Testing the Pieces

Before we get to the display and Thunderbolt specific testing I wanted to confirm that the individual controllers in the Thunderbolt Display were just as capable as those in the Mac it was connected to. For the most part, they are.

Following in Brian Klug's footsteps, I created two RAM disks - one on a MacBook Air and one on a MacBook Pro. I connected both systems to my local GigE network and copied giant files between them. I measured over 780Mbps going from the MacBook Air to the MacBook Pro, and 832Mbps in the other direction (images below). That's about as good as it's going to get.

Next I tested performance over FireWire 800 and USB 2.0. For FireWire 800 I used a Western Digital My Book Studio Edition II configured in RAID-1 and measured peak read speed from the device. For USB testing I turned to a Corsair Flash Voyager 3.0 (USB stick) and a SF-2281 SSD connected to a SATA-to-USB bridge. In both of the USB tests I measured write speed to the USB 2.0 devices. Apple appears to have chosen its FireWire controller well as performance was only off by 2MB/s compared to the FW800 port on the 15-inch MacBook Pro. USB 2.0 performance wasn't nearly as good however, I maxed out at 16.4MB/s and saw typical rates closer to 15MB/s:

Transfer Rate Comparison
  FireWire 800 USB 2.0 (stick) USB 2.0 (SSD)
Apple Thunderbolt Display 70.0 MB/s 14.1 MB/s 16.4 MB/s
Apple 15-inch MacBook Pro (2011) 72.0 MB/s 21.2 MB/s 32.2 MB/s

Both the audio controller and FaceTime HD cameras interface via the Thunderbolt Display's internal USB controller. It's likely that one of those devices is forcing the controller to negotiate at a lower speed and thus ultimately limit peak USB 2.0 performance through the display. Note the gap in performance is much smaller if you're looking at transfers to a USB stick vs. an SSD. I happen to have a lot of SSDs around so I tend to use them as glorified USB sticks, I suspect the majority of users won't notice much of a difference due to the lower overall performance of standard USB sticks.

FaceTime HD Camera

Although Photo Booth in Lion captures at 1080 x 720, using iSightcapture I was able to confirm that the sensor in the Thunderbolt Display appears to be able to capture 1280 x 720 natively. Quality is what we'd come to expect from the current generation of FaceTime HD cameras.

I tested the camera both in Photo Booth and in a FaceTime chat with our own Brian Klug. The experience worked fine in both cases.

FaceTime seems to have issues when one party is in a noisy environment but that doesn't appear to have anything to do with the Thunderbolt Display hardware as I duplicated the issue on a MacBook Air as well. If you're curious, the problem I'm talking about occurs when the party in a quiet environment is trying to talk to the person with a lot of background noise. The quiet party will hear audio just fine but the noisy party will get a lot of broken up audio from the other side. It seems like FaceTime is trying to do some active noise cancelation that ends up doing more harm than good. I confirmed it's a FaceTime software problem by calling Brian via Skype without any issues.

The Changing Role of Displays Thunderbolt Performance
POST A COMMENT

275 Comments

View All Comments

  • vision33r - Sunday, September 25, 2011 - link

    Apple products gets a decent review and people in this forum jump on it like every review is about Apple.

    Imo, the whole PC industry is in serious decline lately. No leadership and no innovation.

    Every company just decided not to do anything new and just copy Apple or wait for Apple to come out with something and then follow them.

    Every PC geeks ask for ports, ports, ports! However in the Apple world it's all about minimalist.

    Less ports and ports that multi-function inter-connect ports like Thunderbolt is the way to go to have a clean design.

    That's why Apple is selling a ton of macbooks and making a good profit and all these PC notebook makers are crying about making a tiny profit and they still can't make a cheap Macbook air clone.
    Reply
  • repoman27 - Sunday, September 25, 2011 - link

    The lack of headphone/optical digital audio output and audio line in/optical digital audio input minijacks on the ATD is a little puzzling. Not seeing the technical reasons why these couldn’t have been included leads me to chalk it up as a cost issue. The ATD contains a hell of a lot more circuitry than the previous 27-inch ACD and retails for the same price. Apple most likely took a haircut on their margin for this device and might have just hit the threshold of pain.

    I think the inclusion of only three USB ports is very reasonable, since there is only one USB 2.0 host controller and you already have two isochronous devices attached to it—the FaceTime HD camera and audio controller. Five devices sharing less than 40 MB/s of bandwidth is bad enough, giving people a little more rope to create unsatisfactory performance is probably not necessary. And if you really do need more ports, you can always use a hub. Adding a second USB 2.0 host controller could have provided more ports while avoiding some of the performance issues, but I don’t think there is enough headroom left on the Thunderbolt channel for any more stuff.

    Some quick and dirty math provides us with the following: 5800 Mbps + 1.6% for overhead for the video stream = 5895 Mbps. 1000 Mbps for the Gigabit Ethernet + (2 x 786.43 Mbps ) for the FireWire 800 ports + 480 Mbps for the USB 2.0 host controller + 25% for PCIe protocol overhead = 3816 Mbps. Together that comes to a little more than 9.7 Gbps, which is awful close to 10 Gbps. Although the PCIe traffic would probably rarely achieve this maximum, the only way to always ensure that there is enough room for the video signal is to limit the cumulative throughput of all included PCIe devices to less than 4.1 Gbps. Adding so much as one more USB 2.0 controller or a memory card reader would have bumped it into that grey area where there could be potential for interruption of the video signal. This makes me particularly curious about the performance of a Thunderbolt device in a chain that contains 2 ATD’s.

    Including USB 3.0 in a device like this at this point in time is just unrealistic. If Apple believed that their drivers and the currently available host controller silicon was mature enough to include in a product, it would be in the new Macs. The first product they release with USB 3.0 support would not be a display. The ATD is the first display to include a USB host controller, all other displays with USB ports merely contain hubs, and as far as I’m aware, there are no displays currently shipping which contain USB 3.0 hubs. Furthermore, a single USB 3.0 host controller can generate 5.0 Gbps of PCIe traffic, which would not fit on a 10 Gbps Thunderbolt channel along with a 5.8 Gbps video stream.

    If you think that including a GPU in a display like this would be a good idea, then you haven’t really considered the engineering. First of all, the only point in including a GPU would be if it provided better performance than the GPU that’s already in the host PC. If you wanted to retain the features that are in the ATD as well, this would limit you to a PCIe 2.0 x 1 connection for the embedded GPU. In order to not make the form factor totally wonky, you would have to go with a mobile GPU, but even still you’d need to beef up the power supply considerably and provide a custom cooling solution akin to what’s in the 27-inch iMac. So now we’ve made this display considerably more expensive, louder, heavier and power hungry, and the only Macs that will see any noticeable performance improvement are those that are limited to Intel HD 3000 graphics. The only Thunderbolt Macs that don’t have the option for a dGPU are the MacBook Air, 13-inch MacBook Pro and Quad-Core Mac mini, all of which are in the $999-$1699 price range. Including a GPU in the ATD would make the display cost more than most of the Macs that could even see any benefit from it. Meanwhile, of course, everyone would be paying the higher price, unless Apple produced and supported multiple models. This would also lead to a much faster hardware refresh rate for Apple because GPU technology matures much more quickly than display technology does. A user upgradeable display is probably not forthcoming from Apple, and BTO displays are not on the table for any rational OEM, so don’t even go there.

    As for the height adjustment, while Anand goes for the adjustable desk, and those with the dough might shell out for an adjustable VESA mount, I suggest a good old fashioned block of wood. You can also use a wedge shaped block for displays lacking tilt adjustment. If you’re not as concerned about appearances, you can even throw a couple phone books under there until you get the height just right, or add a three ring binder for tilt. Plop it on a lazy susan and you’ve got swivel. Apple’s design cuts costs due to its simplicity, looks attractive due to its form, and is less likely to suffer failure due to its construction. The latter reason being particularly relevant for a 23 lb., $1000 panel covered by a sheet of glass.

    For those who can’t figure this out, Apple’s displays are “glossy” because they have a big slab of glass in front of them. To create a matte finish option would require applying an antiglare film, or removing the glass, either of which you’re free to do yourself. This is not at all like cheap displays with a glossy film applied to the panel. Apple is obsessed with glass. You can be pretty sure that their current displays weren’t engineered during a coffee break, and that they have looked into a good number of the antiglare treatments currently available. If you require a display without a sheet of glass in front of it, don’t buy one of these. Despite all the protests in forums, unit sales of Apple products with glass clad panels are higher than they were when they didn’t have the glass. People clearly like the way these look on the sales floor, even the insanely brightly lit sales floors of the Apple Stores.
    Reply
  • Constructor - Sunday, September 25, 2011 - link

    While I agree with most of your reasoning in principle, you're making one incorrect assumption:

    Furthermore, a single USB 3.0 host controller can generate 5.0 Gbps of PCIe traffic, which would not fit on a 10 Gbps Thunderbolt channel along with a 5.8 Gbps video stream.


    Thunderbolt actually provides 20 Gb/s in each direction; 10 Gb/s are intended for PCIe and the other 10Gb/s are intended for Display Port tunneling, so with only one display connected, the display signal will have no impact on PCIe performance whatsoever (see the Thunderbolt performance test in the article).

    Since one TBD consumes at least slightly more than half of 10 Gb/s, there will apparently be some capability for an overlap (dual TBDs should indeed have a slight to moderate impact on PCIe data throughput), however.

    To create a matte finish option would require applying an antiglare film, or removing the glass,


    No, you'll have to remove the glass in any case, since a matte coating on a glass significantly in front of the actual display's surface would blur the display beyond recognition of any reasonably-sized text.

    Matte displays are by necessity bare panels with no protective glass in front of them whatsoever.
    Reply
  • repoman27 - Monday, September 26, 2011 - link

    Thunderbolt provides 2 x 10 Gbps, full-duplex channels per port / cable. A Thunderbolt device can only use one channel at a time, the other channel is to provide a switching path to ensure bandwidth to other devices in the chain. To quote an Intel slide from IDF "Each direction in each channel can be data and / or display." All of the PCIe and DP packets for the ATD are on a single channel. Of course, due to the simplex nature of the DisplayPort main link, there would be sufficient bandwidth in one direction, but there's not much point in having a USB 3.0 controller that offers 370 MB/s reads and 40 MB/s writes.

    Think of a TB daisy chain as a chain of five-port 10-gigabit ethernet switches with two uplinks between each switch and a device connected to the remaining port.

    The Pegasus RAID in the test was minimally affected because it's traffic was on the other channel. I'm guessing that it was the PCIe switch in the MacBook Pro's Thunderbolt host controller that actually caused the minor performance hit.

    I agree with you about the glass issue, and that was kinda what I was driving at. As long as Apple's design language is all about glass in front of their displays, there can be no matte option. I think a lot of people feel that Apple could just offer the same display but in an antiglare version, but in order to do so, they would have to come up with a totally different design for the front of the device.
    Reply
  • Constructor - Monday, September 26, 2011 - link

    A Thunderbolt device can only use one channel at a time, the other channel is to provide a switching path to ensure bandwidth to other devices in the chain.


    That couldn't work with a chain of up to seven devices. The first would then carry the stream for the second one while using up one channel for itself, and the second device would use up the second channel with nothing left for the other five devices. It can't be quite that simple.

    It looks as if dynamic partitioning of the stream should in fact be possible.

    The test on page four saw almost no performance hit on the Pegasus RAID when enabling the display (meaning the display was operating independently, outside of the 10Gb/s available for PCIe), and only gigabit Ethernet, FireWire and USB in the TBD actually took data throughput away from the RAID.

    This means that PCIe and Display Port are not necessarily on the same channel. The dynamic allocation seen in the test would not be possible in that case.

    However, PCIe performance with dual TBDs connected would be quite interesting, since that combination could actually max out the total output bandwidth of 20Gb/s. Maybe if Anand still has the review unit from Apple around, he can repeat the test (with a MacBook Pro, obviously, since his Air won't be able to drive both TBDs anyway).

    http://www.anandtech.com/show/4832/the-apple-thund...

    As long as Apple's design language is all about glass in front of their displays, there can be no matte option. I think a lot of people feel that Apple could just offer the same display but in an antiglare version, but in order to do so, they would have to come up with a totally different design for the front of the device.


    Indeed. I personally wouldn't mind a matte iMac or TBD, even though the glossy one is far less of a problem than many people appear to believe, and it actually has some advantages of its own.
    Reply
  • repoman27 - Monday, September 26, 2011 - link

    [QOUTE]That couldn't work with a chain of up to seven devices. The first would then carry the stream for the second one while using up one channel for itself, and the second device would use up the second channel with nothing left for the other five devices. It can't be quite that simple.

    It looks as if dynamic partitioning of the stream should in fact be possible.[/QUOTE]

    It's a switched fabric architecture with full-duplex links. Refer to my Ethernet analogy again. With more than two devices in the chain, each one gets the full 10 Gbps pipe, but not 100% of the time. With a little bit of buffering, everything is copacetic and you don't take a huge latency hit. I'm guessing that Thunderbolt is almost like an external version of DMI, and much of its logic comes from Intel's PCH.

    The big question is, of course, what happens to performance with two ATD's plus other TB devices in the same chain. They'll clearly be limited to the remaining available bandwidth, but how much will the switching architecture impact things?
    Reply
  • Constructor - Monday, September 26, 2011 - link

    It's a switched fabric architecture with full-duplex links.


    Not according to Intel's documentation, even if that's still relatively sparse at this point.

    They say it's a packet-switched architecture, which is much closer to what it seems to look like in practice. Apparently it is mixing PCIe packets and Display Port packets as needed, for which it is convenient that both are packet-based anyway.

    The interesting open question is if on this level the existence of two separate physical channels still has any real relevance or if the controllers treat both channels combined as a single logical 20Gb/s channel with only the PCIe interface being limited to 10Gb/s...
    Reply
  • repoman27 - Monday, September 26, 2011 - link

    I was actually quoting the Intel "Technology brief", which is definitely sparse, and more than a little glossy. Yes, Thunderbolt also uses a packet based transport layer.

    I'm pretty sure DP is limited to 10 Gbps as well, hence no DP 1.2 devices (not that there are any). DP 1.1a tops out at 8.641 Gbps without the 8b/10b encoding so it will always fit on one channel. At the physical layer, what they've done is taken the conductors used for the 4 simplex lanes of the DisplayPort main link and used them instead for 2 full-duplex TB channels. If you think of the channels as PCIe lanes, they could theoretically operate as two x1 links or be bonded into one x2 link with the data being striped across both channels. All the evidence seems to point to x2 operation not currently being supported. What's odd is to consider what these channels would represent when using an optical interconnect.
    Reply
  • repoman27 - Monday, September 26, 2011 - link

    I think I see what's going on here. The cross-bar Thunderbolt protocol switch in the Light Ridge chip is an 8-channel switch. 2 channels are connected to each Thunderbolt port, and 1 each is connected to the PCIe protocol adapter port, the 2 DisplayPort Rx protocol adapter ports and the DisplayPort Tx protocol adapter port. Any two Thunderbolt protocol adapter points in a chain can be linked via a single channel by the switching fabric, so long as they support compatible protocols. Which physical layer channels are used is thus completely arbitrary.

    My previous assumptions about the bandwidth limitations of the ATD were thus completely wrong.

    The limitation here is the single PCIe protocol adapter port in the current Thunderbolt controllers. Thus you can only move 10 Gbps of PCIe packets to and from the host PC. With the bare minimum of PCIe overhead on a Sandy Bridge system that would work out to a theoretical maximum of 1028 MB/s of payload throughput. Gosh, Anand clocked 1002.7 MB/s in his first go round with the Pegasus R6! Intel does hint that a Thunderbolt controller can contain "one or more PCI Express protocol adapter ports", so this limitation might be overcome by future controllers.
    Reply
  • Constructor - Tuesday, September 27, 2011 - link

    Yes, it does look quite promising for future development, indeed! :-)

    Just a few points:

    As far as I'm aware the four Display Port simplex channels (or two full duplex channels) just remain the same when used for Thunderbolt – only the data rate is increased. So there should still be two output and two input channels for two full duplex channels in total.

    And do you have any real information about the internal routing mechanism in the controller chips? As far as I understand, Anand's graphic was also rather speculative and indirectly derived from what is now known from the outside, so to say.

    As far as I can see, both Anand's and your conjecture appear plausible in light of the little official information we've got. A real, substantial manual for these controllers would still be a fantastic resource here. ;-)

    Or at least some official clarifications from Intel about the true nature of their capabilities...
    Reply

Log in

Don't have an account? Sign up now