Thursday, August 24th 2017
AMD Defends MCM Approach for Building 32-core EPYC: Huge Cost Savings of ~41%
AMD, presenting at a HotChips talk, packed with VLSI gurus, defended its decision to make AMD EPYC enterprise processors in the socket SP3r2 package a multi-chip module (MCM) of four 8-core "Summit Ridge" dies, rather than building a monolithic 32-core die. For starters, it stressed on the benefits of a single kind of silicon, which it can use across its performance-desktop, high-end desktop, and enterprise product-stacks; which translates into higher yields. "Summit Ridge" is the only CPU silicon based on the "Zen" micro-architecture, with the company giving finishing touches to the second silicon, codenamed "Raven Ridge," which will power mobile and desktop APUs. AMD has to pick the best-performing dies out of a common bin. The top 5% dies go into powering the company's Ryzen Threadripper HEDT processors, and a higher percentile go into making EPYC.
The relatively smaller 8-core common die has an inherently higher yield than a larger chip due to the rule of inverse-exponential reduction in yield with increasing die-size. This, coupled with the R&D costs that would have gone into developing the hypothetical monolithic 32-core "Zen" based silicon, works out to a significant cost saving for the company. A 4-die/32-core MCM is 0.59X the cost of a hypothetical monolithic 32-core die, according to the company, which is a cost-saving that enables the company to aggressively price its products. The slide AMD used in its presentation also confirms that each 8-core "Summit Ridge" die features four external Infinity Fabric links, besides the one that connects the two CCX units with each other. On a 4-die EPYC MCM, three out of four of those external IF links wire out to the neighboring dies, and one link per die probably wires out to a neighboring socket on 2P machines.
The relatively smaller 8-core common die has an inherently higher yield than a larger chip due to the rule of inverse-exponential reduction in yield with increasing die-size. This, coupled with the R&D costs that would have gone into developing the hypothetical monolithic 32-core "Zen" based silicon, works out to a significant cost saving for the company. A 4-die/32-core MCM is 0.59X the cost of a hypothetical monolithic 32-core die, according to the company, which is a cost-saving that enables the company to aggressively price its products. The slide AMD used in its presentation also confirms that each 8-core "Summit Ridge" die features four external Infinity Fabric links, besides the one that connects the two CCX units with each other. On a 4-die EPYC MCM, three out of four of those external IF links wire out to the neighboring dies, and one link per die probably wires out to a neighboring socket on 2P machines.
46 Comments on AMD Defends MCM Approach for Building 32-core EPYC: Huge Cost Savings of ~41%
The approach does have some downsides, but it seems to be a fair trade off if the cost savings are that high.
AMD states that the price is 0.59x of what the monolithic die would be.
So the cost savings are actually 41%
Thank you for your articles and everything. Have a nice Day everyone.
B costs 59%
What's the cost saving, TPU?
TPU: 59%, dúh! :roll:
Good job TPU! Good job!
No offense, but the quality in your news articles took a big nose dive in the last few months.
If only everyone was like that...
Edit: And also, welcome to TPU! We guess you didn't just start an account to correct our math, and hope you become a member of our awesome community =) Hit our forums for all the guides and stickies you might need.
Coffee Lake and 6-core mainstream was already announced some time ago. The only kneejerk from Intel was them pulling it in 3Q of a year, but they were definitely already planning and designing a similar move to higher core counts. I find it funny that people always consider the market leader less intelligent than themselves. Besides, Intel has known ever since Sandy Bridge that the Ghz wall was 4 Ghz and going over was going to be painful for perf/watt, proof being 6 generations of Intel Core products after that, and since Kaby Lake they're passing the 4Ghz barrier base/boost clock and look at those glorious TDPs, even on Intel's superior process. Its not a coincidence Coffee Lake is the gen that releases higher core counts. If Intel was really panicking they'd not have postponed so much over the years, from full node shrinks to whole generations of refreshes as with Kaby Lake. Everyone has known Ryzen was in the pipeline for several years now and AMD already pushed higher core counts during Sandy Bridge.
Multi GPU idea came from a lack of processing power, its as simple as that. Not 'users'. Just demand, like in every marketplace. Right now we have single cards that can push enough pixels to saturate the cable the data's sent through, hell we can even add endless layers of post processing on top and STILL render 'in-time'. GPU processing power on a single die has leaped ahead of I/O standards. It makes a lot of sense that we're now moving away from multi GPU again. We've seen its drawbacks and all sorts of new tech really can't deal with those drawbacks (latency, frametime variance, high degrees of support from drivers). At the same time, just like @ Intel Core and AMD Ryzen, modular design now steps in because we now have access to fast interconnects and are rapidly exhausting the available 'tricks' to further increase perf/watt and perf/mm of silicon. On top of that, node shrinks increase the yield risk of larger dies even more; so really, modular is inevitable and you don't need to be rocket scientist to understand this.
AMD isn't handing out stuff, its just following the obvious line of thought, one that all leading hardware companies follow, one that's already widely adopted in other markets too, such as ARM. Its a well known fact that in all ages of our history, a lot of people at similar timeframes think of similar things. And they really have never spoken to each other. It just made sense at that time. What matters is the first fellow who can market that idea as a usable, effective product.
It's just PC enthusiasts that know how awesome this technology is and why everyone should support it.
The truth is: investors understand Zen pretty well. So well, in fact, that they've priced AMD very well at the day of release already.
As you can see below - there were no large shifts in stock price after Ryzen reviews (beside the temporary drop after Intel licensing fiasco).
And since this is MCM and AMD told us how much they save on manufacturing, it's not that difficult to estimate how different CPUs will perform and how much they will cost.
It's quite unlikely that they'll attract investors by repeating what we already know. There's only one way: improving single-thread performance and fixing platform quirks that AMD fans here underestimate so much (and maybe also selling RTG altogether :-D).
And yes, homegrown investing gurus on this forum were expecting $30 by now. :-)
@Vayra86
No, that's apparently what you think they are doing...
On one hand you don't believe in stocks and you don't care whether someone wants to give AMD money or not.
On the other hand you're constantly moaning that all these problems are a result of their small R&D budget.
You know, I think there is a name for your affliction. And I don't need a large R&D budget to notice it...
Its cheaper for them, thus cheaper for us and performance is plenty close enough.
I also hope that Intel will keep doing what it is doing so we have a nice devided landscape in terms of tech.
However Intel being Intel I would not be surprised if they suddenly do this for some of their processors as well because ya know, Intel is rather cheap that way.
They're not at 2x the cost, they lost in almost every benchmark metric capable.
anandtech did a good review on epyc vs xeon (latest vs latest) and the database benchmark is bogus but the rest is good.
Database dataset was chosen at a size where it cannot fit in a L3 cache of Epyc, before you say that shows a weakness.
I've never seen any production databases at those sizes, only lab databases thus it's bogus.
Who runs a 8000\4000 usd cpu and have a 10-20mb database?...
What Intel and AMD are doing is not a race for CPU supremacy. It's not about surprising products, great reviews, leading in benchmarks or anything like that. This war only exists in heads of people like you, usually writing on forums like this one.
What actually matters for these companies is being profitable and having a stable business.
So yes, AMD was leaking badly about Zen, giving Intel time to prepare, but it doesn't matter. They had other tiny problems, like whether they will survive or not. And the whole leaking before and "defending" now has only one purpose: to show investors that AMD can make money. And it seems they actually can - that's why their market cap went up from total garbage to more expected level (considering their revenue).
I know you would love both companies to put all their effort in developing better CPUs, so that you would have more and more fps in games. But frankly, most of the world doesn't care about your fps. Even Intel and AMD don't. AMD could make Zen better for games, but why bother? This is a battle for a tiny portion of market. There's no point in becoming a leader at a huge cost. AMD simply wanted to become relevant once more - to be able to earn money on CPUs. And it seems they succeeded.
Powerful desktop CPUs are not the future. Most likely after Zen gets traction in the desktop world, AMD will put R&D money in other fields: AR/VR, AI, IoT and all these nice abbreviations. At this point they also have absolutely no product that could compete with Intel's ULV lineup and they won't cut it from their current Zen die.I did.
You see... there are 2 problems with the server segment that you didn't consider:
1) Intel actually sells those Xeon CPUs, while EPYC exists only in benchmarks at this point.
2) CPU performance is not as highly prioritized as in consumer segment (let alone gaming). EPYC has to deliver in other properties as well.
www.anandtech.com/bench/product/1832?vs=1851
If anything RR is the reason Intel might've been spooked & released mobile parts first, before desktop CFL.
I mean glued or not, it works.... And you're still paying more for the crappy TIM'ed chips.
AMD is making good, efficient use of what they came up with, can't blame them for that at all and it's something for now they need to capitalize on. They don't have the R&D budget Intel certainly has and this is AMD taking their best shot and making it count every way they can - Let's hope they continue doing so because in the end everyone wins, even the Intel guys do with lower prices.
Competition is good for everyone.
And I didn't say they're more efficient than Zen. I said that they exist and work beautifully. So Intel can make a good purpose-built CPU for passively cooled notebooks.
Lets face it. Intel mobile CPUs work beautifully in the frugal 7.5W mode, while it's difficult to imagine a Zen+Vega package sipping under 15W (but I'd love to be surprised).
You see: since this is MCM, the bigger it is, the more efficient it gets. But it's also pretty complicated as a result: Infinity Matrix, SenseMI, large cache... a lot of things that are important for the efficiency of the whole package, but need energy as well.
The result of this could be that Zen is very effective where it's not that important (gaming desktops) and not very effective where it really matters (ultrabooks).
Look at this Ryzen review:
www.techpowerup.com/reviews/AMD/Ryzen_3_1200/18.html
Ryzen CPUs are efficient under load, but suck at idle. All of them.
And now Vega:
www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_64/29.html
Vega sucks at both idle and load.
BTW: Intel HD is 1-2W. I bet Infinity Matrix itself draws more.Mobile KL and Broadwell were also released before desktop variants, Skylake and Haswell - together.
Mobile segment is way more important for Intel. That's where the new tech goes first.