Late 2016 13" MacBook Pro (nTB) + VEGA64@32Gbps-TB3 (Netstor HL23T) + macOS10.12 & Win10 [goalque]  

 

goalque
(@goalque)
Honorable Member Admin
Joined:1 year  ago
Posts: 779
September 10, 2017 6:37 pm  

As we’ve seen, some enclosures are struggling with power requirements of the RX Vega 64. Furthermore, my Akitio Node’s twisted 4-pin cable is dangerously close to the hot backplate. It actually touches the RX Vega’s backplate. I wonder how these enclosures have passed AMD’s R9 390X eGFX thermal testing? I found a common factor: those listed AMD compatible cards do not have thick backplates. In my DIY Thunder2 project, that 4-pin cable had distinguishable damage due to the backplate’s hot touch. Even though the yellows are 22AWG 90C and blacks marked as 80C, the wire plastic started to melt. Therefore, I’ve been very careful thereafter, because RX Vega’s hot spot can reach over 90C temperatures.

Fortunately, HL23T has a bit more room between the GPU and the TB3 controller. I placed a temperature sensor between these two PCIe x16 slots and a LCD screen to the front to keep track of temperatures in my various cooling configurations. I realized that Noctua NF-R8 REDUX-1800 has a self-stabilising oil-pressure bearing that uses an additional magnet. It snaps to the green metal grill of the HL23T, and fans won’t fall even if you turned it upside down. Screwless operation provides also the best acoustics. I ended up doing the following:

 

I upgraded HL23T’s original 300W PSU to Seasonic’s 500W, with 80 PLUS Gold efficiency rating. It wasn’t an easy task. I sacrificed my old NA211TB shells, drilled two side screw threads from the back off, and cut an opening for the PSU power switch (not finalized yet).

 

A “paperclip trick” was also necessary to power on the Seasonic SS-500 L1U.

 

 

In these tests, I used Netstor’s eGFX firmware (not publicly released) where the upper TB3 port should guarantee ~22Gbps H2D performance. Not sure if this is necessary because the factory firmware gave the same good CUDA-Z reading on Win10, regardless of the TB3 port. It likely depends on the hosting computer as well. I’ve not yet tested factory firmware in AIDA64 Extreme GPGPU benchmark but here first the eGFX firmware results.

EDIT: The factory firmware is recommended. Take a look at the new benchmarks:
https://egpu.io/forums/implementation-guides/late-2016-13-macbook-pro-ntb-vega6432gbps-tb3-netstor-hl23t-macos-win10-goalque/#post-20080


Netstor’s 40Gbps 1m cable Belkin’s 40Gbps 0.5m cable

 

What matters is the length of the TB3 cable. I tested with the Netstor’s 40Gbps 1m cable and compared to Belkin’s 40Gbps 0.5m cable and the latter performed noticably better in AIDA64 memory read/write.

These benchmarks are conducted with the Netstor’s 1m cable only:

 

 

Will let you know if there is significant performance impact with the shorter cable in real gaming and computing tasks.

Edited: 3 months  ago

automate-eGPU.shapple_set_os.efi
--
late-2016 13" Macbook Pro nTB + Vega64@32Gbps-TB3 (Netstor HL23T) + macOS & Win10
late-2016 13" Macbook Pro nTB + GTX980/RX580@32Gbps-TB3 (Netstor HL23T) + macOS10.13 & Win10


nando4 and theitsage liked
ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined:1 year  ago
Posts: 1580
September 10, 2017 7:36 pm  

goalque, thank you for your implementation guide. Q: did you obtain the Seasonic 1U PSU from your old Powercolor Devilbox?

Also, your AIDA64 memory write (equiv H2D) of  2260MB/s is less than what @jefniro‘s 2507MB/s here  is. He used an 2016 13″ MBP + AKiTiO Node + 0.5m 40Gbps cable + RX480. Maybe some info you can provide to Netstor for tweaking of their firmware?

eGPU Port Bandwidth Reference TableeGPU Setup 1.35


ReplyQuote
goalque
(@goalque)
Honorable Member Admin
Joined:1 year  ago
Posts: 779
September 10, 2017 9:29 pm  
Posted by: nando4

goalque, thank you for your implementation guide. Q: did you obtain the Seasonic 1U PSU from your old Powercolor Devilbox?

 Yes. It was the hardest part.

Also, your AIDA64 memory write (equiv H2D) of  2260MB/s is less than what @jefniro‘s 2507MB/s here  is. He used an 2016 13″ MBP + AKiTiO Node + 0.5m 40Gbps cable + RX480. Maybe some info you can provide to Netstor for tweaking of their firmware?

Interesting. I would suggest repeating the test because my Node & Akitio 0.5m cable did not perform as well. I tested both the RX 480 and RX Vega 64.

I’ve observed that the first run from the cold boot sometimes gives nearly theoretical max read performance [2697MB/s] in CrystalDiskMark with the Samsung NVMe SSD 960 PRO and NP631N installed on the HL23T:

On the second run, the result can suddenly drop:

Edited: 5 months  ago

automate-eGPU.shapple_set_os.efi
--
late-2016 13" Macbook Pro nTB + Vega64@32Gbps-TB3 (Netstor HL23T) + macOS & Win10
late-2016 13" Macbook Pro nTB + GTX980/RX580@32Gbps-TB3 (Netstor HL23T) + macOS10.13 & Win10


nando4 liked
ReplyQuote
goalque
(@goalque)
Honorable Member Admin
Joined:1 year  ago
Posts: 779
September 11, 2017 11:22 pm  

The fan configuration represented in the video might be the quietest but regarding GPU temperatures, I think it was not the best idea.

One 8cm fan that points directly towards the GPU blower fan (as Netstor built), yields GPU temperatures of ~75C in Valley benchmarking, after 10 minutes. This seems to be on a par with Tom’s Hardware gaming loop (air, balanced), in GPU open case (or a bit cooler).

http://www.tomshardware.com/reviews/amd-radeon-rx-vega-64,5173-18.html

Edited: 5 months  ago

automate-eGPU.shapple_set_os.efi
--
late-2016 13" Macbook Pro nTB + Vega64@32Gbps-TB3 (Netstor HL23T) + macOS & Win10
late-2016 13" Macbook Pro nTB + GTX980/RX580@32Gbps-TB3 (Netstor HL23T) + macOS10.13 & Win10


ReplyQuote
goalque
(@goalque)
Honorable Member Admin
Joined:1 year  ago
Posts: 779
September 17, 2017 12:03 pm  

With the factory macOS certified (dual TB3 port firmware), AIDA64 GPGPU benchmark displayed improved H2D numbers:

 

1m cable 0.5m cable

 

 

The cable length did not make much difference. However, it did in macOS:

 

1m cable 0.5m cable

Device: AMD Radeon HD Vega10 XT Prototype Compute Engine

    Driver version  : 1.2 (Aug 24 2017 22:22:19) (Macintosh)

    Compute units   : 64

    Clock frequency : 1630 MHz

 

    Global memory bandwidth (GBPS)

      float   : 340.48

      float2  : 365.44

      float4  : 372.66

      float8  : 237.04

      float16 : 111.97

 

    Single-precision compute (GFLOPS)

      float   : 4896.93

      float2  : 4818.56

      float4  : 4763.06

      float8  : 4740.79

      float16 : 4693.50

 

    Transfer bandwidth (GBPS)

      enqueueWriteBuffer         : 2.13

      enqueueReadBuffer          : 2.69

Device: AMD Radeon HD Vega10 XT Prototype Compute Engine

    Driver version  : 1.2 (Aug 24 2017 22:22:19) (Macintosh)

    Compute units   : 64

    Clock frequency : 1630 MHz

 

    Global memory bandwidth (GBPS)

      float   : 334.26

      float2  : 364.00

      float4  : 369.92

      float8  : 233.90

      float16 : 111.79

 

    Single-precision compute (GFLOPS)

      float   : 4893.35

      float2  : 4806.80

      float4  : 4756.72

      float8  : 4730.31

      float16 : 4689.05

 

    Transfer bandwidth (GBPS)

      enqueueWriteBuffer         : 2.40

      enqueueReadBuffer          : 2.81

 

By using a shorter 0.5m cable, we can observe that H2D(enqueueWriteBuffer) increased by 270 MB/s and D2H(enqueueReadBuffer) speed increased by 120 MB/s. The performance difference was clearly visible in bandwidth-sensitive Metal API:

 

1m cable 0.5m cable

 

These results were repeatable.

 

HL23T & undervolted RX Vega 64 Fire Strike 1.1 result was quite equal to Leoben’s Mantiz Venus, even though he has a bit better CPU at 28W TDP.

Graphics score: Mantiz Venus [19 662 (undervolted) / 18 729] vs HL23T [19 708 (undervolted) / 19 224 (balanced)].

Edited: 5 months  ago

automate-eGPU.shapple_set_os.efi
--
late-2016 13" Macbook Pro nTB + Vega64@32Gbps-TB3 (Netstor HL23T) + macOS & Win10
late-2016 13" Macbook Pro nTB + GTX980/RX580@32Gbps-TB3 (Netstor HL23T) + macOS10.13 & Win10


theitsage liked
ReplyQuote
caogui
(@caogui)
New Member
Joined:2 months  ago
Posts: 1
December 15, 2017 12:24 am  

anyone who can tell me this font name? looks nice. Thanks


ReplyQuote
chinsteig
(@chinsteig)
Eminent Member
Joined:4 months  ago
Posts: 21
December 19, 2017 7:09 pm  
Posted by: caogui

anyone who can tell me this font name? looks nice. Thanks

Off topic, but the font is something called Menlo.

Mid 2012 15" MacBook Pro > Thunder2 > GTX 1070 > 34" 2560x1080 Ultrawide
Mid 2017 15" MacBook Pro > Thunder2 > GTX 1070 > 34" 2560x1080 Ultrawide


ReplyQuote
theitsage
(@itsage)
Noble Member Admin
Joined:1 year  ago
Posts: 2017
December 19, 2017 7:42 pm  

 @goalque I wonder if this is why Apple went with a rather unusual length of 0.8m for its Thunderbolt 3 cable.

Numerous implementation guides


goalque liked
ReplyQuote
goalque
(@goalque)
Honorable Member Admin
Joined:1 year  ago
Posts: 779
December 19, 2017 9:36 pm  
Posted by: theitsage

 @goalque I wonder if this is why Apple went with a rather unusual length of 0.8m for its Thunderbolt 3 cable.

Very possible. I remember that CUDA-Z showed similar performance loss from using a longer cable.

automate-eGPU.shapple_set_os.efi
--
late-2016 13" Macbook Pro nTB + Vega64@32Gbps-TB3 (Netstor HL23T) + macOS & Win10
late-2016 13" Macbook Pro nTB + GTX980/RX580@32Gbps-TB3 (Netstor HL23T) + macOS10.13 & Win10


theitsage liked
ReplyQuote
  
Working

Please Login or Register