Enclosure & Hardware Discussions
A Call For Measurements: Isolating the Thunderbolt Effect.
 

A Call For Measurements: Isolating the Thunderbolt Effect.  

 of  6
  RSS

wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

@TheITSage: If i look the overall Intel Thunderbolt implementation on Sky/Kaby Lake:

 

Almost all the pictures i looked at seems to show that the TBE controller is linked to the PCH with 4x pci-e link and receive DP signal from the CPU.

If on your Asus X99, the TBE add-in card do not work on all pci-e port, it could mean that it must use a specific port linked to the PCH pci-e root.

So on overall, really astonished about, i didn't find any major information, i would just simply conclude it is far easier on a desktop motherboard to add I/O through the PCH instead to go mess with the CPU pci-e config. Since on desktop motherboard the TBE controller is mainly used to care display signal rather used like I/O to attach hardware at.

However i was pleased to find at least one motherboard that state from where the alpine ridge controller is fed: Gigabyte's GA-X99-Designare EX

 

Dunno how reliable this picture is, also how reliable usual X99 diagram are, but i would say on old system the TBE controller would be on PCH, on newer it could be integrated to the CPU lanes, but you need to carefully check each pci-e link config, for each high end motherboard integrating an alpine ridge controller. It seems that connecting the alpine ridge to the cpu could be an implementation done by the manufacturer and not basic specifications gave by intel.

The best way to do that would be to mail the usual friendly brand, ask some information about, as you usually do for reviews, it would be the best to get something reliable.

Or even if you want could asses different kind of motherboards using different pci-e wiring setup for the TBE controller.

 

This is the basic picture of a Z270 chipset:

The Gigabyte GA-X99P-SLI should have a fully working TBE without using any optional add-in card, if i got it well.

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


itsage, gtosi and vava726 liked
ReplyQuote
goalque
(@goalque)
Noble Member Admin
Joined: 3 years ago
 

Generally TB3 (up to 22Gbps) bandwidth does not increase external FPS. You have to choose apps/games that are highly bandwidth-intensive and non-CPU intensive.

And this is what I got after Node firmware upgrade:

Ryse: Son of Rome.

Moreover, seems that Apple's Metal API with minimized CPU overhead yields better performance when more bandwidth is available. 

automate-eGPU EFIapple_set_os.efi
--
2018 13" MacBook Pro + Radeon [email protected] + Win10 1809


itsage and gtosi liked
ReplyQuote
gtosi
(@gtosi)
Eminent Member
Joined: 3 years ago
 
Posted by: goalque

 

Generally TB3 (up to 22Gbps) bandwidth does not increase external FPS. You have to choose apps/games that are highly bandwidth-intensive and non-CPU intensive.

And this is what I got after Node firmware upgrade:

Ryse: Son of Rome.

Moreover, seems that Apple's Metal API with minimized CPU overhead yields better performance when more bandwidth is available. 

   

Thanks. Ok so if the game is bandwidth intensive, the bandwidth matters. And for non-bandwidth-intensive games/benchmarkers, should we still see some "performance drop" in egpu compared to dgpu? I guess that is where we come back to the purpose of this thread and the itsage's testing.

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


ReplyQuote
goalque
(@goalque)
Noble Member Admin
Joined: 3 years ago
 

Bandwidth/CPU usage % matters.

Very difficult to find games that are highly bandwidth-intensive with lower CPU usage. For example, Hitman (2016) was not a good candidate to indicate H2D bandwidth impact on a dual core 2016 13” MBP, because the CPU became bottleneck first. It is a highly CPU intensive game.

automate-eGPU EFIapple_set_os.efi
--
2018 13" MacBook Pro + Radeon [email protected] + Win10 1809


ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

Thank you for the additional research and recommendations. I tried all PCIe slots on this Asus Rampage V Extreme. Thunderbolt connection only works through the PCH PCIe slot. The Falcon Ridge controller on this board definitely limits the connection at Thunderbolt 2 speed. I’ve used this Intel Thunderbolt tool to confirm. There are only three X99 motherboards that are certified for Thunderbolt 3: Asus X99 Deluxe II, Gigabyte X99 Designare EX, Gigabyte X99P SLI.

I will try the Gigabyte X99P SLI next. Hopefully this computer can then be a proper Thunderbolt 3 host to do further testings.

Best ultrabooks for eGPU use

eGPU enclosure buying guide


gtosi, enjoy and vava726 liked
ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

Well another assumption i can make, if the full TB3 alpine ridge is present and placed near the CPU slot, it could be that the controller is linked to the CPU lanes, since from TB2 to TB3 you should increase connection speeds, smaller distances are required to get a good output signal. If the controller is near the PCH, i could presumably think that it is linked to the PCH.

Where the controller is linked could impact on performances, if linked to the PCH there would be any configuration difference with usual laptop one, if hooked on CPU you remove the bottleneck of the DMI PCH/CPU connection.

Quoting the Designare EX review:

As our PCI routing diagrams reveal, Gigabyte hooked up Intel's Alpine Ridge controller to four PCIe lanes direct from the CPU. (You can see that chip sitting behind and to the right of the USB-C connector in the image above). That's because this board is Thunderbolt 3 certified, meaning it can deliver 40 Gbps of potential bandwidth to peripheral devices through the USB-C connector. Owners can daisy-chain up to six Thunderbolt devices off that single port, including displays.

 

I think the controller on your Asus X99 and old MB supporting the add-in card, the integrated TBE controller is mainly used to route the DP over an USB-C, where the newer one offer full TB3 compatibility and have been approved by intel.

But you should mail Giga to ask where the Alpine Ridge is hooked, if you see it near the CPU, surely it is hooked on it. I didn't found any info/diagram of the Giga one.

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


itsage liked
ReplyQuote
vava726
(@vava726)
Estimable Member
Joined: 3 years ago
 

http://www.tweaktown.com/reviews/7680/gigabyte-x99p-sli-intel-x99-motherboard-review/index4.html

According to my opinion  the x99p sli is what we are looking for.... the position of the controller is the same of  the nuc

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


itsage liked
ReplyQuote
enjoy
(@enjoy)
Reputable Member
Joined: 3 years ago
 
Posted by: itsage

 

Thank you for the additional research and recommendations. I tried all PCIe slots on this Asus Rampage V Extreme. Thunderbolt connection only works through the PCH PCIe slot. The Falcon Ridge controller on this board is definitely limits the connection at Thunderbolt 2 speed. I've used this Intel Thunderbolt tool to confirm. There are only three X99 motherboards that are certified for Thunderbolt 3: Asus X99 Deluxe II, Gigabyte X99 Designare EX, Gigabyte X99P SLI.

I will try the Gigabyte X99P SLI next. Hopefully this computer can then be a proper Thunderbolt 3 host to do further testings.

   

I think it is great after all, because we will have Thunderbolt2 (16Gbps) and Thunderbolt 3 (32Gbps) tests not only TB3 🙂 But the results will be the same for 980Ti - I'am sure, and @goalque shows that 🙂 

ϟ AKiTiO Thunder2 + EVGA GTX 1060 6GB SC Gaming (macOS Sierra 10.12.4 and Windows 10)
MacBook Pro (Retina, 15-inch, Later 2013) 3.2GHz Quad Core Intel i7-4750HQ / 8 GB 1600 MHz DDR3 / 256GB SSD + 1TB
mini eGPUPCI Express vs. ThunderboltMac CAN gameGaming Laptops vs. MacBook Pro with eGPU


gtosi liked
ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 
Posted by: vava726

http://www.tweaktown.com/reviews/7680/gigabyte-x99p-sli-intel-x99-motherboard-review/index4.html

According to my opinion  the x99p sli is what we are looking for.... the position of the controller is the same of  the nuc

 

   

Not so good, as i said there is 2 way to link the alpine bridge controller:

  • To the CPU, hard way.
  • To the PCH, easy way.

Only with a controller that is linked to the CPU you could claim to get the same performances as pci-e gpu. If the controller is linked to the PCH you are still limited by the DMI bus, that link CPU/PCH, that is itself a 4x pic-e 3.0 32gbps.

On your nuc, if the diagram is right, the alpine ridge is linked to the PCH. Your nuc motherboard is very small compared to an ATX counterpart.

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


ReplyQuote
vava726
(@vava726)
Estimable Member
Joined: 3 years ago
 
Posted by: wimpzilla

 

Posted by: vava726

http://www.tweaktown.com/reviews/7680/gigabyte-x99p-sli-intel-x99-motherboard-review/index4.html

According to my opinion  the x99p sli is what we are looking for.... the position of the controller is the same of  the nuc

 

   

Not so good, as i said there is 2 way to link the alpine bridge controller:

  • To the CPU, hard way.
  • To the PCH, easy way.

Only with a controller that is linked to the CPU you could claim to get the same performances as pci-e gpu. If the controller is linked to the PCH you are still limited by the DMI bus, that link CPU/PCH, that is itself a 4x pic-e 3.0 32gbps.

On your nuc, if the diagram is right, the alpine ridge is linked to the PCH. Your nuc motherboard is very small compared to an ATX counterpart.

   

For skylake, my nuc:
The processor is connected to the chipset by the four-lane DMI 3.0 interface. The DMI 3.0 protocol is an upgrade over the previous generation which used DMI 2.0 – this upgrade boosts the speed from 5.0 GT/s (2GB/sec) to 8.0 GT/s (~3.93GB/sec), essentially upgrading DMI from PCIe 2 to PCIe 3, but requires the motherboard traces between the CPU and chipset to be shorter (7 inches rather than 8 inches) in order to maintain signal speed and integrity.

When an enclosure reaches the 2.1 -2.2 GB/sec (cudaz results) we can be happy and celebrate lol, so I don't see the bootleneck..... 

I don't know what king of dmi the x99p is using but the trace is shorter then 5 inches.   The real problem would be the dmi 2.0 with a minimum expected drop of 0.1-0.2 GB/s

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 
Posted by: vava726

 

Posted by: wimpzilla

 

Posted by: vava726

http://www.tweaktown.com/reviews/7680/gigabyte-x99p-sli-intel-x99-motherboard-review/index4.html

According to my opinion  the x99p sli is what we are looking for.... the position of the controller is the same of  the nuc

 

   

Not so good, as i said there is 2 way to link the alpine bridge controller:

  • To the CPU, hard way.
  • To the PCH, easy way.

Only with a controller that is linked to the CPU you could claim to get the same performances as pci-e gpu. If the controller is linked to the PCH you are still limited by the DMI bus, that link CPU/PCH, that is itself a 4x pic-e 3.0 32gbps.

On your nuc, if the diagram is right, the alpine ridge is linked to the PCH. Your nuc motherboard is very small compared to an ATX counterpart.

   

For skylake, my nuc:
The processor is connected to the chipset by the four-lane DMI 3.0 interface. The DMI 3.0 protocol is an upgrade over the previous generation which used DMI 2.0 – this upgrade boosts the speed from 5.0 GT/s (2GB/sec) to 8.0 GT/s (~3.93GB/sec), essentially upgrading DMI from PCIe 2 to PCIe 3, but requires the motherboard traces between the CPU and chipset to be shorter (7 inches rather than 8 inches) in order to maintain signal speed and integrity.

When an enclosure reaches the 2.1 -2.2 GB/sec (cudaz results) we can be happy and celebrate lol, so I don't see the bootleneck..... 

I don't know what king of dmi the x99p is using but the trace is shorter then 5 inches.   The real problem would be the dmi 2.0 with a minimum expected drop of 0.1-0.2 GB/s

   

It is a bus bottleneck in a wide sense, compared to a basic 4x pci-e CPU lanes, because the DMI is the bus that connect the CPU to the PCH, so all the data that need to travel back/forth from the CPU/PCH pass through there, not only your TBE.

But i agree if intel use a 4x DMI 3.0, it mean it is surely enough, but on desktop MB remember that the gpu use the cpu lanes, so it free resources from your PCH that manege usually the rest of your computer!

You can't compare your cuda result to the DMI bus, cmon lolz, what you get is the bandwidth between the PCH/TBE/GPU, you can't asses like this the weigh of the TBE controller being on the PCH instead of the CPU. These are two distinct things, that why the best would be a TBE on CPU lanes, if TBE on PCH, you will get the same results, imo, than on your laptop without being cpu limited.

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


vava726 liked
ReplyQuote
vava726
(@vava726)
Estimable Member
Joined: 3 years ago
 

My rough simplification only means that at the state of art the benefits of a TB controller on cpu's lane are wasted  by actual generation of enclosures... It would be nice to experiment if there is an effective difference of performance in gaming where the principal load is generated by gpu. I think that a cpu TB controller can express all its power during the rendering process or in massive TB3 storage systems. 

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

The Gigabyte X99 Designare EX may be the best option. I looked at the board layout more and found it has a Thunderbolt 5-pin header (X99P-SLI does not have it). What’s more interesting is the proximity of this header to the CPU. On my Alienware Area 51 R2’s original board and Asus Rampage V Extreme board, the Thunderbolt header is near the bottom right.

Another attractive feature of this X99 Designare EX is the PEX8747 chipset. It enables triple X16 PCIe for SLI/CrossFire for 28 lane CPUs. Would you guys know how this may play in the whole scheme of things as far as bandwidth goes?

Gigabyte X99 Designare EX layout

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

Well both the Giga card are TB3 ready to go, imo.

Effectively the Designare have an additional THB_C header, i think it could be used if you would add another TB port using their add-in card. Dunno if the header placement would be as important as for the TB controller. 

The SLI+ one have the TBE in the same spot near the USB-C, so i hope it is hooked to the CPU like the Designare. I hope and it should, because premium MB = premium features, have the TB3 hooked on the CPU lanes is a small example, imo, also a wise engineering move.

But again i would possibly assume, being the Alpine Ridge already present hooked to the CPU, if true, the second add-in card could be used like on the Asus MB, with dedicated PCH pci-e lanes.

Unfortunately for you the X99 platform is old, really old even being the top Intel tier, compared to newer chipset like the Z170/270. You could wait and invest into newer MB coming out in a couple of month with the newest high end PCH X299 if i remember well. My advice is to get the newest X99 MB as possible, with the TB3 controller on CPU.

It seems only recently intel allowed full TB controller capabilities on desktop motherboard, i find out that TB3 firmware updates unlocked features on desktop MB controllers, hopefully implementing the TBE winthin the USB-C, again smart choice!

About the pci-e lanes, with 28 lanes you can get only a 3 way SLI under 3*8X pci-e or basic SLI under 1*16X/1*8X pci-e.

With 48 lanes you get a miraculous 4 way SLI  under 4*8X pci-e or 3 way SLI under 1*16X/2*8X or a basic SLI under 2*16X pci-e.

The PLX chip just route the pci-e lanes depending if used or not, by splitting the 16X under 2*8X or 1*8X and 2*4X pci-e.

Obviously if you didn't use any M.2/U.2 slot hooked on the PLX, what would be sad since no RAID. If you use the M.2/U.2 slots under RAID configuration, you need to steep down to 3 way SLI max even with 48 lanes

But i think you noticed unfortunately the MB does not allow you to get a 4 way SLI with 28 lanes, the PLX can't/wouldn't split the 8X into 2*4X.

In few word just consider the total lanes of a CPU, we packed them into 16/8/4x to be easier to deal with i think, but the total amount of bandwidth the CPU can manage is 48 1X lanes. Also for the CPU it does not matter how they are split, i think, because at the end all the lanes fall on the same area on the CPU.

 

See here for more exemples on MSI boards.

https://forum-en.msi.com/index.php?action=dlattach;topic=251355.0;attach=20943;image

Edit: spellcheck! ^^

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


itsage liked
ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

I installed the Gigabyte Designare EX board tonight. Here’s the HWInfo64 screen capture. I’m not sure how else I could go about determining the Thunderbolt controller attachment to CPU.

Here’s Intel Thunderbolt Software showing “External GPU Support: Yes” on this motherboard. I’m currently encountering error 12 with this GTX 980 Ti eGPU. It’s pretty surreal to relive all of these Thunderbolt 3 eGPU findings and issues on a desktop. 😀

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined: 3 years ago
 

itsage, the TB3 controller is on the 0:1.1 PCIe port. It’s using a faster TB3-CPU link as is found on 15″ MBPs only.

For the error 12, proceed to:

i. Download and install Windows Binary Tools (WBT), extracted to c:\dsdt directory

ii. Dump your ACPI tables to disk files (dsdt.asl and dsdt.dat) with these commands at Command Prompt (run as administrator). 

c: & cd \dsdt
acpidump -b -z

iii. Upload your c:\dsdt folder so can create a DSDT override for loading as a registry override.

eGPU Setup 1.35    •    eGPU Port Bandwidth Reference Table

15" Dell Precision 7510 (Q M1000M) (6th,4C,H) + GTX 1080 Ti @ 32Gbps-M2 (ADT-Link R43SG) + macOS 10.13.6 & Win10


itsage liked
ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

@Nando Thank you for the confirmation it’s connecting to the CPU similar to the late-2016 MBP 15″. I found these settings in the BIOS. “Thunderbolt Gfx Support” option communicates to Intel Thunderbolt Software whether the system is eGPU certified. I had more time this morning and was able to locate the controllers on this motherboard. USB-C controller is Texas Instrument TPS65892. Thunderbolt 3 controller is Intel DSL6540.

There are plenty of changes from when this motherboard was released last summer to now in terms of Thunderbolt firmwares. Even though I had all these settings right, the AKiTiO Node wouldn’t work with this motherboard initially. The TB3/USB-C port behaved like it was only a USB-C connection. I had to update the Thunderbolt firmware on this X99 Designare EX and all was well.

 

Best ultrabooks for eGPU use

eGPU enclosure buying guide


vava726 and enjoy liked
ReplyQuote
vava726
(@vava726)
Estimable Member
Joined: 3 years ago
 
We stay tuned. 

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

Hi @Nando. Here’s the dsdt folder.

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined: 3 years ago
 

@itsage, pls give this a go:

1. Extract your DSDT override the c:\dsdt folder:

https://egpu.io/nando4-shared/DSDT_overrides/Gigabyte_Designare_EX_theitsage.zip

2. Right-click  c:\dsdt\load.bat and Run As Administrator to load it to the registry override and turn test signing mode on.

3. If need to back these out,  Right-click  c:\dsdt\unload.bat and Run As Administrator

eGPU Setup 1.35    •    eGPU Port Bandwidth Reference Table

15" Dell Precision 7510 (Q M1000M) (6th,4C,H) + GTX 1080 Ti @ 32Gbps-M2 (ADT-Link R43SG) + macOS 10.13.6 & Win10


itsage liked
ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

I performed the DSDT Override and confirmed the system has Large Memory. However, the eGPU still experience error 12. I tried booting with the eGPU plugged in and encountered ACPI_BIOS_ERROR. The system no longer boots into Windows. 

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
ed_co
(@ed_co)
Reputable Member
Joined: 3 years ago

ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 
Posted by: ed_co 

Any news here??

I was taking care of some personal matters and have not had time for this project. I will be back full speed this week.

Best ultrabooks for eGPU use

eGPU enclosure buying guide


gtosi, ed_co, enjoy and 1 people liked
ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

Error 12 persists after several different attempts. Coincidentally, @BlueRaider got an X99 system running with Thunderbolt 3. I tried a couple of suggestions mentioned in his thread but no progress. I may have to exchange the X99 Designaire EX for X99P-SLI motherboard.

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

Well Linus latest video is about it. The topic get hot, it seems.

LinusThechTips

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


nando4 liked
ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined: 3 years ago
 

Linus’ Alienware Graphics Amplifier (x4 3.0  CPU) versus Razer Core (TB3) comparison

 

@wimpzilla, I’ve transcribed Linus’ video from 4:39 onwards below.  He’s identified that Thunderbolt 3 doesn’t deliver expected performance and given some reasons why.  There are some holes in his analysis which then make his conclusion questionable.

 

* he notes the AGA-CPU versus TB3-PCH connectivity difference. Though doesn’t identify if the Alienware 13’s TB3 controller is connected to the CPU or PCH when doing a comparison of AGA vs TB3 performance on that platform

 

* mistakenly believes Thunderbolt 3 is a 32Gbps bandwidth link. It’s 22Gbps.

 

.. the AGA has the potential to absolutely spank the Core delivering buttery visuals compared to the questionably playable performance on Razer’s solution. And as a troubleshooting step then [they connect using the internal LCD rather than external LCD and find the there is now an even greater performance difference favoring the AGA]

So what gives? Well, we could only think of two differences between these solutions that could affect performance:

1. Thermals. Either the CPU in the laptop or the graphics card in the enclosure.
2. The connection to the graphics card

[ does AIDA64 monitoring concluding thermals are not a problem ]

Let’s move onto the connection then. The theoretical bandwidth of the two should be equal. In GPU-Z, both have the same number of PCIe and the same reported 40Gbps bandwidth.

Is there some ways we could pin this on Thunderbolt 3? What if the Gigabit ethernet on the Core is sucking up resources? We used disabling device manager and .. it wasn’t very effective. 

So it was time to go splunking deep into the depths of the specs searching for a single difference between the two. And finally we found one, in the dark crevices of the device manager.

On the AGA, the GPU gets connected directly to the PCIe controller on the CPU. So the lanes that it gets are solely devoted to the graphics.

The Razer Core on the otherhand, it gets connected through the motherboard chipset meaning the graphics card has to share, and perhaps most importantly, the DMI interface with things like networking, the USB ports and the biggest bandwidth sucker, storage.

But can we prove once and for all that the lanes are causing the bottleneck?

For a complete Apple-to-Apple comparison, we needed a way to connect the Core to the Alienware 13, something that Dell doesn’t want you to do inspite of the 13’s higher end 4-lane implementation of the Thunderbolt 3 standard. The good news is that if you disable enough of the Alienware services, it will begrungingly let you use a Razer Core and in-game the Alienware 13 connected to the Core recorded results within our margin-of-error.

This proving then that the bottleneck wasn’t in the laptop but in the [Thunderbolt 3] connection itself.

 

 

eGPU Setup 1.35    •    eGPU Port Bandwidth Reference Table

15" Dell Precision 7510 (Q M1000M) (6th,4C,H) + GTX 1080 Ti @ 32Gbps-M2 (ADT-Link R43SG) + macOS 10.13.6 & Win10


ed_co and itsage liked
ReplyQuote
itsage
(@itsage)
Famed Member Admin
Joined: 3 years ago
 

@Sky11 Sorry I accidentally removed your post because it appeared as a double-post. Please repost when you have a chance.

I will try to do a comparison this weekend between AGA vs. Sonnet Breakaway/AKiTiO Node.

Best ultrabooks for eGPU use

eGPU enclosure buying guide


ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

Well i agree with you @Nando4, on this review Linus group could had done a better job showing the in game test, i have the feeling too something is missing.

Generally they spend some time to test and setup to be sure, to not get caught misinformed. It's like they got something but they couldn't/didn't spent much time over it.

At the end it came out  from the video, that the average 97 percentile LOW FPS, is almost the same on both laptops while gaming. It fit with the good overall result offered by the TB3 we notice.

Then he maybe spend some more time where the games were really demanding bandwidth side, I.E, using ultra high textures, using a lot of CPU resources, etc.

I was surprised too, but i think the AGA have no TB3 controller at all, it is just the box connected to the CPU pci-e lanes like our loved PE4C/EXP!

That's explain the price drop, on this one my admiration goes to Dell, i can check the schematics if i found them to be sure, but pretty sure Dell implemented a big EXP/PE4C.

That,'s why you need to reboot the laptop, like we do with the GDC.

Smart move again, 1st time from long i get astonished about something and found out really nice from a company to do so.

 

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


enjoy and vava726 liked
ReplyQuote
ed_co
(@ed_co)
Reputable Member
Joined: 3 years ago
 

@nando4, I think that he mistakenly believes that is 40Gbps, no 32Gbps...

Which I don't  understand, is why intel is limiting to 22Gbps instead of having 32Gbps.

At the very beginning, we all thought that was 40Gpbs, but then we realize that where 32Gbps + 8Gbps, and finally 22Gpbs...

There is any reason?

Maybe changing the firmware to 32 could be an option in order to get the full bandwidth?

mid-2017 Macbook Pro RP560 + MSI GTX 1080Ti Gaming [email protected] (Mantiz Venus) + macOS10.13 & Win10


ReplyQuote
enjoy
(@enjoy)
Reputable Member
Joined: 3 years ago
 

more and more results for Thunderbolt 3 fail to reach 32Gbps are great for Thunderbolt 2 users 🙂 ( )

@ed_co - @nando4 explains that many times it's because only MacBook Pro have direct connection - Thunderbolt 3 port to the CPU (only on the left ports of 15'' MBP 2016 TB)

15" Macbook Pro is highest performance TB3 eGPU notebook to date

@Lacin, all TB3 PC notebooks so far have their TB3 controller attached to the PCH, a x4 3.0 link which shares traffic with other high throughput PCIe devices like wifi, lan and noteably PCIe SSDs. The sole exception being the 15" Macbook Pro which attaches the TB3 controller to the CPU. This is shown pictorially at:

https://egpu.io/macos-pascal-drivers-gtx-1080-ti-2016-macbook-pro/

If wanting the highest performance TB3 platform then that should be at the top of your list. One eGPU.io user seeing a significant FPS increase in changing from a TB3-PCH PC notebook to a TB3-CPU 15" MBP:

https://egpu.io/forums/builds/mbp-2016-razer-core-titan-x-pascal-better-than-blade-and-gs40/

The 15" Macbook Pro also commands a premium price.

ϟ AKiTiO Thunder2 + EVGA GTX 1060 6GB SC Gaming (macOS Sierra 10.12.4 and Windows 10)
MacBook Pro (Retina, 15-inch, Later 2013) 3.2GHz Quad Core Intel i7-4750HQ / 8 GB 1600 MHz DDR3 / 256GB SSD + 1TB
mini eGPUPCI Express vs. ThunderboltMac CAN gameGaming Laptops vs. MacBook Pro with eGPU


ReplyQuote
ed_co
(@ed_co)
Reputable Member
Joined: 3 years ago
 
Posted by: enjoy

@ed_co - @nando4 explains that many times it's because only MacBook Pro have direct connection - Thunderbolt 3 port to the CPU (only on the left ports of 15'' MBP 2016 TB)

But I am talking about in the best case scenario (mbp 2016 15" left side), not in others... and is confirmed that is 22Gbps, not because of the direct connections:

Posted by: nando4

Thunderbolt separates it's bandwidth into DATA (up to 22Gbps) and DISPLAY. The onboard USB/SATA/Ether runs off the Thunderbolt3 USB controller which is part of the DATA bandwidth allocation, shared with the eGPU.

If you want full 22Gbps bandwidth for eGPU use as well as USB/Ether/SATA then it requires:

* a separate dock to host USB/Ether/SATA  off it's own USB3/TB3 port
*a TB3 enclosure that has a switched USB controller like the Asus XG Station 2. There you attach a USB-B cable to your system USB 3.0 port to activate the USB switch.

 

Which I don't understand, if the DISPLAY is not being used, why not use the full bandwidth for DATA... and get the full 32Gbps... or even the full 40Gbps, if we were not using the USB-SATA-ETH...

mid-2017 Macbook Pro RP560 + MSI GTX 1080Ti Gaming [email protected] (Mantiz Venus) + macOS10.13 & Win10


ReplyQuote
enjoy
(@enjoy)
Reputable Member
Joined: 3 years ago
 

Thunderbolt 3 bandwidth allocation in various use cases.  Video will always take priority over data and will get the full bandwidth allocation it needs. The remaining bandwidth is allocated to data transfers from the PC to a dock or display. The inbound data back to the PC is at full Thunderbolt 3 data bandwidth of 22 Gbps.

ϟ AKiTiO Thunder2 + EVGA GTX 1060 6GB SC Gaming (macOS Sierra 10.12.4 and Windows 10)
MacBook Pro (Retina, 15-inch, Later 2013) 3.2GHz Quad Core Intel i7-4750HQ / 8 GB 1600 MHz DDR3 / 256GB SSD + 1TB
mini eGPUPCI Express vs. ThunderboltMac CAN gameGaming Laptops vs. MacBook Pro with eGPU


ReplyQuote
nando4
(@nando4)
Noble Member Admin
Joined: 3 years ago
 
Posted by: ed_co

@nando4, I think that he mistakenly believes that is 40Gbps, no 32Gbps…

Which I don’t  understand, is why intel is limiting to 22Gbps instead of having 32Gbps.

At the very beginning, we all thought that was 40Gpbs, but then we realize that where 32Gbps + 8Gbps, and finally 22Gpbs…

There is any reason?

Maybe changing the firmware to 32 could be an option in order to get the full bandwidth?

TB3 is marketted as the one port that can do it all – USB 3.1, DP or Thunderbolt.

Only reason I can think Intel limited TB3 (PCIe) to 22Gbps is because these TB3 controllers are dual-port. Intel guarantees 10Gbps when it’s used as a USB 3.1 port. As there is 32Gbps total traffic to the x4 3.0 linked TB3 controller, when we deduct this 10Gbps reserved bandwidth it leaves 22Gbps for PCIe data on the second TB3 port.

Of course, this doesn’t help eGPU users who want their TB3 port to be dedicated fully to eGPU traffic. Nor does it appear to take into consideration if the system only has 1 TB3 port so has no need to reserve 10Gbps for USB 3.1.

Maybe Intel will release a 32Gbps TB4? John_Mantiz indicated that the 22Gbps limit is set across the transmission link. Setting the enclosure firmware to 32Gbps would not release the desired +10Gbps bandwidth.

 

Posted by: enjoy

more and more results for Thunderbolt 3 fail to reach 32Gbps are great for Thunderbolt 2 users 🙂 (devil emoticon)

@ed_co – @nando4 explains that many times it’s because only MacBook Pro have direct connection – Thunderbolt 3 port to the CPU (only on the left ports of 15” MBP 2016 TB)

@enjoy, the 15″ MBP uses a dual-port TB3 controller, one per side (4:0.0 and 7a:0.0). Each of these dual-port TB3 controllers connects at x4 3.0 to a dedicated PCIe port on the CPU (0:1.1 and 0:1.2) respectively. The dGPU is attached to the 0:1.0 CPU PCIe port. See the Macbook 13,3 lspci.txt for confirmation.  It’s the highest performance TB3 architecture on the market.

I’ve elaborated on the differences in bandwidth of TB3-CPU and TB3-PCH architectures at:
https://egpu.io/setup-guide-external-graphics-card-mac/#cpu-vs-pch

eGPU Setup 1.35    •    eGPU Port Bandwidth Reference Table

15" Dell Precision 7510 (Q M1000M) (6th,4C,H) + GTX 1080 Ti @ 32Gbps-M2 (ADT-Link R43SG) + macOS 10.13.6 & Win10


enjoy liked
ReplyQuote
vava726
(@vava726)
Estimable Member
Joined: 3 years ago
 
Posted by: ed_co

Which I don't understand, if the DISPLAY is not being used, why not use the full bandwidth for DATA... and get the full 32Gbps... or even the full 40Gbps, if we were not using the USB-SATA-ETH...

Because it can reach 40 Gbps -> 5GB/s only when the connected device is a display, when you need to control a gpu you send and recive data so tb3 can only reach 22 Gbps as described by the picture, 32 Gbps are a chimera

Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts


ReplyQuote
enjoy
(@enjoy)
Reputable Member
Joined: 3 years ago
 

so we can finish with that conclusion:

  • Thunderbolt 1 - full 10Gbps
  • Thunderbolt 2 - full 16Gbps
  • Thunderbolt 3 - full 22Gbps

Can we?

ϟ AKiTiO Thunder2 + EVGA GTX 1060 6GB SC Gaming (macOS Sierra 10.12.4 and Windows 10)
MacBook Pro (Retina, 15-inch, Later 2013) 3.2GHz Quad Core Intel i7-4750HQ / 8 GB 1600 MHz DDR3 / 256GB SSD + 1TB
mini eGPUPCI Express vs. ThunderboltMac CAN gameGaming Laptops vs. MacBook Pro with eGPU


ReplyQuote
wimpzilla
(@wimpzilla)
Honorable Member
Joined: 3 years ago
 

Well i could simply suppose that the initial major issues for apple engineers was to find a way to deal with very high definition/quality display so an increasing amount of data.

Since display is one of the major flagship hardware for apple, it would be normal to be sure that the display signal is good enough over data bandwidth.

Aside the eGPU world, i think TB3 bandwidth is plenty enough for other consumer and pro portables devices.

 That's why there was no initial support for the eGPU i think.

2012 13-inch Dell Latitude E6320 + R9 [email protected] (EXP GDC 8.4) + Win10
E=Mc²


ReplyQuote
 of  6