A Hardware Wonk's Guide to Specifying the Best BIM and 3D Workstations, 2017 Edition

MATTHEW STACHONI: All right, let's get started. So, hi, everyone. Thanks for making this class. I can only assume that all the good ones were taken, Right?

Yeah. It's no shame in that. It's fine. I understand.

My name is Matt Stachoni. I am currently the BIM Manager for Tutor Perini Building Corporation. We are located in Philadelphia, and we build very, very, very big things.

If you've walked the Strip in the past couple of days, you've probably seen a lot of the stuff that we've built. The Paris, Caesar's Palace, Luxor, all of city center, Cosmopolitan Aria, Verdant-- a lot of stuff. Luckily, I wasn't involved in any of it.

And you might be wondering, what's a construction guy doing talking about computers and BIM, and stuff like this. Well, I used to be an IT guy. Since then I've gotten better. And now I have somebody else doing IT for me, which is awesome.

So we're here to talk about specifying the best BIM and 3D workstations. Did everyone have a chance to--

[LAUGHTER]

Everyone have a chance to download the handout? If you've had enough bandwidth to do that? It's 100 and, I don't know, 30-- I lost count. After awhile I was, like, just typing away. It's huge. There's a lot of detail in it we're not going to get in this class, because I've only gotten 90 minutes. So I've tried to compress the presentation down to just the facts of what's going on out there.

So, what we're going to be talking about today in this wall of text that you see here. Essentially, what we're going to be doing is looking at all the different components that make up your PC. And how they affect your performance in AEC applications, like Revit, and Navisworks, 3D MAX, and so on.

This year was a particularly huge year for hardware. When I taught this last year, I said that's it, we've hit maximum hardware. I don't have to teach this class anymore. There's really nothing new that is going to come out.

Every year it seems like it's just dribs and drabs of slightly better stuff, right? Just enough to keep Intel and NVIDIA and AMD in business. This year was different. This year was a huge year for hardware, especially on the AMD side.

So we're going to be talking about all the new stuff that's come out in the past year-- some of the rebranding and refocusing of things that Intel has been doing. And we're going to be looking at, specifically, processors and graphics cards, or GPUs as I refer them here, in serious detail. Excruciating detail, I would say.

So some key learning objectives here-- understand the relative computing demands of the software that you're using. This actually-- I'm going to skip this first one in class. It is in the handout. It's 30-some odd pages of text and graphics and all kinds of stuff.

I'm going to skip over that just because of time constraints for the purposes of discussion right now. But we are going to learn all the latest and greatest stuff in CPUs, graphics, memory, storage, and so on, how to intelligently specify a workstation, some buying tips-- things like that. OK?

So here's the sort of session outline where we're going to look at some of the stuff that's happened in 2017 specifically, and really talk about, what are the things driving what's happening in the hardware industry. For example, the gaming industry is huge. How is that driving everything? How are different things really sort of stacking up to drive prices upward or pull prices down?

Then we're going to look at the different platforms that you can specify a workstation in. There are currently four of them. Sort of a mainstream desktop kind of a thing is the Best Buy special of the week kind of machine that you would look at. Although, for our purposes, mainstream desktop is really high-end mainstream desktop.

The next step up above that is high-end desktop, or HEDT as Intel calls it. And it was really invented by Intel. And what Intel did was essentially took their Xeon line of processors and turned them into consumer grade processors that you can buy off the shelf. So with that comes a lot of opportunity to build some really exciting systems for a lot less than you could if went out to a Dell or at HP and bought a precision workstation or something from them.

We're going to look at then, the professional workstations or the xenon based workstations. And then we'll look at the mobile workstations as well. System memory, graphics cards, storage, so and so, such and such, and such.

So we're going to try to get into peripherals too. I always end up having to skip that for time. OK. So those are the four things we're going to be basically hitting really hard in this class.

So obligatory disclaimer-- no one's paying me to say nice things about them, which is really unfortunate.

[LAUGHTER]

I don't officially endorse anything here, although I do have my own personal likes and dislikes, I should say. Just because of being in IT for so long, I've had to deal with a lot of this stuff in the past. And I know where the sore spots are.

But of course, if anyone, say from Intel, or NVIDIA or AMD is listening to this right now, and they want to send me stuff for me to say nice things about them, I probably will. I have standards. You know-- integrity. It's just down here.

[LAUGHTER]

OK. So how many people are using Revit? Everybody, great. How many people are not using Revit? Yeah, sorry.

[LAUGHTER]

It doesn't matter. We'll get to you. How many people are using 3ds Max? Quite a few.

How many people are using Navisworks? Everybody, of course. Yeah.

How about Lumion? Or Rhino? Or V-Ray? Some of those other third party applications that tax your system just as much so, if not more so, than Revit. OK.

I'm assuming the people that use Lumion and V-Ray are creating renderings, animations, things of that nature. How many people are running hardware that's about two or three years old, and you're not really sure where you're supposed-- OK. Then you're in the right room. All right. Good.

How many people are shopping for a workstation right now? OK. Very good. All right. Well, hopefully this will work out for you. I'm not making any promises, but let's see what happens.

Let's see what's going on in 2017. A couple of big things happened this year. First of all, the continuing reason for a lot of things happening is that Intel, AMD, NVIDIA, which are the big players and pretty much the only manufacturers we're going to really talk about, are mainly involved in raising the performance per watt. In other words, keeping the performance steady, but lowering the power draw of everything that they make.

And they do that by making transistors smaller, coming up with a smaller process, coming up with a smarter way of putting things together. But that's the idea, is to maximize that performance. Running, sort of, with that is Moore's law, which we'll cover a little bit in detail, which is essentially specifying that we double the number of processors or the number of transistors on a processor every 18 to 24 months or so.

So we're looking at some of the limitations of Moore's law. We're looking at parallel processing. multiprocessing, multithreading These are important, because as Autodesk applications mature, they get more and more able to run across multiple cores of execution-- multiple threads of execution-- which can be sent across multiple cores.

That's really good, because what's happening in the industry by and large, from a CPU standpoint, is that four-core system that you have on your desktop is an antique. We're going to six, eight, 10, 12, up to 18 cores that you can just buy off the shelf. And is that a good investment? Is that not a good investment? Stay tuned. That's what we're here to talk about.

We also have, on the other side of things, with the graphics cards that are out right now, massive numbers of cores on those graphics cards-- thousands of them. So it makes that paltry 16-core CPU that you just got look like a child's toy, because these things have massive numbers of cores. They're simpler, obviously. They don't do as many things as a CPU can, but because there are so many of them, the processes that you have that are highly parallel in nature can be sent to those thousands of cores and be done a lot quicker.

So, now, that we have programmable graphics cards, we can send problems to those graphics cards-- some of which aren't even graphic in nature. They can crank away on things like echo location problems, and all kinds of different kinds of ray casting problems, and scientific problems that a CPU would choke on. So we have a little, tiny supercomputers in our workstations nowadays. The power is largely going untapped.

Gaming engines-- how many people are using something like Stingray or Unreal Engine in their work? Gaming engines are basically the tools that game developers use to create their games. Well, those same tools can be used in AEC visualization, where you export out a Revit model. The main way you do it is throw it Max, do some things with it there. Thrown it into Unreal, throw it into Stingray. or just go right directly out of Revit to Revit Live, and you get an interactive environment.

Virtual reality plays into that, because all those gaming engines can feed right into a virtual reality-- the big sunglasses that you see people wear-- and augmented reality. Then we also have other things that are playing into this. The increase in using virtual machines, virtual design, infrastructure, and cloud computing. Actually, that should say virtual desktop infrastructure. That was late. I can tell already.

Then we'll look at price-performance compressions. AMD, obviously, is back. That's a big one. And this is a new one-- this came up this year. Anyone price out graphics card recently?

[INAUDIBLE]

Yeah. They're hundreds of dollars more than you would have expected. $800 is a good starting point nowadays. The reason that's happened is because of the damn cryptocurrency miners. The thing is-- I don't know if you know about bitcoin or any of those other things, but basically, you can buy a high-end rig, stuff it with a bunch of graphics cards, and mine currency-- mine cryptocurrency out of it.

Now, you'll make like $0.20 a day, so it's going to be a while before you ramp up enough to do something decent with it, like buy a car or a boat. But essentially, you've got guys that are buying truckloads of graphics cards, depleting the supply, the price goes up, and until it equalizes a little bit we're stuck in a bubble right now. OK.

All right. So at a core of what is driving a lot of this stuff is this performance per watt kind of thing, lowering the cost of components, getting more power out of what we're working with. And it really comes down to how transistors work. And it's of a magical kind of thing.

Basically, you have a source and a drain and they're embedded in silicon. And the source and the drain are infused with phosphorous, which basically makes them semiconductive-- which is why we call them semiconductors. And then it's essentially an electrically driven switch. You apply a voltage to a gate that sits on top of it. And when you apply that voltage, it creates an electric field. And then underneath of it, little electrons will pass from the source to the drain.

They'll pass underneath, under here. They'll start to collect here, and then when you apply the voltage down on the gate, they'll transfer across. So you've got this switch. And with that switch, you can create logic circuits. And with logic circuits you can create Grand Theft Auto-- stuff like that.

So that's what is happening. So the process that you see here-- this process width-- is basically, generally governed as the size between the source and the drain. Although, that's it's a variable thing. But it's presented as a process node. And right now, our process node of choice for all of the major CPUs out there is 14 nanometers. That's how big that little thing is across there.

We're getting down to a situation where the smaller and smaller we go with this, the harder and harder it is to go smaller. So we've been on 14 nanometers for about two years now. And Intel and AMD are trying to get down to 10 nanometers, which is the next milestone. But the problem is that as you get closer and closer, you start to get out of the classical mechanics-- classical physics-- realm, and into the realm of quantum mechanics, where electrons will just sort of pop up out of thin air and sort of ruin the idea of being able to control the electrons across the transistor.

So as we shrink down from as we double the number of transistors every year-- or every 24 months, or whatever it is-- it's a logarithmic function. So you'll notice that the numbers here go up very fast along this straight line. So you end up basically running out of atoms.

So we're down to, I think, 14 nanometers is something like-- I did the math-- 37 atoms or something really small. And at that point, you're going to get this quantum tunneling effect that they call where the electrons just pop up, and you have to build a very tall gate-- basically a tall wall-- that people can't-- "people"-- electrons can't jump over.

[LAUGHTER]

Sorry I had my head in the news too much.

[LAUGHTER]

All right. So, basically, the scope of this though is pretty amazing. Because in 1971, you could fit about 2,300 transistors on a chip. And in 2017, you have almost a billion and a half. And the analogy I like to use is that Boston Symphony Hall, in Boston-- I used to live right there and right across the street from it-- it holds about 2,300 people, conveniently, during pops season.

So if you think of Boston Symphony Hall as a transistor-- 2,300 people-- it works a lot like that. Now, 1.3 billion is the population of China. It's like stuffing the population of China inside of Boston Symphony Hall. And the lines for the bathroom were long enough as it were with 2,300 people. With 1.3, it's going to be really difficult.

It's like I'm shrinking a five foot six person down to a grain of rice. That's basically the idea that we here. That's the scope of Moore's Law over time.

So we've talked a little bit about some of the different things going on with parallel processing. Now, you've probably heard the term multitasking, which is sort of the old school way of doing more than one thing at one time. Although, it really wasn't. It was basically slicing up your computer time into small slices, and one program will do something here, another program would do something here.

So it was really an illusion of parallelism. It wasn't true parallelism. Now, we have multiprocessing where a software can spin off bits and pieces of itself to different cores. And with multithreading especially, that's where you have something like rendering being able to push off its problem onto as many cores as it possibly can and produce that image that much faster.

In fact, if you ever see rendering done in Mental Ray, you may notice that it has little boxes that fly around the screen as you're rendering this image. Those boxes represent each CPU core on your system rendering that box-- that bucket. And so the more cores you have on your CPU, the faster it will go.

A lot of people ask me, what kind of graphics card should I get to render something? It doesn't matter, because it doesn't touch rendering at all depending on these kinds of rendering engines like Mental Ray and Art Render-- a lot of the stuff that's out there-- are purely CPU-based. So if you're doing a lot of rendering, and you're waiting hours, minutes, days for your rendering to go, it's because you don't have enough horsepower behind it, and number of cores to put at the problem.

So with that, we've seen a big change going to cloud computing-- especially cloud rendering. How many people use cloud rendering out of Revit, and Max, and places like that? That's really weird. Last year, that number was way up. Cloud rendering is the best thing you could do in Revit, because you send that rendering off to the cloud, Autodesk has thousands of CPUs just cranking at that problem away, and it can return that rendering really fast. You don't as much control over it, but the imagery is actually pretty good.

And I've worked in plenty cases where we had a meeting at one o'clock. At 9 o'clock, I was sending renderings, 50 of them. We'd all get them back in about an hour, and then we could assemble the PowerPoint do the presentation by that. So it really has a big effect on how much you can get done in a short amount of time.

When we're talking about judging CPU horsepower, one of the benchmarks that you'll see used a lot if you look at reviews and things like that online, and what I use in this handout, is Cinebench, built by the guys that make Cinema4D. It has a single multithreaded performance benchmark. It's pretty accurate in terms of judging the relative speeds between multiple different CPUs and so on. So you can really judge a CPU on its merits based upon that benchmark.

The one thing about benchmarks though, is that it's not AutoCAD-- it's not Revit. It's not Navisworks. It's not 3ds Max. The thing about buying a CPU nowadays is that, especially with the AMD CPUs that are out, you may think, oh, well, AMD has these great CPUs for a good price, maybe I'll look at those. Well, that's fine. The problem is that we see, especially in certain benchmarks that you can find online, and I document them in the handout, is that Autodesk has had about 10 years to fine tune its applications for Intel CPUs.

And so, while you may see multithreading benchmarks for a Ryzen CPU that beat an Intel CPU, when you actually run that rendering in Revit, it might come out the other way around. OK? So the handout has specific website pages and things like that that show that action going on.

Like I said, Unreal Engine 4, you get imagery like this coming out for almost nothing. In fact, Unreal Engine is free. You can just download it, and run it, and do as much as you want to with it. Epic Games only charges you if you publish a game with their engine. But for visualization guys that want to use it all day long, no cost whatsoever.

Also Stingray has a nominal cost. I think it's $30 a month, which still isn't bad considering the power you have. And of course Lumion, which is the old standby. Lumion is a gaming engine. I documented it in the handout. I forgot what it is exactly

VR is another big thing. It's coming on strong this year. The prices have dropped considerably, making the cost of entry much lower. Anybody has some VR gear? A couple people.

It's really cool. I first got my VR gear, and I spent all day in my office climbing Mount Everest, right? And I'm doing this-- for hours. People are walking by, and they're, what are you doing? And I had explain to them I was-- , research or whatever.

[LAUGHTER]

Doing this. Like, what are you doing? Stabbing somebody? No, I'm planting the flag, and I can't quite get it.

It's weird, but it's a great experience. And we're using it in our presentations to clients and things like that as we look for new work. Because we can show somebody exactly what the hotel room that we're going to build them is going to look like. It's complete with finishes.

Augmented reality too, by the way-- anybody play Pokemon Go? Right? You had a little guy-- you're looking at your phone and you see what's behind it, but you see a little creature or something like that.

Anyone ever watch football? You see the line that goes across? And you start thinking, how do they put that in there? That's all part of augmented reality. That's the idea, is that you combine reality with a transparent screen that it can project something on it.

Pricing compression and volatility-- we've had, like I said, serious upticks in prices recently. RAM prices have gone through the roof. This difference here-- I think this is, what? From about $60 to over $180 for the same amount of RAM. I think it's 16 gigs, four, four-gigabyte sticks.

Video cards too. This is a $220 range right here, just to give you an idea of the scope of the change in graphics card prices. Again, driven by things like this these idiot cryptocurrency miner guys. I shouldn't say that. They're probably smarter than I am.

OK. Other things that are happening. We get, like I said, new hardware every year, which means the old hardware has to get off the shelf somehow. We do have a few-- very rare do we have something that is completely game changing. We're just getting so good at how we build processors, there's not a whole lot of headroom left to have something great and special happen.

We have pricing that's all over the place. And, again, the cryptocurrency guys. All right, let's look at the platforms we're going to look at.

The first thing is, in 2017, AMD struck a blow. They're the rebels blowing up the Death Star. And then Intel came back, and struck back. So essentially, what happened with AMD is they came out with these new Ryzen 7 CPUs-- actually, Ryzen 5 and 7 CPUs-- for the mainstream desktop, and as well as this new Threadripper line of high-end desktop CPUs with massive numbers of cores-- 12 and 16-cores on a chip for less than $1,000.

Intel's there with an 18-core chip for $2,000, going, ooh, we got to do something about this. So what they did, basically-- well, let me finish this up. Basically, AMD's strategy is more cores for less money. Not as fast as Intel in terms of raw clock speed, but more cores for the money. They have a different manufacturing process. They fabricate their chips a little bit differently from Intel, and it allows them to do these kinds of these kinds of things.

Intel's strategy was to tweak out this latest architecture that they've had called Skylake. They've had it since early last year. And instead of giving you a new architecture every year, they're just tweaking the old one and calling it something new. But it's really the same old thing with some tweaks.

So we've got what we call sixth generation, seventh generation, an eighth generation within the past two years. And they're called Skylake, Kaby Lake and Coffee Lake with this new version. They put more cores on a single die, whereas AMD will put fewer cores on a die and put more dies together.

In fact, the Threadripper CPU, if you opened it up-- if you take that metal lid off of it-- it's actually got four cores in it. I mean four dies. Four physical chips, but they're all glued together. Two of them are disabled, and the other two are eight-core CPUs. That's how you get the 16.

The other two that are disabled are only on there for stability, so that the heat sink doesn't wobble. It's just-- it's nuts.

[LAUGHTER]

Literally. That is literally true. Yeah. There's a picture if it in the handout. I didn't put it in the presentation, but there's a picture in the handout where somebody put the chip in a microwave, or whatever it is, and popped the lid off. And he's got four cores there, and he's like, well, what are these other four for. Can we turn them? Can we do something-- No. They're dead.

But it does hint at what AMD could possibly do. And just flick a switch, and turn those little puppies on, and now you've got 32 cores, 64 threads of execution going on. It'll probably run at 500 megahertz, because it would cook itself to death. But you never know.

[LAUGHTER]

So the idea with Intel's strategy is that what they want to do is they want to make sure that their chips are running as fast as humanly possible. And they have this differentiation between multithreaded performance and single-threaded performance. Single-threaded performance is important for certain applications. It's hugely important for games.

So if you're looking at reviews online, most reviews are game driven. How fast does Call of Duty run in this thing? Things like that.

Games are single-threaded for the most part. So they do measure-- when they do benchmarks, you have to think of them, well, OK, they're measuring single-threaded performance. But they will always throw in a Cinebench score that says, OK, this is what it does multithreaded. This is how fast this chip will render, which is a huge metric.

So what Intel is doing is they've got the mainstream desktop. They've popped the mainstream desktop CPU up to six cores now, instead of four. So you get 12 threads of execution instead of eight. Keeping the clock speed very high-- well over 4 gigahertz, up to 4.7. They've also come out with an expanded high-end desktop range that goes anywhere from six cores all the way up to 18, in two core increments.

They've done a lot of weird stuff in the Xeon space in terms of rebranding and re-jiggering around things like that. Because essentially the high-end desktop space is eating into the Xeon space. So there's a problem there that they have to differentiate these things a little bit better. They also have these aggressive turbo boost or internally overclocking their chips to make them perform better.

So the idea behind Intel is let's just make these puppies run as fast as humanly possible. We can beat out the competition on benchmark scores. And we can do decently well in multithreaded to keep things going.

What Intel had done in the past is they had this tick-tock model. What they would do is they would have a new process, they would shrink something down here say to the 32 nanometers, and then they'd come out with a new microarchitecture for that process. So the differences between microarchitectures are pretty huge in terms of the internal organization of the chip. And then once they get that microarchitecture working, then they will attack the problem of shrinking it down the next step.

So they go from 32 down to 22 nanometers, down to 14 down here. So basically, as they do the process shrinks, they would have another cycle where they would do a new microarchitecture on that new process, and then get the bugs out, shrink it down, and so on and so forth. So the difference between, say, a Sandy Bridge chip, the i7 2600K in an Ivy Bridge, which came out a year later, is not much, except that the Ivy Bridge is a smaller process. It's the same microarchitecture, just a smaller process. So it runs cooler, it runs at higher speeds, and so on.

So that was up until 2016. OK. After 2016, Intel hit a wall. They couldn't shrink this thing down fast enough. So they came up with this hokey kind of process architecture optimized thing, where they do the process, they shrink it down, then they give them new microarchitecture on it.

And then they just optimize the microarchitecture. They just tweak it out. Not a whole lot of performance difference over their chips over the past couple-- like, two years. What they've done instead is tweak the configuration. So instead of two cores, you get four. Instead of four, you get six, and so on.

OK. So what makes a good CPU AEC applications? Obviously, there is no CPU fast enough to run Revit-- not one I found anyway. Not on my models anyway. Let's put it that way.

There are CPUs that are good enough, obviously. But the best way to figure out-- when you're really getting into CPUs, go ark.intel.com. It's a phenomenal site. They've done a really good job of putting all of their stuff together. And you can get all the details and all the gory techno stuff about a CPU on that site. I have a lot of screenshots in this class on that, so I can probably save you a couple of trips anyway.

But the things to look for are the number of cores. You want to get the most cores for your money. You don't to spend way too much money for an extra two cores. That's what you'll see in the pricing structure is the pricing goes two, four, six, 10, 16, and so on. So the prices just go right through the roof.

But also pay attention to the core clock speed as well. If you've got two machines or two CPUs that are 3.7 gigahertz, and one six-core and is eight-core, you would think the eight-core runs better. Probably true. However, you might also have two CPUs that are at six cores, and one's running it 3.2 and one's running at 3.7. That can be a huge, huge point of difference in overall Revit performance.

Turbo Boost is another problem, or another situation you want at. Look for the highest Turbo Boost you can get out of it. Turbo speeds-- and I'll talk about this later-- but that basically gives you a judge of how fast the machine will run a single-threaded application.

The L3 cache size is also a big metric nowadays, because Autodesk, in their latest performance technical note, the link of which is included in the handout, talks about the level 3 cache and how important that is for certain operations in Revit. The Maximum supported RAM-- most people are probably OK with 32 gigs to start off with. But that is basically half of what a mainstream desktop processor can handle. They can handle 64.

If you need more than that. If you're really individualization and you're running huge renderings and things like that that take days or whatever, you might have to go more than 64 gigs. And then you're into looking at a different platform altogether.

The inclusion of an integrated graphics processor or IGP is completely unnecessary. Don't ever use it for mainstream work. They are fundamentally horrible. However, if your graphics card dies, then you have one built-in that you can use. OK? I've had a couple of times, actually, where that happened. And having a GPU on the CPU itself got me through a day without having to take a whole system offline.

Like I said, ark.intel.com has basically everything that they built under the sun with all the gory details of everything. So I invite you to go check it out. Turbo Boost-- this is a really interesting technology. Essentially what you have is a situation where you have multiple cores and they don't all run at the same speed. They don't have to. Which is a good thing, because there are times where you're running a single threaded application like a game, or minesweeper, or something like that, and it doesn't need all six cores or all four cores, right?

It only needs one. So what a Turbo Boost situation will do is it will overclock that one core. It'll ramp that one core up. If it had to ramp all of up the same time, the machine would cook itself to death. So the chip would break.

So instead, what Intel had in what they called Turbo Boost 2.0 technology is that the chip would have a multiplier number attached to it. This little number that you see-- where is it right-- right here. Like, three, four, five, six-- those are called turbo bins. And what they would do is the numbers would indicate how many hundreds of megahertz that particular core would ramp up.

So basically what happens is if you've got all four cores active, in this particular case, the number of turbo bins for all four cores being active is three, which means you get a base clock of 3,400 megahertz, or 3.4 gigahertz, plus 3 times 100. So you can end up with 3,700 megahertz, 3.7 gigahertz. However, if you've got one core active the number is six. Which means that it ramps that up to 4.0 gigahertz. That's what that whole Turbo Boost means.

Now, the thing about this is that this number that you see here, this little guy right there, varies from CPU model to CPU model. It depends on how Intel wants to basically tune their chips or set the clock speeds. So some of those numbers might be even zero, just because Intel's ramping the clocks speed up as far as it can go right out of the gate for market-driving purposes.

So that's what Turbo Boost 2.0 means. For Turbo Boost Max 3.0 is a new feature that came out just recently on Intel's Skylake-X chips, which are the high end desktop chips. Which is essentially a situation where Intel grades each CPU core. Gives it a score, and says, OK, this one can run this fast. This one can run this fast, or whatever. And picks the two best ones on a CPU.

Because they are all different. They're all slightly different. It picks the best 2. It grades them and records that. The OS can read that with a special driver you have to install. And the OS can actually send a single threaded application to that specific core, or those two specific cores. And you it can fine-tune performance so that when it needs it, it sends it not only ramps up a core-- jacks up the clock speed-- but it also jacks up the right core. So that it knows that that can certainly handle the highest performance, the highest megahertz that it possibly can.

It's, like I said, only on the latest and greatest Skylake-X high-end desktop. The problem with this, though, is it's really new. It requires a driver. It actually requires the motherboard manufacturer to have their act together when it comes to implementing this stuff. Which doesn't always happen. So if you're looking to have this work, be prepared for a bunch of BIOS updates as they finally get this stuff right.

Level cache-- when you're looking at specifying a CPU, one of the big things is how much cache is on-board. Typically speaking, the size of the level 1 and level 2 cache are going to be fixed-- probably 32k and maybe 256k for level 2 cache. Both of these caches are sitting in the core itself.

Here you see an x-ray of the actual CPU core where the caches are sitting in here. Level 1 is the fastest cache. It's also the largest in terms of the circuitry size. And it consumes the most power, so that's why it's really small. That's why it's only about 32k.

Level 2 is smaller than that, runs a little bit slower. And it basically handles whatever the level 1 doesn't. So basically, what happens when the CPU is looking to do something and it tries to find a piece of data, it doesn't want to run the system memory, because that's really slow. What it wants to do is find memory that's really close by, which is why these caches are on the core itself.

So it just has to go here for the memory. It doesn't have to go over here, right? This is where system memory is over here. So that's what level 1 and level cache do.

The level 3 cache is this big, huge box down here that is basically shared by all the other CPUs. So if this core needs a piece of information here, it has, maybe, some information that would normally be going to this guy over here, and it can go across here and find it over here. What you want to do is make sure that you're always maximizing how much cache you can access. Because if you have to go to system memory, it slows everything down.

Hyper-threading is a situation-- that's why you always see CPU cores given with a number of threads, and the number of threads is double the number of core. This is how you get eight threads on a four core CPU. That's why when you fire up Performance Monitor, you'll see eight bars instead of four.

Hyper-threading is basically what I call the granny in the checkout line kind of situation. You're in line-- I'm in line at the grocery store, and the woman in front of me, who is anywhere from 80 to 3,000 years old, has to write a check.

[LAUGHTER]

Right? And Mary Sue cashier comes over and says, I'm open. And then, I try to get out, but everybody else goes around. So I'm still stuck behind the check writer. Anyway, that's-- my life. Anyway-- but the idea is that opening of the second checkout line is essentially what happens inside the core. And the core now expresses itself to the operating system as two CPUs instead of one. Which is why you have this two cores four threads kind of situation going on.

The other consideration when you're talking about CPUs is if you want to run multiple graphics cards. How many people have multiple graphics cards in their station? Five or six lucky people. OK.

Multiple graphics cards are great. They are. If you're doing certain graphics operations like rendering, they are awesome. However, you have architectural limitations with how many you can stuff into a machine before everything starts to slow down. And that is basically dependent upon the number of PCIe lanes-- I should say lanes that have slots-- coming off of the CPU.

The CPU has, at minimum, 16 lanes of PCIe 3 bandwidth coming off of it, which can drive the graphics cards. Those graphics cards are tied directly to the CPU. They don't have to go through the chipset at all. But basically, what happens is if you have one graphics card, it'll run at what we call by 16 speed, or by 16 bandwidth. 16 lanes are going to that graphics card.

If you put two graphics cards in there, because you only have 16 lanes to play with, it splits it off into eight and eight, or by eight by eight. Now, the good thing about that is that these are PCIe 3.0 lanes, not PCIe 2.0. PCIe 3.0 doubled the bandwidth on each of those lanes-- or doubled the speed-- over PCIe 2.0.

So what ends up happening is eight lanes of PCIe 3.0 is equal to 16 lanes of PCIe 2.0. And that's about as much speed as a graphics card needs. So graphics card will run just fine in by eight mode. It doesn't have to be in by 16. So you can have those two cards running side by side and they will run just fine. There's maybe a 3% difference in overall performance, which is well within the margin of error of testing.

However, if you put a third one in, it has to split it up into one by eight lane-- or one by eight card-- and then the other two are going to be it by four. And that's when you start to see things start to drop. So you want to make sure that-- in fact, I think there's-- modern graphics cards, I don't think even fire up if you put them into a by four slot. So they won't even work.

So it's important that if you're going to run multiple graphics cards, that you take that into consideration when you're picking your CPU. Don't pick a CPU that's only has 16 lanes if you want to stuff five graphics cards in it. Because that ain't happening.

All right. Let's look at some specific processors now OK. So let's look at the eighth generation, what we call Coffee Lake CPU line.

What Intel did was fired up the marketing machine and said, what can we do to make this exciting to people buying a machine. Well, let's just add a couple of plus signs to that little 14 nanometer deal right there, and we'll just make-- that's like new technology. It's not. It's just tweaked out a little bit, OK?

So the big news is that they've upped the number of cores to six instead of 12. They've upped the cores to six instead of 12. That is the new baseline now. When you're looking at CPUs, you have a whole range of CPUs to look-- core i5, core i7s. They also have core i3s.

I'm discounting, right off the bat, all the core i5s and i3s. You can run Revit on a core i5 system. It will run just fine. It just won't be what you can run it on an i7 system.

So it doesn't make sense to buy an i5 system, because for a little bit more money you get a lot more performance. You get multithreading-- or hyper-threading. You get more cores. You get a whole lot of other stuff. So it makes sense to only focus on the i7s.

They also have up to 4.7 gigahertz boost clock, which is insane. It used to be you couldn't get anywhere close to that unless you seriously water-cooled or dunked it in a vat of mineral oil or something like that. So there's a lot of shenanigans going on behind the scenes to get those boost clocks up high.

So it is, basically, the fastest single-threaded performance CPU you can buy. If you're buying a system for games, this is what you want. OK? If you're buying a system for games and Revit, this is going to do pretty well. If you're buying a system for Revit only-- or rather a system for Revit and 3ds Max, you're not going to go wrong doing this-- going this way-- but there are other options depending on what your particular workloads are going to be. It also does have solid multithreading performance as well.

Here's the lineup, basically, all the i7s and i5s. You can see the i5s here are six-core, six threads, they don't have multithreading enabled. So that's why-- on these. I'm really liking the 8700k. That's the one I talk about in the handout.

The K on the end of that means that it's unlocked. The multiplier in there is unlocked. And when a multiplier is unlocked, what you can do is you can go into the bios of your system and up the multiplier, which cranks up the whole speed of the system all at once.

And so, that's called overclocking. And Intel took that to another step by adding internal overclocking with the boost frequency. But basically, you can overclock your system and make it run faster for the cost of better cooling apparatus, or whatever you need. So that's why water-cooling is a big deal right now, because they've got all-in-one water coolers that you can just buy off-the-shelf, shove in there, and get the better cooling performance.