Opinion: Why Counting ‘Platform’ PCIe Lanes (and using it in Marketing) Is Absurd

It’s at this point that I’d like to take a detour and discuss something I’m not particularly happy with: counting PCIe lanes.

The number of PCIe lanes on a processor, for as long as I can remember, has always been about which lanes come directly from the PCIe root, offering full bandwidth and with the lowest possible latency. In modern systems this is the processor itself, or in earlier, less integrated systems, the Northbridge. By this metric, a standard Intel mainstream processor has 16 lanes, an AMD Ryzen has 16 or 20, an Intel HEDT processor has 28 or 44 depending on the model, and an AMD Ryzen Threadripper has 60.

In Intel’s documentation, it explicitly lists what is available from the processor via the PCIe root complexes: here 44 lanes come from two lots of sixteen and one twelve lane complex. The DMI3 link to the chipset is in all but name a PCIe 3.0 x4 link, but is not included in this total.

The number of PCIe lanes on a chipset is a little different. Chipsets are for all practical purposes PCIe switches: using a limited bandwidth uplink, it is designed to carry traffic from low bandwidth controllers, such as SATA, Ethernet, and USB. AMD is limited in this regard, due to spending more time re-entering the pure CPU performance race over the last few years and outsource their designs to ASMedia. Intel has been increasing its PCIe 3.0 lane support on its chipsets for at least three generations, now supporting up to 24 PCIe 3.0 lanes. There are some caveats on what lanes can support which controllers, but in general we consider this 24.

Due to the shared uplink, PCIe lanes coming from the chipset (on both the AMD and Intel side) can be bottlenecked very easily, as well as being limited to PCIe 3.0 x4. The chipset introduces additional latency compared to having a controller directly attached to the processor, which is why we rarely see important hardware (GPUs, RAID controllers, FPGAs) connected to them.

The combination of the two lends itself to a variety of platform functionality and configurations. For example, for AMD's X399 platform that has 60 lanes from the processor, the following combinations are 'recommended':

X399 Potential Configurations
  Use PCIe Lanes Total
Content Creator 2 x Pro GPUs
2 x M.2 Cache Drives
10G Ethernet
1 x U.2 Storage
1 x M.2 OS/Apps
6 x SATA Local Backup
x16/x16 from CPU
x4 + x4 from CPU
x4 from CPU
x4 from CPU
x4 from CPU
From Chipset
52 Lanes
Extreme PC 2 x Gaming GPUs
1 x HDMI Capture Card
2 x M.2 for Games/Stream
10G Ethernet
1 x M.2 OS/Apps
6 x SATA Local Backup
x16/x16 from CPU
x8 from CPU
x4 + x4 from CPU
x4 from CPU
x4 from CPU
From Chipset
56 Lanes
Streamer 1 x Gaming GPU
1 x HDMI Capture Card
2 x M.2 Stream/Transcode
10G Ethernet
1 x U.2 Storage
1 x M.2 OS/Apps
6 x SATA Local Backup
x16 from CPU
x4 from CPU
x4 + x4 from CPU
x4 from CPU
x4 from CPU
x4 from CPU
From Chipset
40 Lanes
Render Farm 4 x Vega FE Pro GPUs
2 x M.2 Cache Drives
1 x M.2 OS/Apps
6 x SATA Local Backup
x16/x8/x8/x8
x4 + x4 from CPU
x4 from CPU
From Chipset
52 Lanes

What has started to happen is that these companies are combining both the CPU and chipset PCIe lane counts, in order to promote the biggest number. This is despite the fact that not all PCIe lanes are equal, they do not seem to care. As a result, Intel is cautiously promoting these new Skylake-X processors as having ’68 Platform PCIe lanes’, and has similar metrics in place for other upcoming hardware.

I want to nip this in the bud before it gets out of hand: this metric is misleading at best, and disingenuous at worst, especially given the history of how this metric has been provided in the past (and everyone will ignore the ‘Platform’ qualifier). Just because a number is bigger/smaller than a vendor expected does not give them the right to redefine it and mislead consumers.

To cite precedent: in the smartphone space, around 4-5 years ago, vendors were counting almost anything in the main processor as a core to provide a ‘full core count’. This meant that GPU segments became ‘cores’, special IP blocks for signal and image processing became ‘cores’, security IP blocks became ‘cores’. It was absurd to hear that a smartphone processor had fifteen cores, when the main general purpose cores were a quartet of ARM Cortex A7 designs. Users who follow the smartphone industry will notice that this nonsense stopped pretty quickly, partly due to anything being called a core, but some hints towards artificial cores potentially being placed in the system. If allowed to continue, this would have been a pointless metric.

The same thing is going to happen if the notion of ‘Platform PCIe Lanes’ is allowed to continue.

Explaining the Jump to Using HCC Silicon Test Bed and Setup
Comments Locked

152 Comments

View All Comments

  • Spunjji - Monday, September 25, 2017 - link

    Super relevant, because they indicate how badly thermally limited the CPU is - which is hella good info to have if you're, say, considering delidding a $1999 processor because the manufacturer used toothpaste under the IHS.
  • tricomp - Monday, September 25, 2017 - link

    Poor AMD...
    No chance they are going to supply (even more) cpu's demand after posting this article..
    I am trying to purchase at list 7 systems for my customers in my country but there's nowhere I can find them beasts here..
  • iwod - Monday, September 25, 2017 - link

    I wish someone could do an article on that too. GF doesn't seems to be the limitation here. GF, should in theory more then enough capacity in their Fab 8 for AMD. Unless GF have some other big customers, otherwise AMD should really be bumping out as much unit as possible.
  • Atom11 - Monday, September 25, 2017 - link

    Can we please see one test (!), if you could possible manage, that shows the advantage of AVX-512 in compare to AVX2 when doing:

    1.) matrix multiply
    2.) FFT
    3.) convolution
  • ZeDestructor - Monday, September 25, 2017 - link

    Give us a comparison to AVX1 and SSE4 too!
  • Gothmoth - Monday, September 25, 2017 - link

    threadripper delivers 80+% of the perfromance for less than 50% of the price.... you don´t have to be a genius to see what the better deal is (price germany: TR 1950x = 950 euro, 7890xe =2300 euro)
  • Spunjji - Monday, September 25, 2017 - link

    Don't let that stop them equivocating about how companies who need that power yet somehow have no need for ECC don't care about cost because something something software TCO blah blah.
  • spdragoo - Monday, September 25, 2017 - link

    I'm trying really, really hard to think of a company that, at some point or another, doesn't say, "Equipment X may outperform Equipment Y, but the extra cost to buy Equipment X is too much, we'll just make-do with Y instead." Especially since 100% of companies have a limit on their budgets.

    What's that, you say? Multi-billion dollar corporations don't have to worry about the money they spend? Someone apparently didn't pay attention in their Econ 200 class, or their Introduction to Accounting coursework.

    By definition, every business has a *finite* amount of money they can spend, based on a) how much money they collect from their customers, b) how much they can recoup on the sale of assets (tangible or intangible), & c) how much they can get from "other sources" (mostly bank loans or by selling stock shares, or sometimes government grants, but you might find the occasional situation where a generous benefactor just bequeaths money to a company...but I doubt you'll even see that happen to 1% of the companies out there -- & no, venture capitalists pouring money into a company is *not* a situation where they "give the money away", they're getting something for their money, usually stock shares or guarantees of repayment of the loans). Of that money, some of it is earmarked for employee compensation (not just the executives, but the office drones & lower-level employees that do 99% of the actual work), some of it goes towards taxes, some of it pays for rental payments, some for loan payments, some for utilities (telephone, Internet, electricity, gas, water, etc.), some of it may get set aside for "emergencies", some gets earmarked for dividends to the shareholders, etc. That means that a (relatively) small portion is set aside for "equipment replacement". Now, if the company is lucky, the lion's share of that budget is for IT-related equipment...but that covers more than just the office drones' machines, that covers everything: server racks, storage services, cloud vendor payments, etc.

    And that is where the price comes into play. For probably 90% of office users out there, not only is Threadripper an overpowered product, so are these products. Heck, we're in the middle of an upgrade from Windows 7 to Windows 10, & they're taking the opportunity to replace our old Sandy Bridge i5 machines with Skylake i7 machines. Sure, they're running faster now...but the main reason they're running faster is because we went from 32-bit Windows to 64-bit Windows, so our PCs now have 8GB of RAM instead of 4GB. That helps with our workload...which primarily revolves around MS Office & using browsers to access & modify a number of massive databases. Having an 8C/16T CPU, let alone a 16C/32T CPU, wouldn't provide any boost for us, since the primary slowdown is on the server side.

    These are going to be expensive systems for specialized purposes...& those individual companies are going to look at their budgets very closely, as well as the performance benchmarks, before deciding to purchase these systems. Sure, they may hold the performance crown...but not by that big of a margin, & especially when compared to the margin that gives them the "most expensive price" crown.
  • BrokenCrayons - Monday, September 25, 2017 - link

    Human labor is more expensive than hardware. The 20% additional performance for $1000 more can be earned back quickly by the increased productivity of your workforce (assuming your management staff is effective enough to keep the employees gainfully employed of course and that's certainly not always the case).
  • vladx - Tuesday, September 26, 2017 - link

    Indeed the difference in price is pretty much negligible in a professional setting.

Log in

Don't have an account? Sign up now