MSI's GUS II: External GPU via Thunderbolt

47 Comments

View All Comments

• phatboye - Monday, January 9, 2012 - link

  From the above picture it does in-fact look like there is a power core going into the GUS II external enclosure. So what I am guessing is that what the article meant was you could only use whatever power that the power cord supplies through the PCI-e bus and that the enclosure does not supply aux power connectors.
  
  Still though that puts a large limit on what GFX cards this thing can work with. More than likely you will only be able to use lower performance GFX cards with this thing which makes it kind of pointless as in that case you might as well just stick to the integrated GPU that comes with most CPUs these days.

• SlyNine - Tuesday, January 10, 2012 - link

  Even low end GPU's can beat the crap out of IGPs, Plus this can be upgraded and you will never upgrade your IGP.

• ickibar1234 - Sunday, June 16, 2013 - link

  Not Intel's HD 5200 Iris integrated graphics.

• Bull Dog - Monday, January 9, 2012 - link

  Admittedly somewhat confusing the way it was written. The Enclosure receives power from an external adapter. The cord is visible in both pictures.
  
  Inside the enclosure however the means to deliver power to the GPU is through the physical PCI-e connector. So only GPUs without extra 6-8 Pin power connectors will work in it.
  
  Does that make sense?

• Sttm - Monday, January 9, 2012 - link

  There might not be a point in using a faster gpu that would require more then 75W as the bandwidth you get over Thunderbolt is going to be a limiting factor.

• zanon - Monday, January 9, 2012 - link

  Honestly it really shouldn't be much of a problem, surprisingly enough. It seems to have become very widespread "common sense" that GPUs need a lot of bandwidth, and it's certainly what I assumed. But in fact short of a multi-gpu multiple-monitor setup it makes little to *zero* difference for even x16 vs x4. Take a look at results like the tests HardOCP did:
  http://www.hardocp.com/article/2010/08/25/gtx_480_...
  
  Definitely not what I would have predicted, but there it is. The full size TB chips found in something like the MBP can put out 20 Gbps each way, the half size ones 10 Gbps, although this thing may only use 10 anyway. But that's something like x3 to x5 PCIe 2.0 equivalent, which based on testing should be plenty for any single screen setup even with a pretty high end GPU.
  
  Obviously, it's going to become more of an issue if people want to start combining or chaining a lot of stuff. Running an external GPU and trying to chain on a highspeed RAID for example, or wanting to have a multi-PCIe box, certainly looks like it could overload a single 1.0 TB port. It may be possible to bridge multiple ports, but longer term we'd have to hope it scales as Intel claims (at 100 Gbps per channel it shouldn't be a limiting factor under even near future use cases). Nevertheless, it is a pretty compelling use for the present. Being able to have full, standard PCIe slots for an ultrabook or equivalent would further knock down any remaining use cases for desktops. Docks that have enhanced active cooling that let the system's chips rev up higher would also help.

• repoman27 - Monday, January 9, 2012 - link

  All of the current Thunderbolt controllers only provide a single PCIe to Thunderbolt protocol adapter, as far as I can tell. Thus total PCIe bandwidth is limited to the equivalent of PCIe 2.0 x2.5. This would begin to bottleneck higher end cards pretty severely, I'd think. Still better than integrated graphics, but a pretty crazy expensive solution for gains that could be achieved vs. just buying a system with a better dGPU.

• zanon - Monday, January 9, 2012 - link

  >All of the current Thunderbolt controllers only provide a single PCIe to Thunderbolt protocol adapter, as far as I can tell.
  I don't know enough at all myself to whether this is a current hard limit or whether it could be worked around somehow in situations where no DP is needed. Maybe Anand could look into that, but at any rate even if it is hard limit it's not the end of the world necessarily (see below), and furthermore something that manufacturers should have been on earlier :).
  
  >This would begin to bottleneck higher end cards pretty severely, I'd think.
  Yeah, "I'd think," but again see my link above. I honestly truly would have thought that a drop on a GFX 480 at 2560 resolution from x16 to x4 would have made a difference, but in many of the games it literally had ZERO effect, nada. I would want to see some real tests about exactly what kind of real-world effect a bandwidth drop to have. Don't forget, 1920 is plenty common, and even mid range dGPU boards would completely, utterly slaughter any integrated or even higher end mobile graphic solutions.

• JNo - Tuesday, January 10, 2012 - link

  this.

• Penti - Tuesday, January 10, 2012 - link

  Actually they provide 4 lanes (x4) PCI-e (but can not communicate at full speed 4x PCI-e 2.0 obviously) except on Macbook Air where they have half the physical speed. If it where not physically connected to x4 PCI-e/PCI-e 2.0 on the motherboard and instead PCI-e 2.0 x1 you wouldn't get up to ~800MiB/s on products like say Pegasus R6 which is physically an x2 PCI-e external SATA-raid card in a box connected over Thunderbolt which tunnels the PCI-e protocol. Extra bandwidth is just there to provide the needed room of the 8b/10b encoding. Add to that the the Thunderbolt chip is an PCI Express switch which means that you can physically connect and establish a connection to more then one full speed x2 PCI-e 2.0, two x1 PCI-e 2.0 or up to four PCI-e 1.x full speed devices which means they can be bandwidth starved when daisychained. Effective bandwidth for GPU is the 8Gbit/s that PCI-e 2.0 x2 provides. System might see it as x4 device though, but it won't perform like an x2.5 device. Should still fine use for gpu use though, but not so much in the consumer space. Though it would be fast enough for some gaming compared to what you could otherwise achieve in a 11.6-13.3" case. For a 15" notebook just buy a faster discrete gpu as said though.