eGPU vendor demonstrates AMD Vega card working in High Sierra on Thunderbolt 3 MacBook Pro...

Posted:
in macOS
External GPU enclosure developer Mantiz posted a Tweet on Thursday of the new AMD Vega PCI-E video card working in its enclosure, connected to a MacBook Pro -- and that likely means that the card will work on Apple's last PCI-E Mac Pro as well in High Sierra.




The Tweet featured two pictures -- one of the new card inside the Mantiz MZ-02 eGPU enclosure, and the second a picture of the System Profiler showing the card inside the system.

Vega RX 64 is on High Sierra Beta 5 pic.twitter.com/kEn6GOT4Yc

-- Mantiz eGPU (@Mantiz_John)
No benchmarks of performance were published by the enclosure developer. However, the card appears to be functioning, and identified as a Vega card by macOS High Sierra's latest beta release.

AMD is on the cusp of shipping a trio of cards based on its Vega architecture. The card is either the RX Vega 56 or RX Vega 64 -- both of which have been slated to make an appearance in the iMac Pro.

The $399 RX Vega 56 has 56 compute units, with the GPU running at 1156MHz at base speed, and 1471MHz in boost. HBM2 memory in the Vega 56 allows for 410GB/sec of memory bandwidth.

The Radeon RX Vega 64 has 64 compute units, and clock speeds of 1247Mhz and with a peak speed of at 1546Mhz, with 484GB/sec of memory bandwidth. The RX Vega 64 retails for $499.

At the time of announcement, it wasn't clear if the cards could be used in an eGPU enclosure, or the last Mac Pro tower with PCI-E slots. However, at present, the Vega PCI-E card in High Sierra seems to work through Thunderbolt 3, and the Mac Pro 5,1 is listed as one of the machines that supports the operating system so it should work fine there as well.

In both iterations of the Vega GPU, half-precision 16-bit calculations (FP16) such as that used for some image and graphic processing, ray tracing, artificial intelligence, and game rendering can be optimized to hit a peak of 22 tflops, exceeding that in the Nvidia GTX 1080. But, the advantages of FP16 for the consumer aren't quite clear, or easily comparable, to Nvidia cards as of yet.

The $379 Mantiz MZ-02 eGPU enclosure has five USB 3.0 type A ports, a Gigabit Ethernet port, and a 2.5-inch mounting plate for a SATA drive in addition to the PCI-E slot for a graphics card. It can accommodate GPU cards demanding up to 375W, and provides a full 87W of charging power back to the connecting computer, assuming it hasn't been adapted to an older version of Thunderbolt.

AppleInsider reviewed the Mantiz MZ-02, and found it to be a solid, stylish choice for users needing eGPU functionality, for not much more than a Thunderbolt 3 dock's retail price.

Comments

  • Reply 1 of 11
    So I get how Mac support Thunderbolt PCI Express enclosures now, but where are these video card drivers coming from? Are we back to vendor-supplied drivers? Weren't they historically pretty buggy?
  • Reply 2 of 11
    Mike WuertheleMike Wuerthele Posts: 6,861administrator
    Eric_WVGG said:
    So I get how Mac support Thunderbolt PCI Express enclosures now, but where are these video card drivers coming from? Are we back to vendor-supplied drivers? Weren't they historically pretty buggy?
    The Nvidia drivers come from Nvidia and are decent, but not fabulous -- particularly with the 10-series cards. 

    The Vega drivers are likely Apple's own, intended for the iMac Pro but usable with the card.

    And to be clear, Apple doesn't support video on Thunderbolt 3 enclosures until the spring of 2018 with High Sierra. The testing I'm doing on the enclosures in Sierra relies on a hack.
    edited August 2017 RacerhomieXwatto_cobra
  • Reply 3 of 11
    I'm surprised you're not using the HS public beta… regardless, thank you for the info. Strange, exciting days ahead…
    watto_cobra
  • Reply 4 of 11
    mdriftmeyermdriftmeyer Posts: 7,503member
    With AMDGPU being public source code, Apple can incorporate that into their stack and also contribute back. It all depends on their contractual obligations. Apple will be making Vega Metal API 2 compliant w/o AMD writing a single piece of code.
    watto_cobraquadra 610
  • Reply 5 of 11
    Mike WuertheleMike Wuerthele Posts: 6,861administrator
    Eric_WVGG said:
    I'm surprised you're not using the HS public beta… regardless, thank you for the info. Strange, exciting days ahead…
    I'm using both. There are no Nvidia drivers for PCI-E cards in High Sierra, and the Nvidia card in the developer's kid has no Sierra driver.
    edited August 2017
  • Reply 6 of 11
    hugorfghugorfg Posts: 1unconfirmed, member
    "AMD is on the cusp of shipping a trio of cards based on its Vega architecture. The card is either the RX Vega 56 or RX Vega 64" - It's a Vega 64. Mantis said so on the picture's description, plus only the 64 will come in this metallic finish. The Vega 56 and 64 (non-Limited) will come in a regular, black plastic shroud finish.
    edited August 2017
  • Reply 7 of 11
    So it'll be cool if a 4,1 (flashed) or 5,1 Mac Pro (maybe even a SIP-disabled 3,1 running HS) can at least recognize this card, but with only PCIe 2.0 x16, it would seem that it will be bottlenecked. I might be wrong, but just like the 3,1  (even though it also has PCIe 2.0 16x) really only fully benefits from a GTX 960 or under before hitting a bottleneck,  if I've read things properly, the 5,1 is also a bottleneck for the 1080ti, so I suspect the same for at least the top Vega. I hope I'm wrong. 

    I'm also seeing lots of PCIe 3.0 cards not properly negotiating in MP 2.0 slots, either downstepping to basic 1.0, or reducing lanes by half.  It is extremely frustrating to be artificially crippled with zero chance that Apple would work with a vendor to correct the issue on their end, let alone issue firmware for 5-7+ year old machines that otherwise remain viable workhorses awaiting a 2018-2019 rethought Mac Pro. 
    cornchip
  • Reply 8 of 11
    tipootipoo Posts: 1,142member
    Frederico said:
    So it'll be cool if a 4,1 (flashed) or 5,1 Mac Pro (maybe even a SIP-disabled 3,1 running HS) can at least recognize this card, but with only PCIe 2.0 x16
    , it would seem that it will be bottlenecked. 

    I think you'd be surprised how little PCI-E bandwidth they actually need. 2.0 x16 is still a lot more bandwidth than Thunderbolt 3 (40Gb/s = 5GB/s, PCI-E 2 x16 is 8GB/s), plus without the added round trip latency of TB. If TB3 is providing 85% of a GPUs performance, I'd expect PCI-E 2.0 won't be a worry for another few high end generations. PCI E 2.0 x16 would be 64Gb/s the way Thunderbolt lists it (I'm sure using Gb instead of GB started out as marketing to sound as good as PCI-E, lol). 


    Particularly as Vega isn't really pushing past 1080TI territory. 


    edited August 2017
  • Reply 9 of 11
    nhtnht Posts: 4,522member
    With AMDGPU being public source code, Apple can incorporate that into their stack and also contribute back. It all depends on their contractual obligations. Apple will be making Vega Metal API 2 compliant w/o AMD writing a single piece of code.
    If you compare the performance of AMD opens source OpenGL Linux driver performance against their closed source driver performance on Windows there's a huge performance hit.  Likewise against nVidia's closed source Linux drivers.  You see parity between AMDs open and proprietary drivers on Linux but both are slow in comparison to nVidia Linux and AMD windows.

    Apple is better off getting AMDs proprietary  driver source as the basis of any MacOS support.  AMD would be stupid to not provide it.
  • Reply 10 of 11
    cornchipcornchip Posts: 1,950member
    Frederico said:
    So it'll be cool if a 4,1 (flashed) or 5,1 Mac Pro (maybe even a SIP-disabled 3,1 running HS) can at least recognize this card

    Would be sweet!

  • Reply 11 of 11
    Mike WuertheleMike Wuerthele Posts: 6,861administrator
    Frederico said:
    So it'll be cool if a 4,1 (flashed) or 5,1 Mac Pro (maybe even a SIP-disabled 3,1 running HS) can at least recognize this card, but with only PCIe 2.0 x16, it would seem that it will be bottlenecked. I might be wrong, but just like the 3,1  (even though it also has PCIe 2.0 16x) really only fully benefits from a GTX 960 or under before hitting a bottleneck,  if I've read things properly, the 5,1 is also a bottleneck for the 1080ti, so I suspect the same for at least the top Vega. I hope I'm wrong. 

    I'm also seeing lots of PCIe 3.0 cards not properly negotiating in MP 2.0 slots, either downstepping to basic 1.0, or reducing lanes by half.  It is extremely frustrating to be artificially crippled with zero chance that Apple would work with a vendor to correct the issue on their end, let alone issue firmware for 5-7+ year old machines that otherwise remain viable workhorses awaiting a 2018-2019 rethought Mac Pro. 
    The AMD RX 560 in the developer's kit works in a 5,1 with High Sierra without strangeness related to PCI-E 2.0 versus 3.0, so I'm not expecting any problems. We'll see.

    Even if Apple was still actively supporting the 5,1 with firmware updates, it wouldn't rectify the PCI-E issue. That's new hardware, not a firmware update.
    edited August 2017
Sign In or Register to comment.