US$300 AKiTiO Node TB3 eGFX box (32Gbps-TB3)

[Splitframe]

On 11/10/2016 at 0:39 AM, DanKnight said:

Benchmarks will be delayed. Been working on other tests ever since Apple released their MBPs. Sorry for the delays.

 

 

Do you know if the Node has any major differences from the Thunder3 aside from
the PSU and the two missing ports, or will it in general perform like a Thunder3 just
in another chassis? Just asking because I could buy a Thunder3 right now, but if
the Node performs considerably better I might wait.

[Imagios]

a bechmark of the two would be nice, I got my Thunder3 today, but have to wait for a GPU, so I will not be able to contribute any time soon, as well still waiting for cables to the DELL brick

 

[ld0891]

Got mine today, hands-on with benchmarks later.

 

Package and size comparison with a 9.7' iPad Pro:

 

 

 

 

Standard Akitio front face, TB3 port and a PSU with a separate switch, which is nice since the PSU fan would continue spinning if you leave the power cable connected.

 

 

 

Internal

 

 

2pin connectors for power LED and front fan, which means the latter would not adjust speed.

After first boot, I decided to pull it out to switch the fan off to eliminate the noise.

 

 

Design flaw, the 6pin connector to power the TB3-PCIe board would conflict with those cards with thick backplates.

You would have to push the card a little bit to properly install the screws.

The card is a MSI GAMING X GTX 1060 6GB.

 

 

Card installed and the space left is considerable, a water-cooling radiator is definitely possible.

 

 

TB3 cable is 50cm long made by Akitio.

I ordered a Belkin 2m 40Gbps cable and it works perfectly.

 

 

The front fan comes with a tool-free mount and it blows out air by default, which is good for those dual or triple fan cards.

 

 

PSU is a customized model and unfortunately, it does not come from any major suppliers.

 

 

PSU fan is too noisy so I just switch it off by unplugging the connector.

IMO fanless would not be a problem for a 400W PSU while delivering 120W.

 

 

Inside PSU, for reference.

 

Edited November 26, 2016 by ld0891

[ld0891]

Some benchmarks, the graphic card used is MSI GAMING X GTX 1060 6GB.

Midtower - i7 6700K/8G DDR4-2400 *2/Crucial M500 240GB SSD

NUC - Intel Skull Canyon/8G DDR4-2133 *2/Samsung SM951 256GB SSD

 

Fire Strike Extreme GPU score:

Midtower @ PCIe 3.0 x16 - 6482/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 6016/92.81%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 5707/88.04%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 3890/60.01%

 

Time Spy GPU score:

Midtower @ PCIe 3.0 x16 - 4193/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 3990/95.16%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 3632/86.62%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 2974/70.93%

 

CUDA-Z output:

NUC with PE4C V4.1

Spoiler

CUDA-Z Report
=============
Version: 0.10.251 64 bit http://cuda-z.sf.net/
OS Version: Windows x86 6.2.9200 
Driver Version: 375.95
Driver Dll Version: 8.0 (6.14.13.7595)
Runtime Dll Version: 6.50

Core Information
----------------
    Name: GeForce GTX 1060 6GB
    Compute Capability: 6.1
    Clock Rate: 1784.5 MHz
    PCI Location: 0:62:0
    Multiprocessors: 10
    Threads Per Multiproc.: 2048
    Warp Size: 32
    Regs Per Block: 65536
    Threads Per Block: 1024
    Threads Dimensions: 1024 x 1024 x 64
    Grid Dimensions: 2147483647 x 65535 x 65535
    Watchdog Enabled: Yes
    Integrated GPU: No
    Concurrent Kernels: Yes
    Compute Mode: Default
    Stream Priorities: Yes

Memory Information
------------------
    Total Global: 6144 MiB
    Bus Width: 192 bits
    Clock Rate: 4004 MHz
    Error Correction: No
    L2 Cache Size: 48 KiB
    Shared Per Block: 48 KiB
    Pitch: 2048 MiB
    Total Constant: 64 KiB
    Texture Alignment: 512 B
    Texture 1D Size: 131072
    Texture 2D Size: 131072 x 65536
    Texture 3D Size: 16384 x 16384 x 16384
    GPU Overlap: Yes
    Map Host Memory: Yes
    Unified Addressing: Yes
    Async Engine: Yes, Bidirectional

Performance Information
-----------------------
Memory Copy
    Host Pinned to Device: 2285.9 MiB/s
    Host Pageable to Device: 1959.17 MiB/s
    Device to Host Pinned: 2758.53 MiB/s
    Device to Host Pageable: 2216.55 MiB/s
    Device to Device: 66.9245 GiB/s
GPU Core Performance
    Single-precision Float: 4859.02 Gflop/s
    Double-precision Float: 153.951 Gflop/s
    64-bit Integer: 344.96 Giop/s
    32-bit Integer: 1600.64 Giop/s
    24-bit Integer: 1222.12 Giop/s

Generated: Sat Nov 26 13:38:06 2016

 

NUC with Akitio Node

Spoiler

CUDA-Z Report
=============
Version: 0.10.251 64 bit http://cuda-z.sf.net/
OS Version: Windows x86 6.2.9200 
Driver Version: 375.95
Driver Dll Version: 8.0 (6.14.13.7595)
Runtime Dll Version: 6.50

Core Information
----------------
    Name: GeForce GTX 1060 6GB
    Compute Capability: 6.1
    Clock Rate: 1784.5 MHz
    PCI Location: 0:9:0
    Multiprocessors: 10
    Threads Per Multiproc.: 2048
    Warp Size: 32
    Regs Per Block: 65536
    Threads Per Block: 1024
    Threads Dimensions: 1024 x 1024 x 64
    Grid Dimensions: 2147483647 x 65535 x 65535
    Watchdog Enabled: Yes
    Integrated GPU: No
    Concurrent Kernels: Yes
    Compute Mode: Default
    Stream Priorities: Yes

Memory Information
------------------
    Total Global: 6144 MiB
    Bus Width: 192 bits
    Clock Rate: 4004 MHz
    Error Correction: No
    L2 Cache Size: 48 KiB
    Shared Per Block: 48 KiB
    Pitch: 2048 MiB
    Total Constant: 64 KiB
    Texture Alignment: 512 B
    Texture 1D Size: 131072
    Texture 2D Size: 131072 x 65536
    Texture 3D Size: 16384 x 16384 x 16384
    GPU Overlap: Yes
    Map Host Memory: Yes
    Unified Addressing: Yes
    Async Engine: Yes, Bidirectional

Performance Information
-----------------------
Memory Copy
    Host Pinned to Device: 1117.65 MiB/s
    Host Pageable to Device: 1078.95 MiB/s
    Device to Host Pinned: 2491.73 MiB/s
    Device to Host Pageable: 2124.37 MiB/s
    Device to Device: 64.6559 GiB/s
GPU Core Performance
    Single-precision Float: 4757.6 Gflop/s
    Double-precision Float: 156.728 Gflop/s
    64-bit Integer: 353.998 Giop/s
    32-bit Integer: 1637.89 Giop/s
    24-bit Integer: 1225.26 Giop/s

Generated: Sat Nov 26 16:48:49 2016

 

[Imagios]

Congrats !!!

 I have the Thunder3 will buy as well a 1060 6GB and will as well hang it on a NUC, so as soon as I get my GPU will make a test and we can compare, I expect to have a lower performance, due to the Thunder3, with 2 ports and one Display port, but the question is by how much we will see. I'm glad your works well.

[menpasav]

Where did you buy the Node from ?

I am surprised by the benchmark. Maybe I do not read the values right but the Node seems to have a greater performance hit over TB3 than the PE4C ! I was expecting similar values.

What do you mean by eGPU and iGPU ?

Edited November 26, 2016 by menpasav

[Casey_S]

Haha, looks like the Node uses the exact PCB as the Thunder3 with an additional power cable. Guess there's not much point in getting one if you already have a Thunder3.

[ld0891]

14 hours ago, menpasav said:

Where did you buy the Node from ?

I am surprised by the benchmark. Maybe I do not read the values right but the Node seems to have a greater performance hit over TB3 than the PE4C ! I was expecting similar values.

What do you mean by eGPU and iGPU ?

 

In China, Akitio started pre-order on Nov 11th, and shipped on Nov 23rd.

TB3 performs worse than PE4C because the latter connect the card directly to the PCIe interface,

while the former has to pass through Intel Alpine Ridge thunderbolt controller.

This situation will be better since Intel is going to integrate thunbderbolt 3 directly in the CPU instead of using a separate chip.

 

eGPU means the monitor is connected to the GTX 1060 in Node, while iGPU means it is connected to the miniDP port on NUC.

It's like using internal screen or external monitor while using laptops.

[Imagios]

Fire Strike Extreme GPU score:

Midtower @ PCIe 3.0 x16 - 6482/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 6016/92.81%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 5707/88.04%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 3890/60.01%

 

wow a loss of up to 40% of power over TB3  that is terrible, and I expect since its not in the CPU directly but its in trough the  Intel Alpine Ridge thunderbolt controller that this will mean lattencies, so from what you show, eGPU is quile below GPU. 

[ld0891]

1 hour ago, Imagios said:

Fire Strike Extreme GPU score:

Midtower @ PCIe 3.0 x16 - 6482/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 6016/92.81%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 5707/88.04%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 3890/60.01%

 

wow a loss of up to 40% of power over TB3  that is terrible, and I expect since its not in the CPU directly but its in trough the  Intel Alpine Ridge thunderbolt controller that this will mean lattencies, so from what you show, eGPU is quile below GPU. 

 

The 40% loss happens when I connect the monitor to the miniDP port on the NUC, so it's normal to have much worse performance than output by GTX 1060 directly.

The actual loss over TB3 is 10% to 15%.

[Imagios]

its unfortunate that you have to switch cables then, but I assume no one is much carrying the cases around any way  

[Splitframe]

23 hours ago, ld0891 said:

Fire Strike Extreme GPU score:

Midtower @ PCIe 3.0 x16 - 6482/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 6016/92.81%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 5707/88.04%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 3890/60.01%

 

Time Spy GPU score:

Midtower @ PCIe 3.0 x16 - 4193/100%

NUC with PE4C V4.1 @ PCIe 3.0 x4 - 3990/95.16%

NUC with Akitio Node output by eGPU @ PCIe 3.0 x4 - 3632/86.62%

NUC with Akitio Node output by iGPU @ PCIe 3.0 x4 - 2974/70.93%

 

 

Are these the overall scores or just the "graphics score" part of the benchmark?
You wrote GPU score, but just to make sure.

 

Also, huge thanks for the values, my Thunder3 comes on Monday and I'll post
benchmarks with my XPS 15. But if I look at people with a Thunder3 case, it seems
like the performance between the Node and the Thunder3 are is the same.

Edited November 27, 2016 by Splitframe

[Kuinox]

Hi !

13 hours ago, ld0891 said:

In China, Akitio started pre-order on Nov 11th, and shipped on Nov 23rd.

TB3 performs worse than PE4C because the latter connect the card directly to the PCIe interface,

while the former has to pass through Intel Alpine Ridge thunderbolt controller.

This situation will be better since Intel is going to integrate thunbderbolt 3 directly in the CPU instead of using a separate chip.

 

eGPU means the monitor is connected to the GTX 1060 in Node, while iGPU means it is connected to the miniDP port on NUC.

It's like using internal screen or external monitor while using laptops.

Does this mean i should wait before buying all the stuff to make an eGPU ?

I need a good GPU when i'm at home and a laptop for the school...

[Imagios]

To be honest I make my DYI eGPU from Thunder3, but  I would not recommend that to any one, its a lot of money for the Akitio, for the NUC, and in the end you have 2 powerbricks, the Dell brick for the Akitio is humongous, you need custom cables, sure you can buy them as I did, here from members on the forum, but if you are buying a laptop, my personal suggest buy one with thunderbolt and GTX 1060 the difference between desktop and mobile is today negligible, and wait till a really good eGPU will hit the market, with the 1060 you will be totaly covered for years considering you dont want to play 4k, which on a NTB is a overkill, in that case I would rather buy a OLED NTB, but this is just my view, and frankly I have a different used case, build around the NUC skull canyon, but even then today I would buy the Zotac mangus 1070. 

[ld0891]

On 11/27/2016 at 11:05 PM, Splitframe said:

 

 

Are these the overall scores or just the "graphics score" part of the benchmark?
You wrote GPU score, but just to make sure.

 

Also, huge thanks for the values, my Thunder3 comes on Monday and I'll post
benchmarks with my XPS 15. But if I look at people with a Thunder3 case, it seems
like the performance between the Node and the Thunder3 are is the same.

 

Yes they are the "graphics score" parts.

On 11/28/2016 at 5:01 AM, Kuinox said:

Hi !

Does this mean i should wait before buying all the stuff to make an eGPU ?

I need a good GPU when i'm at home and a laptop for the school...

 

I don't know your performance need so I don't have a answer.

Actually the performance hit using external monitor (10% ~ 15%) is acceptable to me.

 

Maybe next year when Intel integrates TB3 controller into their CPU the efficiency would be over 90%, but that's a small increase.

[Ness]

2 hours ago, ld0891 said:

Maybe next year when Intel integrates TB3 controller into their CPU the efficiency would be over 90%, but that's a small increase.

Can you share a link proofing this info? Which one will it be? Kaby Lake or Coffee lake?

[ikir]

Any update on macOS compatibility since the release of TB3 enabler?

[Slovedon]

On 11/29/2016 at 6:00 AM, ld0891 said:

Actually the performance hit using external monitor (10% ~ 15%) is acceptable to me.

Isn't that roughly the performance loss for Thunderbolt 2? I thought TB2 was ~15% for external monitor, but I could be wrong

[bloodhawk]

Here is a performance comparison between eGPU over TB3 VS. the same GPU Over m.2 PCIe x4 

 

 

[ld0891]

On 11/29/2016 at 9:51 PM, Ness said:

Can you share a link proofing this info? Which one will it be? Kaby Lake or Coffee lake?

 

http://beebom.com/intel-kaby-lake-vs-skylake/

 

Quote

Kaby Lake processors will also add native support for Thunderbolt 3.0, which in the case of Skylake processors, could only be supported on motherboards equipped with Alpine Ridge Thunderbolt Controllers.

 

[ld0891]

18 hours ago, Slovedon said:

Isn't that roughly the performance loss for Thunderbolt 2? I thought TB2 was ~15% for external monitor, but I could be wrong

 

Maybe, the bandwidth itself is not a big problem here while using eGPU.

IMO, the real problem lies in the cost of converting protocols between TB and PCIe, as well as the lack of a direct route to CPU instead of a separate controller bypassing the chipset.

 

Based upon this benchmark: https://www.techpowerup.com/reviews/NVIDIA/GTX_980_PCI-Express_Scaling/21.html

PCIe x4 3.0 has a efficiency of over 95% comparing to PCIe 3.0 x16, anything lower would be extra cost then.

[menpasav]

Already December and still available nowhere out of China

Will have to find other alternative for Xmas

[MVC]

It's probably worth waiting for.

 

Razer Core is pretty nice for a little more.(OK, a lot more)

[menpasav]

I received this answer from Akitio (I was asking about availability of the Node in France or Singapore).

They do not trust compatibility with Dell XPS 13 which is strange as this notebook is confirmed working with Razer Core (after driver update) and Acer Graphics Dock !

 

 

From: AKiTiO

Sent: Monday, December 5, 2016 6:11 PM

To: xxxxxxxxxxxxxxx

Subject: [ALL-0XX9S1QEIB] Akitio Node

 

----------Please reply above this line and remove unnecessary text----------

Hello xxxxx.

You have a new message from .
Re: Akitio Node
Message: Hi xxxxxxxxx, The Node is compatible only with the systems listed at https://www.akitio.com/information-center/node-gpu-compatibility. The Dell XPS 13 does not support external GPUs, so the Node will not work with that laptop. Regards, Stefan

When you reply to this message, please send it using the same e-mail address as before. You can also view your ticket and reply online here:
XXXXXXXXX

[DanKnight]

On 11/30/2016 at 11:38 AM, ikir said:

Any update on macOS compatibility since the release of TB3 enabler?

Apple hasn't changed anything on their end and we're not allowed to change anything on our end either... so... no. No good news yet. Still no support for MacOS even with the enabler. :'(

Thank you @ld0891 for the benchmarks. Yeah we know about that little design flaw with the pin connector. I already pointed it out to the team, so I'm hoping that get's fixed in the second wave of production.

 

Reason I haven't provided any benchmarks yet is because the the models that are currently being sold in China/Taiwan are the first wave of production. The second wave of production has a slightly different PSU. Different PSU may mean different voltage outputs. Every little factor counts, so I want to wait before releasing any benchmarks. :'(