together using some interposer magic but interposers between complex chips are super expensive.
Then again, with their roadmaps in mind, is AMD researching viable interposer technologies?
Yes, they did state that they will be making Zen APUs, they will be using Zen for all their CPU products, eventually.
I can't think of them using interposers in that manner in the near future. IIRC there was a video floating about speculating on the use of a multiple GPU die architecture but there is no real evidence to support that theory. However if such a thing were to be true, I could see them making interposers for that. Of course this is assuming also a lot of other factors, such as being able to increase IC production reticules such that massive areas of silicon are able to be patterned. I'm sure there a ton of other issues I haven't considered.
Haha, I'm sure that if AMD actually developed something like this, they'd be rolling around in money simply by licensing this technology off
They're not a chip or packaging company anymore though...But yeah, about multiple GPUs: I suppose it would work if they weren't going for direct chip-to-chip connections typical of interposer technology but instead have something that mimics a PCI-e or similar interface.
SenjutsuSage
Banned
Yep, it would seem like a major missed opportunity to release late 2017 and somehow not be packing a Zen based APU of some sort.
chaosblade
Unconfirmed Member
AMD's mindshare is non-existent in the CPU market, they have to shake things up. If their 4C parts are the same price as an i5 with better single threaded performance they are not going sell at the rate they would like. 4C needs to undercut i5s, 8C needs to undercut i7s. The bigger the disparity in single threaded performance, the more they need to undercut them by.
I agree $150 is definitely really low (it's why I said in theory) but I don't think they can go much over $200 with 4C parts and see success. Especially since Intel can take the hit and reduce their profit margins to go head to head with them.
Well, they used to have Bulldozer family cores for desktop and higher-end laptop and Bobcat family cores for cheap laptops/tablets, but they're dropping the low-end x86 cores in favour of ARM, so Zen is their only x86 core for late-2016 onwards (until the next-gen version of Zen).
In terms of physical chips, it looks like there's only a single 8 core die being manufactured for Summit Ridge, available either in the full 8C/16T configuration or a cut-down 4C/8T configuration. There's no indication that this die will be used in any mobile chips, but that's more down to the fact that there's zero interest for mobile CPUs without IGP. There are also Zen based APUs due next year code-named Raven Ridge, which will probably consist of a couple of dies which will be binned to cover everything from ultra-thin laptops to desktops and media PCs. The cores shouldn't be any different from the ones used in Summit Ridge, though.
AMD have confirmed that Zen based APUs will be releasing next year, although they haven't given any more info than that. I would highly suspect, though, that these will use fewer cores (likely 2 or 4) and incorporate much more modest GPUs than will be used in Scorpio. My guess would be that we'll see a 2C/4T die with relatively basic IGP for ultra-light laptops and cheap desktops, and then a 4C/8T die with RX 460 level IGP for more performance-focussed market segments.
It's funny you mention interposers, though, as there are two ways they could come into play with Zen-based chips. The first is the rumour that AMD's 32 core Zen is actually an MCM with four 8 core dies (i.e. the same die used in their consumer chips). This makes a lot of sense to me, as a 32 core single die solution would be an absolutely monstrously large die to try to make this early on on an immature node. (For reference it's taken Intel about 18 months to get >4C dies out on their 14nm process) I would imagine their experience with Fiji has also been helpful in learning how to overcome the hurdles of these kinds of monolithic MCMs, and the fact that they'll be charging thousands of dollars a piece should make the expense and manufacturing complexity worthwhile.
The second is the persistent speculation (I wouldn't even call it a rumour) that AMD will use HBM with their Zen-based APUs. This does make some degree of sense, in that the limited bandwidth of DDR3/4 is a big bottleneck for IGP, but I can't really see the economics working out. Even if you only use a single 4GB stack of HBM2, you still have to pay for the interposer and the package assembly, and I can't see that being financially viable for a price sensitive entry-level chip. In the longer run as costs some down it's sure to happen, but with even the new Titan X using GDDR5X it would definitely seem than HBM is still going to be very expensive for the foreseeable future.
MCM, that's something I haven't heard of in a very long time... I mean, it makes sense if they're willing to tie multiple cores together, but I would gather that the communication to and from each chip would be the main bottleneck.
Personally, I don't think that HBM will be used with APUs since APUs have historically used system memory. I mean, that's the whole point of an APU, right? To reduce overall system cost while maintaining (reasonable) performance?
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Crystal Well chips from intel, some of IBM's power, and nintendo's wiiU are all MCMs.
Sure. However if they were to increase the graphical capabilities of an APU they'd have to do something similar to crystalwell, which in AMD's case could be HBM or their own eDRAM solution.
I agree they need to be aggressive, but there's no need to go crazy. If they can provide ~30% better performance per dollar than Intel then that's enough for them to gain significant marketshare while still making reasonable profits. I also think you overestimate Intel's willingness to drop prices to compete. These are markets where they have literally 100% market share, and when you're in that position it isn't financially sensible to enter a price war with a competitor which has only just entered the market. Basically it makes more sense for them to let AMD take 20% of the market than it does for them to take a 40% hit to their margins to try to keep AMD out.
Yes, but system memory isn't designed to accommodate the bandwidth requirements of modern video games. Even AMD's cheapest dedicated GPU, the RX 460, has over 100GB/s of bandwidth available to it, while an APU on DDR4 at 2133MT/s has to make do with sharing just 34GB/s with the CPU. Even moving to 3200MT/s RAM gives it just 51GB/s of shared bandwidth.
Actual benchmarks bear this out, with AMD APUs getting near linear performance increases from faster memory. This shows that the limited bandwidth is the main bottleneck in an APU system, and there's a hard limit on how powerful an integrated GPU they can use if it's not going to be held back by the system memory.
This is why Intel use large eDRAM pools in their Iris Pro range of chips, as it effectively removes the bandwidth bottleneck for the IGP (another example of an MCM, by the way). If Zen performs reasonably in games, I can imagine there would be interest in a 2C/4T APU with ~4TF GPU and 4GB of HBM2, for compact gaming HTPCs, but I don't see AMD being able to do that for a price that would make any kind of sense.
THE:MILKMAN
Member
ARM... I completely forgot about AMD using them. Now it makes sense. I do wonder though given that the PS4 and X1 are using a CPU that takes up just 15% of the APU die just what CPU the real next-gen consoles will use? I mean even a 10nm 4-core Zen+ would likely be bigger than the ~50mm2 the Jaguar is wouldn't it?
I guess they might have to bite the bullet and budget more die area for the CPU in future or else design a unique CPU for console APUs?
The original fat PS3 pulled a couple hundred watts in games. And it was usually pretty quiet. Even with an internal PSU.
Wow, really? I should put those on my slides
Yes, that's very true. Current generation APUs perform much better with faster RAM. I remember a benchmark years ago that compared DDR3 vs DDR4 performance with and AMD APU...
I'm just looking up the Crystalwell stuff with Iris Pro and I find it quite interesting that eDRAM is on a separate die. Yields would certainly be better if the eDRAM isn't on the same die as the CPU, but integrating off-die eDRAM would certainly be an engineering marvel.
You typically don't mix DRAM production processes with standard logic IC production methods. They're fairly different, cell libraries will have different constraints. AFAIK there aren't any that integrate DRAM onto the same die as the CPU,what would be the point anyway, if you're using the same die you might as well use SRAM which is wholly compatible with CPU fabrication processes.
THE:MILKMAN
Member
Sure but I think a Scorpio with a full fat 8-core Zen and 6TF Vega (or even the mysterious RX490?) would actually be closer to 250W than 200 like PS3. Anyway the bigger issue is that the BOM of OG PS3 was estimated to be ~$900 at launch and that just isn't happening again even at $700. Not even with Microsoft's deep pockets.
full 8 core zen is almost 100 watts and the rx 480 is already just over 150, system consumption including a higher powered vega chip would be 300 watts minimum
it's not happening
IBM have been using embedded DRAM on-die for a long time, and the Wii U incorporates eDRAM on both CPU and GPU dies. Even aside from the density advantages over SRAM (roughly 3x), latency on large pools of eDRAM is actually typically lower than the same capacity of SRAM, due to the reduced propagation delay.
THE:MILKMAN
Member
I know. That is what I've been saying all along. Some posters insist that these new consoles will be dead in the water if they don't have desktop-class i5 levels of CPU performance, though.
Games are now becoming close to being entirely GPU dependent. Why would Sony or Microsoft waste massive die space and TDP on the CPU?
Hellgardia
Member
One possibility for Scorpio could be a underclocked 8-core zen (something similar to Intel's xeon e5-2609v4 but with even a bit lower power consumption) with a polaris gpu (even though rx480 sits a bit below 6tflops). But even then it would be over 200w...
Honestly though i think they will still use a beefed up jaguar/puma variant
Honestly though i think they will still use a beefed up jaguar/puma variant
Power 8 from IBM, the Last level on die cache is eDram I believe - obviously not mainstream consumer focused
if power consumption doesn't rule out zen, cost and die size absolutely would
if they're still on 14nm next year then they're looking at a much larger SoC than they currently have, if 10nm is in production it'll be a brand new process with all of the costs and complications that go with it
One comparison point is to look in the Intel Broadwell-DE Xeon D line and something like the Xeon D-1548 which has a TDP of 45W and turbo clock rate of only 2.6GHz. If Microsoft is going to use Zen, they might consider reducing or even removing the L3 cache, and on top of reducing the power consumption it has the added benefit of reducing the size of the chip and therefore price.
THE:MILKMAN
Member
I think the CPU will be whatever they can fit into give or take 50mm2.
The thing is, generally, if you can 32-parallelize a task, you can also n-parallelize it, and then you push it to the GPU.
There's still problems in pushing things to GPU rather than n-parallelize on the cpu, but the gains are generally around 1000x, soo...
Huh, you're right. I didn't know about the latency issue, I mistakenly thought that clockspeed advantages of SRAM and simply due to the way DRAM works, it would inherently suffer from more latency regardless of capacity (which also would be bounded by die sizes).
I'll also maintain that due to the different production methods typically employed by DRAM manufacturers having on die DRAM is somewhat of an exotic ask (and somewhat limiting for a fabless design perspective).
DonMigs85
Member
True, a well-coded console game will be less affected by reduced caches anyway. Ideally there'll be less mispredicted branches
Sure. But the best, hottest CPUs are not needed and something with lower TDP could certainly be worked up. Intel can get their TDP pretty darn low, by cutting clockspeeds, cores, cache, etc.
I know its fun to have a bunch of cores. But something like the 35w Skylake 6100T 3.2ghz (dual core with hyperthreading for 4 threads), would absolutely smash these Jaguar chips. I just looked, and they even have 15w mobile chips, like the 6600U. 3.4 ghz at 15w. dual core, hyperthreading.
Those are expensive, though. And Intel isn't trying to stuff their stuff, into a console, for a discount.
If Zen comes out and has ivy bridge like performance or better. A 30w chip would again, smash those octo-core Jaguars. And AMD could be willing to take the hit on price, to make it happen. and MS could be willing to pay, to make it happen.
I bet they could get a tri-core zen, down to a 15 - 30w tdp.
chaosblade
Unconfirmed Member
Intel's big focus is on power consumption and efficiency, and it has been that way since Sandy Bridge. I don't think it's wise to assume AMD is going to come even close to matching their performance/watt with Zen, when AMD's focus is primarily on getting relevant again in performance.
If you want low-TDP high-perf notebook CPU for your future consoles, you can look up Intel's Coffee Lake. That thing gives 6 full physical cores plus iGPU for just 35W.
http://www.guru3d.com/news-story/6-...oing-mainstream-in-2018-with-coffee-lake.html
http://www.guru3d.com/news-story/6-...oing-mainstream-in-2018-with-coffee-lake.html
I think that's too simplistic. I can have e.g. a branch-heavy tree or graph traversal which I can easily parallelize for 16 or 32 cores, but which will run terribly on a GPU. Or something like a branch-and-bound algorithm which relies on fast shared cache to update the bound condition.
Hellgardia
Member
Coffee lake is only due 2018 if i am not mistaken. Either way, a 6-core cpu with an integrated gpu (gt2 most likely) within a 35/45w envelope is a nice upgrade over the current laptop offerings.
We don't really know for sure what zen will bring to the table so we can only speculate. But i do hope it shakes things up a bit.
We don't really know for sure what zen will bring to the table so we can only speculate. But i do hope it shakes things up a bit.
eDRAM has definitively higher latency in low capacities than SRAM, which is why you typically only see it in use as an L3/L4 cache (or L2 in a less performance sensitive case like Wii U), but IBM's 22nm eDRAM L3 on the POWER8 achieves almost half the latency of Intel's 14nm SRAM L3, which shows the benefit when you get up to multi-megabyte capacities.
You're quite right that it's exotic from a production point of view, though. None of the four FinFET fabs currently offer eDRAM on their leading-edge processes, and although IBM will be shipping POWER9 with eDRAM next year on 14nm FinFET (in fabs now owned by GloFo but on a process designed by IBM), they've always prioritised eDRAM in their process development. It will also make post-14nm IBM design interesting, as by that point they'll be using standard GloFo fabrication processes, and if they want to continue using eDRAM they'll have to figure out how to integrate it into those processes (although I'm sure this is something they put serious thought into while negotiating the sale of their fabs).