Intel processor binning explained by Guy Therien | Talking Tech
(upbeat thumpy music) - Hi, welcome to Talking Tech. I'm your host, Marcus Yam, and today we're gonna learn about how Intel processors get categorized. You may have heard of the Core i9, Core i7, Core i5, and Core i3.
And we're gonna learn about how these different processors get categorized in their different performance tiers. To tell us more about that today is Intel fellow Guy Therien. Guy, thanks for joining me. For those who haven't seen our last video on Turbo, can you tell us a little bit about who you are and what you do? - Sure. Great to be back to talk to you, Marcus.
So I'm an Intel fellow, and I currently concentrate in the areas of performance and power optimization, specifically in something called performance segmentation. It has to do with making sure that the right capabilities are in specific Intel products across the price points. - Okay, so you're definitely qualified to tell me how these different processors get categorized. Can you tell me what's that process like? Is it arbitrary, is it random? How does that all work? - As you can imagine, there are certain price points in the market that consumers want to spend money, right? On certain capabilities. So these price points are translated into capabilities that we offer in our products. Specifically performance capabilities, IO capabilities, energy efficiency, and things like that.
And so the different numbers that you mention are brands, Core i3, i5, i7, i9, are levels of these capabilities at certain price points that meet the market needs. - Now, of course between core i9, core i7, you can have core counts, graphics, frequency. How do you determine and how do you differentiate which processor goes into which one? You have a wafer, and each wafer gets cut up into separate dyes, how do you figure out which one goes where? - So, at the highest level, all the processors are the same. With regards to a monolithic design, they're all the same.
- Within a generation. - Within a certain kind of generation. So we have certain kinds of dyes that are able to make multiple products from those dyes. Okay? So not every product we sell comes from one wafer, one dye on the wafer, but there are a series of these that we can make multiple products from. So when we have those parts cut from their dyes and packaged, then we classify them through something called class tests to understand what their capabilities are. And based upon those capabilities, which vary, as silicon manufacturing does have variance or manufacturing variability in terms of its leakage, its frequency capability, its voltage requirements in order to achieve certain frequencies, et cetera.
We ascertain what the needs are of these parts to run at certain frequencies, their capabilities, including functionality. Sometimes they don't even function at certain frequencies, or sometimes certain cores don't operate. And so once we know everything about the part that we've made, we decide whether or not that part is a candidate for productization, or it's a candidate to be crushed and go to the silicon graveyard. So- - Recycling.
- Recycling, yes. We don't bury them. - So what you described, I've heard of this term, is it generally called binning? Did you just describe a process called processor bin or silicon binning? - Yes, Yes. Internally we name the process, the classification process, as binning. And so that once we decide what's the characteristic support, then we fuse. Fuse is causing it to become a certain specific product.
- And at at what stage do these things get classified or binned? Can you tell right from the wafer, or does it have to be fused? Does it have to be in a package? At what point do they get designated Core i9, Core i7, Core i5? - Okay, so clearly there's a certain market demand for the price points, okay? We go out to our customers when we're coming out with a new product, we have a forecast as to what that product will achieve, most usually it's versus the previous generation product. And then the customers give us a non-committal request for volume, what they think they can sell, at certain price points. And therefore certain performance points. And so then what we do is we go through a manufacturing analysis to see if those criteria that they are trying to buy at certain price points can be met through an analysis for the volume that's needed. And if they are, then we lock that in and we head towards manufacturing to meet that need. - Is this process the same whether you're talking about chips for desktops or chips for laptops? Are they- - Yes, it's the same kind of process.
There's a set of requirements, a set of volume projections that we need to meet, and we use that in order to determine whether it's manufacturing feasible in order to deliver the volume that's required amongst all our customers. Because the cardinal sin of manufacturing is to tell your customer that you're gonna be able to supply them with product and then in the end not have that product to supply because of manufacturing issues. - You're talking about forecasting for customers and finding out what they need.
Is there a typical sort of distribution that happens out of manufacturing where you can kind of figure, we're gonna get this many Core i9s, and this many Core i7s? How does that work? - It all depends on the criteria that is being specified by the marketing organization in order to make winning products or leadership products across the price points. And so, in some cases the very, very highest price point, therefore probably the highest performance product, it has a certain amount of them that are gonna come out of the manufacturing process. So for example, of all the parts that you manufacture, let's call that 100% of all the material that you're gonna make, anything that's less than a projected 10% of that is something that's not assured, right? Because you can have unknown variation creep into the manufacturing process. And we also manufacture over time, right? So at the beginning, we have a pre silicon analysis of what we're gonna be able to deliver. Then once we get more mature and we start developing material and manufacturing the material, then we can do a post silicon analysis and we have more confidence in the amount of parts that we're gonna have over the long term that meet certain criteria.
But in terms of the top performing parts, or any bin specifically, any product specifically, it has lower confidence to deliver anything that's less than 10%. Because historically we've seen variation happen in these manufacturing processes, and so the lower amount of percentage we project of all the manufacturing yield, the more risk there is that'll actually come out in volume. So we don't wanna disappoint our customers, so we generally keep to around no less than 10% of the distribution for products. We make exceptions on occasion, but generally we feel comfortable that at 10% we can adapt to changes in the manufacturing over time.
- Now, I imagine that anything below that 10%, especially in the extreme high end that you alluded to, means that there's some really high end parts, really perfect choice parts there. I know that there's something we do with those, we call those thin thin bins, but I wanna learn more about what that means in terms of binning and a thin bin. What does it mean? - Okay, so sure.
So if you think about our highest performing parts today, this is our Core i9 parts. As you can imagine, we're making millions of parts. And the i9 parts, the i9 K parts, are a subset of those parts. And the top bin, I get, like I said, is no less than 10% usually of all the parts that we make in a given time period.
But because there are millions, we have to set a criteria that millions of parts can meet. But like I said, there's manufacturing variations. So within those millions of parts, even for the top bin, there are still variation within that bin. So some of the parts are significantly better than others, even within that top bin. So when you're making millions, you can make hundreds of thousands, 100,000, tens of thousands, that have significantly higher performance but they have a higher risk to manifest into actual parts. Your bet is in there as to whether or not they actually come out in the end.
- And we had a recent "Talking Tech" to celebrate the release of the KS series, and I assume KS is one of those thin bin parts, is that correct? - That's right. So the KS represents the highest performance parts that we make currently for client. And it is a lower volume part by virtue of its manufacturing capability, it's diversity amongst other parts. - Is it difficult to forecast how many thin bins you might have given a generation, or in a general, I don't know, manufacturing window? - When we are post silicon, we've actually had our first silicon, and we can look at a sampling of those parts, say a few thousand parts, and we can look at the characteristics of those parts, we can, with more confidence, project how many parts that we would have for a thin bin.
But, that is still risky because manufacturing variation, equipment variation, can still creep into the silicon manufacturing process. So it's higher probability post silicon than pre silicon, but it still carries significant risk. So that's why, when we offer these parts, that risk is associated with those parts. So the volume can vary and we try to have enough volume to meet certain users need, and specifically those are the users who are the most discriminating. The ones that time is money, the ones that have to have, or need, or can take advantage of the highest single thread performance, the highest multi-thread performance, for their application. - So there's actually a surprising amount of science and process that goes into selecting categorizing these chips.
Now, I've heard the term silicon lottery. That's when enthusiasts would buy a chip and maybe they might hit the jackpot and get this really choice chip. They're able to really push the frequencies really high up, or maybe it's really stable at lower voltages.
Does this process, this very involved process that you described, does this is kinda negate the ability for someone to hit the jackpot or is there to be an outside silicon lottery at all? - Yeah, so no it doesn't. So let's explain a little bit more about what you might call the silicon lottery. When we have demand for a certain amount of volumes for chips, we manufacture to meet that.
That's not to say that there aren't more chips that could meet that criteria, right? So it's possible that you'll get a chip that we consider what we call a down bin, a chip that could have been an i9 that we have demand for an i7, but not an i9, and we fuse that down to be an i7. It's in spec for the specifications of i7, not i9. So some people look at it and they say, the silicon lottery is, I get a bunch of chips in, say, i9 and I overclock them to see how fast they go. And some go really fast and I hit the silicon lottery 'cause I have a really good one. But another one doesn't go as fast.
But if you want to use that, you're outta spec because you're overclocking. So the silicon lottery, in some people's mind, is that I try a bunch of chips, I see which ones I can overclock higher, and I cherry pick the best ones and I use 'em. But recognize that, when we do a KS skew, it's in spec performance. So it is guaranteed by our warranty, and manufacturing tests, and capabilities, to be functional at higher performance than the normal K skews.
- Thinking about kind of the release cadence, usually the enthusiast models, the unlocked models, kinda come out first typically, and KS usually follows a little bit after, even after the mainstream parts. Why is that? Why the time difference between the release of a processor and the appearance of a KS? - Okay. So there's a few reasons for that. One of the reasons is is that by coming out with KS after the original one, we have higher confidence in our ability to deliver a specific criteria of the KS. So by coming out with K, looking for volume in the manufacturing, and doing an analysis of that volume, understanding the capability, we can better choose the criteria for KS. Tweak it, refine it, so that we can get the highest performance we can at the volume targets for this, like you said, thin bin skew. - Right.
And going back to, people who wanna have the latest and greatest things, and the fastest things, those K skews that come out first, those are for the overclockers. Now, these KSs, they sound like they're great kind of assured high performing parts. Should overclockers be going more towards the KS skews or should they be looking at K skews and play the silicon lottery? - Yeah, so that's a great question. So, certainly, by virtue of the fact that we have screened the KS parts for higher capability, it is most likely the best part to overclock. But like I said, once we get the amount that we need, that doesn't mean that a K part couldn't outperform a KS part if it was overclocked.
It would be out of spec, right? But they might save some money by doing it. So there's no guarantee that a K won't outperform a KS when it's overclocked. But KS is a excellent starting point for overclocking because in order to get higher performance in general, you typically need parts that require lower voltage, which is a manufacturing capability, and lower thermals, et cetera. Even though sometimes we raise thermals and things like that, it's really the cream of the crop of the parts. But remember, there aren't as many of them in existence, right? So you can't expect all the volume from K to go to KS 'cause there just won't be any parts, - Right. - right?
And they cost a little bit more money usually too. - So besides overclockers, who else would really benefit from having a K, or who would wanna build a system with K, or buy a system with K. And if they're building their own system, are there any special considerations, design, cooling that they need to consider? - So, again, another great question. So the top of the top, the highest performance processors tend to be the ones that have the highest power demands and therefore the highest cooling demands. So the kind of people that might use them, overclockers overclock oftentimes just as a hobby, to overclock to get the speed.
But there are real reasons why people need highest performance in computing. If you think about financial analysis, high frequency trading, oil and gas, formula one simulation, America's cup hole simulations, all these kinda things you might do on workstations, there's still a lot of money to be made by having a fast computer. In other words, when your computer's fast, time is saved. And so it's a trade off between time and money, right? So when you have the highest performance compute that you can possibly get in the KS, you can get your work done faster. And there's a lot of people, a smaller segment than the general population, but there's still a lot of people who need the most highest performing compute they can possibly get.
And that's the target for the KS, the most discriminating users. Now these come with higher system costs, right? Because you need more fans, liquid cooling oftentimes to get the most out of it. We have performance capabilities that are temperature dependent. eTVB is one of 'em.
And so, if you have this part in a system that's designed for it, or you might say over-designed for it, you can get the most performance outta that part by having that investment in the higher thermals, higher cooling, higher current, things like that, even without overclocking. - So that's a lot on binnings and how we classify the processors that they are. Are there any kind of other thoughts, anything else that people really should know when they're shopping for the next one and thinking about how their processor was binned? - Sure.
So, if you're looking for the best processor for the money, it's typically going to be the K skew, right? Because the K skew is assured in volume, a very high performance for a great price. If you're very, very discriminating and you need that extra amount of performance and you wanna make the investment to go to one step further, then there is a certain number of skews, I'm sorry, certain number of parts that we have that because of the large volume that meets the K criteria, we have that small number that meet an even higher criteria that is manufacturing assured to be within spec for the specifications that you read for the KS skew. Typically it's a few extra bins of performance. And this gives you the most performance that you can buy.
So my recommendation is, is to consult your conscience as to what kind of performance that you need or want. And if you really want the best of the best, buy the KS skew in spec. If you really want to experiment at a lower price point, you can still buy a K and overclock it, 'cause that's the nature of it. You can overclock the KS as well, but those are the choices that you have to make.
- So this is a super intricate process, bidding for high performance parts. Is there anything that Intel's learned in this process that can apply to the other processors, mainstream parts, anything else? - Yeah, actually there is. So, my goal, of course, is to deliver the highest performance product that we possibly can.
In this case, these thin bins, or these KS units, are those processors. And I don't manufacture. I don't work on the silicon process.
I don't work on the silicon architecture. But there are many things outside of that, the control capabilities, the Q&R capabilities, or requirements for these processes, I'm sorry, for these processors, that I'm involved with to look at ways to optimize them. So for example, if we are working on certain specifications for current, or for power, and our parts are guaranteed through manufacturing not to exceed those specifications, how accurate are we at tracking to those limits? Of course, every time I improve accuracy I gain more performance and reduce guard band. So guard bands are the enemy of performance. And so by pushing the envelope on these special skews, these KS skews, we are able to learn how to optimize the K skews, or all the other parts as well.
So in order to get the most out of the process by tweaking the KS skew. So yes, we are able to learn a lot about thermals, a lot about power delivery, a lot about power control, current limiting, things like that, power controls, power limits. By analyzing and optimizing them, it actually impacts the entire rest of the population of parts. - Now we've talked a lot about high performance parts, but does binning have another origin? Binning wasn't always just to find the highest performing part. Why does Intel bin, and what was the inspiration or the need to bin in the first place? - Well, okay.
The first thing is that when you make parts, all parts are not the same, okay? So some parts have different characteristics than others. So if you wanted to all make the highest possible performance parts and you only could offer this highest performance part, you might say, that's 10% of our volume and you have to throw away the other 90%, right? - Right. - In which case you wouldn't have the opportunity to sell as many units, and you'd have to charge a lot more money for that top 10%. So it really comes down to selling what you make, or selling what you make at the right price point in order to consume the entire volume.
So binning is the art of dividing up the outcome of manufacturing to understand what product it would be suitable for at what price points, and by virtue of that, consuming all the material that you produce. - And that's not just for frequency, we're talking about a certain number of cores, maybe- - Cache, cores. - Cache, cores, graphics. - Yes. - Everything like that.
And those things are all factors that make otherwise chips that don't fit in that 10% still perfectly- - Fine for a certain purpose. So sometimes things don't perform as well, they function but have a variability in their performance or their power. So for example, a desktop chip's power might be higher than a mobile system's power necessarily. And also it's, like I said, for function. So like you said, if graphics doesn't work or if some core doesn't work on a large core count dye, we can potentially sell that as a lower core count dye.
- Got it. Well, Guy, thanks so much. I can't offer you a thin bin, but I can offer you a Thin Mint. - Oh.
- Would you like a Thin Mint? - Yes, I would. - Okay. Well, please do. - Thank you so much. - Alright, well. Well Guy, thanks so much for talking tech and explaining binning to us.
- You're welcome. - Alright, thanks so much. - Thanks for being here. Bye now. (upbeat music) (bright music)