|
Kopi Luwak Cafe
Join Date: Dec 2008
Location: YellowBrickRd.AU
Posts: 7,421
|
DIY eGPU experiences
This thread is dedicated to the practical application of these products to attach a desktop external Graphics Processing Unit to a notebook via a expresscard, mPCIe slot, or Thunderbolt port. Ideally, this thread is a place for users to share installation and performance details for others to duplicate or reference.
|  |  |  |  |  | | Gaming videos running on DIY eGPU or ViDock implementations | US$71 PE4L-EC2C (expresscard) or PE4L-PM3N (mPCIe) products | $91 PE4H-EC2C (expresscard) or PE4H-PM3N (mPCIe) products | HP 2560P GTX560Ti @x1.2Opt DIY eGPU impl playing NFS Shift on the internal LCD | Sony VPC-Z11 running 3 displays: 1 via onboard graphics, 2 via eGPU | HP 2510P: [mPCIe] Swap wifi card via underside covers |
If you run Windows 7 exclusively, it is very easy to enable the eGPU without a reboot. Just standby the system, plug in the PE4L, resume, Windows 7 detects the desktop GPU and reverts back to how you have your desktop setup (eg: HD5750 as primary or extended display). Easy.
The performance difference b/w the onboard 4500MHD/HD3000 and an attached desktop eGPU is staggering. The Inspiron 1440 using the x1E tweak saw 3dmark06 go from 866 to 9551. RE5 benchmark went from an unplayable 4.1FPS to a very playabe 60.9FPS. The HP 2560P's x1.2Opt link saw 3dmark06 increase over 4 times to 18340 and RE5 gave an average of 148.3FPS(!!), even allowing the internal LCD to be driven by the eGPU.
Pros - Inexpensive: US$78-shipped PE4L with HD5770+120W=$200, GTS450+120W=$160, GTX460-1GB+430W=$210, GT430+72W=$130, HD5670+72W=$140. Add $66 for a x2 capable kit using a PE4H.
- can be used on any system with an accessible expresscard OR mPCIe slot
- has inexpensive eGPU enclosure examples if you want an enclosure
- 10 times or more increase in video performance
- can have Internal LCD-only setup: NVIDIA Optimus or Ultramon/Chung-Gun method.
- provides additional HDMI video+audio, s-video and VGA out
- provides dual-link DVI connection for 2560x1600 LCDs
- allows multi-monitor connectivity:
- easy, on-demand attachment without a reboot if running Win7
- can be reused after a system upgrade, or shared amongst multiple systems
- see an extra 30-60% in performance when attached to an expresscard 2.0 system, or do a x2 1.0 setup
- provides an additional USB port when connected to expresscard slot
- Time-share your desktop video card b/w your desktop and notebook system.
- perform CUDA/OpenCL processing for applications that support it like Adobe CS5/Premiere [NVIDIA graphics only]
- ability to use physX/hybrid physX & game detail
- components can become a stepping stone to a full deskop system if want higher performance
Cons
- requires a 12V PSU to drive the desktop eGPU
- requires an external LCD/CRT/TV to connect to display at highest FPS
- may require booting DIY eGPU Setup 1.x from a disk image or USB stick to overcome error 12 issues
- restricted bandwidth: a x1/x2/x4 link represents 1/16, 1/8 or 1/4 the bandwidth of a x16 slot.
|
Spoiler :
If the system has a dGPU then ensure it's switchable with the Intel 4500MHD/HD/HD3000/HD4000 providing primary video if you want to do a higher performance Optimus x1.Opt setup with allows accelerated graphics to the internal LCD as well as an external LCD.
Pretty much any notebook can be eGPU compatible if it's run with 2GB of RAM. 95% are OK with 3GB of RAM. It's only if it runs 4GB or more that there needs to be particular attention paid to free pci-e configuration space to host it, which can be checked by:
(Update 3-14-2012) A DSDT override extending the root bridge into 36-bit can allow ALL systems with a 64-bit CPU and chipset (anything after 2007) to be eGPU compatible when used with Windows 7 and the latest iGPU/dGPU/eGPU drivers with up to 48GB of RAM installed. So the DSDT override solution eliminates the restrictions listed below.
- if the system has a switchable iGPU+ATIAMD dGPU or switchable iGPU+NVidia Optimus dGPU configuration then can use Setup 1.x to disable the dGPU component to allow a eGPU to use it's pci-e space. This also works with all Intel chipset systems with NVidia or ATiAMD switchable graphics with the exception that Series-6 AMD switchable graphics hasn't been confirmed to be allow such disabling as explained.
- if the system has only a dGPU or iGPU (not switchable) then check the system's TOLUD+Large Memory to ensure there will be no issues running 4GB+. Check in Device Manager. Set View to "Resources by connection" & expand the Memory tree like shown.
If you have the somewhat rare large memory memory entry then the eGPU can be hosted in 36-bit pci-e space.
If not, as is far more common, then the eGPU will be hosted in 32-bit pci-e space where you'll require TOLUD being 3.25GB or less. Identify this by looking for the first PCI bus entry with a 8 digit memory address. Compare it with table below (thanks Kizwan!):
| First PCI bus with 8 digits memory address | TOLUD | E0000000 | 3.5GB | | D0000000 | 3.25GB | | C0000000 | 3GB | | B0000000 | 2.75GB | | A0000000 | 2.50GB |
As long as TOLUD is equal or lower than 0xD0000000 (3.25GB) your notebook will have enough 32-bit PCI space for a eGPU to work with 4GB or more of system RAM. See details. Some examples:
There is some exception to this rule. We've found some systems have bios allocate systemboard devices that sit right in the middle of a "free" 32-bit 256MB pci-e block that cannot be relocated and prevents an eGPU from that space.
To be absolutely sure an eGPU can be accomodated requires booting Setup 1.x and performing an 32-bit PCI "All" compaction, and seeing if you now have 256MB free shown in the Status window and confirming your system can chainload to Windows. If not, but you do have 128MB contiguous free space shown then consider that NVidia Fermi cards use a fragmented 128MB+64MB+32MB PCI BAR so might still be able to be allocated where ATiAMD cards that require a straight 256MB contiguous block cannot.
If this is a bit too confusing then proceed to boot Setup 1.x, run Save Diags->compact_fails, then Save Diags->Windows_fails following the prompts. Boot back to Windows, mount the Setup 1.x image, archive the v:\diag folder with Winrar/WinZIP, upload to mediafire.com and link here for analysis of your diagnostic output.
3. I don't have an ExpressCard slot. Can I still do use a eGPU?
Spoiler :
Yes, if have an accessible mini PCIe (mPCIe) slot via the underside covers that works with your wifi card, as show here, then could use a PE4L-PM3N or PE4H-PM3N mPCIe setup. The example Dell XPS15 L502, Acer 1810TZ, HP 2510P and Lenovo X61T below all use a PM3N mPCIe-based eGPU.
NOTE: WWAN mPCIe slots with only USB pins connected (pin 36+38) rather than PCIe pins (23, 25, 33, 35) cannot do a eGPU.
USB-to-ExpressCard adapters cannot be used since they uses just the 480Mbps USB pins on an ExpressCard slot.
mini PCI slots can't be used. Can see the difference between mini PCI and mini PCIe here.
4. Can I make this work using just my notebook's LCD display?
Spoiler :
The desktop video card outputs to it's HDMI/VGA/DVI connector. The notebook's LCD display uses a LVDS input on your systemboard with no external jack to connect to it. Meaning then they cannot be physically connected together unless adapt a US$30 ebay kits allowing external input to the internal LCD as discussed, which only quite an advanced user would ever attempt.
Another more elegant solution would use a HDMI input expresscard as described at Play PS3/X360 on laptop screen using new HDMI Input Express Card. . Though current costs of $170 for the item makes it unattractive AND you'd need to then do a mPCIe eGPU implementation.
There are however these less drastic ways of getting your accelerated graphics card to render to the internal LCD, all of which will have lower performance than when running using an external LCD: - NVIDIA Optimus driver provides a transparent internal LCD cloning mode for systems with a 4500MHD/HD/HD3000/HD4000 iGPU primary video when using a NVIDIA GTS4xx/GTX4xx card.
- Use Lucidlogix virtu drivers to provide transparent output using the internal LCD if you have a Sandy Bridge cpu. (timohour)
- Ultramon/Chung Gun method can clone from the desktop eGPU's window to the internal LCD for windowed games/apps.
- USB 3.0 framegrabbers have sufficient bandwidth to capture the image from the desktop eGPU and pipe it back to your notebook's display.
It is not feasible to use USB 2.0 framegrabbers. The desktop video card outputs HDMI/S-Video. To try to clone the output from the desktop video card via a USB frame grabber can be done, but consider say 1280x800x32-bit = 4MB per image. If gaming at 30FPS that's 120MB/s bandwidth required. USB 2.0 is 480Mbps (60MB/s in *best case* scenario.. more like 30MB/s in real-life).
5. (Performance) What affects the performance of a eGPU?6. (Performance) Can my system do a PE4H x2 or x4 link as shown in PCIe Scaling Analysis?
Spoiler :
BPlus' PE4H has made provision to allow compatible 2xmPCIe ports, or mPCIe+expresscard ports to be combined to form a x2 port. To identify what is possible, find your Intel I/O chipset below to identify what ports can be combined to create a x2 or x4 width link
| Intel Chipset | max#
pci-e
ports | x2 link | x4 link | | ICH7M | 4 | p1+2, p3+4* | p1+2+3+4 | | ICH8M/ICH9M/HM55 | 6 | p1+2, p3+4*, p5+6 | Series-5
PM55/HM5x/QM57 | 8 | p1+2, p3+4*, p5+6,p7+p8^ | p1+2+3+4, p5+6+7+8 | Series-6
HM6x, QM6x | 8 | NA | NA |
| 
2510P's port | 
DV2000's port |
*=switches [port1+2] into x2 mode. ^switches [port5+6] into x2 as well. An Intel limitation. NA=feature locked into BIOS flash descriptor
Other combinations to get x2/x4 ports such as p1+3, p2+4 do not work! In example above, the 2510P can't do a x2 link since it has only the 1 physical mPCIe slot. You'll notice on the 2510P port1 is enabled - it's a root port that must be enabled for *any* port to work. 2510P's port1 is not physically connected. The DV2000 has ExpressCard=port1, mPCIe=port3, mPCIe=port4. So the DV2000 can do a x2 link using the 2 mPCIe slots as port3+4.
The best way to identify your port layout is to insert a ExpressCard e-sata card and see what port it reports as being on, then checking too the wifi card. If have multiple mPCIe slots, worth swapping the wifi aiming for a [port1+2], [port3+4], [port5+6] or [port7+8] setup.
NOTE: port1 is the root port and must be enabled for any of the ports to work. On quite a few systems this is just a phantom port with no electrical connections to a mPCIe or expresscard slot.
Where can I find schematics of my systemboard to find port layout? Try laptopdesktopschematic, lqv77, GSM-extreme, notebookschematic and a general google search.
Related: How do I enable port1-port6 if bios has disabled them? Use eGPU Setup's 'iport on'.
Related: How do I setup a x2 1.0 link? Use eGPU Setup's 'iport x2 [port]'.
7. (Performance) What recommended candidate notebooks should I buy for best performance?
Spoiler :
Quick answer - choose any from this list: Example ExpressCard equipped SB systems suitable to attach an eGPU.
Performance will be determined by PCIe bandwidth available, which is ranked from highest->lowest below and and if the notebook has Intel 4500MHD, HD, HD3000, HD4000 iGPU primary graphics so is NVIDIA Optimus capable.
1. PCIe 2.0 series-6 chipsets providing equivalent to x2 1.0 bandwidth. Current ATI cards appears to be running crippled half-duplex links on Intel pci-e ports. Compare HD5750@x2 and HD5770@x2E and techpowerup's HD5870 x1 2.0 results. Series-6 chipsets might deliver x1 2.0 *full duplex* with ATI cards, or 4 times more bandwidth than what ATI cards appear to be getting on series-5 or older chipsets.
2. a Intel-chipset system that allows a x2 link by ganging [port1+2], [port3+4], [port5+6], [port7+8] using mPCIe+expresscard or mPCIe+mPCIe slots. Note: using a x2 port3 affects port1/port2 operation since it sets port1 into x2 mode.
3. series-4 or older Intel-chipset system that allows a x1E link off an odd numbered port for ATI cards OR an AMD/NVIDIA chipset that provides full x1 1.0 performance using an ATI card Intel 4500MHD, HD, HD3000 or HD4000 iGPU allows NVIDIA Optimus superior performance with latest Fermi NVIDIA cards.
8. (Performance) Can I do a SLI/Crossfire implementation combining mPCIe+mPCIe or EC+mPCIe ports?
Spoiler :
timohour found this is not possible here.
9. (Performance) How could manufacturers make eGPU x2 2.0 notebooks?
Spoiler :
UPDATE 2-24-2011 The Apple+Intel Thunderbolt port is a x4 2.0 external PCIe port, though looks like will be uprated to x4 3.0. Perfect for high performance eGPU purposes.
As it appears there is no desire to make an external graphics port standard it can do be done effectively another way. All that is required are 12-18.4" notebooks with: - 2x34mm Intel ExpressCard 2.0 ports electrically configured as either [port1+2] or [port5+6]. Doable with series-6 chipsets.
- switchable HDMI input back into the internal LCD to route the external GPU output to like the Clevo X8100 provides.
In addition provide a bios that - can set [port1+2] or [port5+6] as x2 width.
- performs correct PCI allocation when attach a desktop eGPU to it.
Such a notebook will be perfect for eGPU purposes, offering the equivalent of x4 1.0 in bandwidth. Performance then would be ~95% of when using a desktop eGPU with full x16 2.0 bandwidth.
Only dilemma for manufacturers being such notebooks would become a serious competitor to bulky expensive gaming notebooks that have poor battery life.
10. My system has switchable graphics. Can I still do install a DIY eGPU?
Spoiler :
Yes you can. Setup 1.x's (Video card->Hybrid Graphics.dedicated=off) allows the dedicated graphics component to be completely disabled on Series-5 or older chipsets, freeing up 32-bit resources to host the desktop eGPU. If your primary video card is a Intel 4500MHD, HD, HD3000 or HD4000 then can even do a x1.Opt NVIDIA setup.
11. My Villagetronic ViDock product isn't working. Can you help?
Spoiler :
I've never used a Villagetronic Vidock so am not familiar with their hardware. Consider then opening a support ticket on Villagetronic's support page or asking for help on the Villagetronic community Facebook page. If Villagetronic can't provide a working solution then consider obtaining a refund from them.
12. Can I solder my own adapter to host an eGPU instead of buying one from HIT, BPlus, Redtrontech, Villagetronic?
Spoiler :
Yes you can, especially if you have more time for soldering than money for a pre-manufacturered adapter. See examples how to make one by acyd (Russian) &, sonny_mv, blackett &.
13. What do I need to run the fastest x1.2Opt link shown in the implementations?
Spoiler :
The following ingredients are required for a x1.2Opt link:
1. A Series-6 system (Sandy Bridge) or Series-7 (Ivy Bridge) notebook with an expresscard slot^.
2. A pci-e 2.0 compliant PE4L-EC060A 2.1b
3. A NVidia Fermi or Keplar class video card. A s/h GTX460-1GB OC edition with 336 cores being great bang-per-buck.
4. A power supply to drive your video card. A 203W XBOX360 a good to mod for for < 198W video cards and providing portability. Or any ATX PSU with sufficient power on the 12V rail if don't need portability.
5. Latest NVidia desktop video drivers. They provide the ability to drive the internal LCD upon detecting a Intel 4500MHD, HD, HD3000 or HD4000 iGPU. When the iGPU and x1 link is found, pci-e compression engages gaining anywhere from 20-333% better performance over a x1 2.0 link without compression.
^ a PE4L-PM3N 2.1 will be released in coming weeks so could use a less convenient mPCIe slot instead of the expresscard slot. Useful for systems like a Sony Z2 which don't have an expresscard slot
See Techradar's How to make an external laptop graphics adaptor article for an overview.
1. What parts do I need and where can I buy them?
Spoiler :
    
Above: Fujitsu S710 eGPU implementation - US$85 PE4H-EC2C(comes with SWEX and PCIEMM-030) plus a HD5770+ATX PSU
1. The expresscard or mPCIe to PCIe adapter. See related 2. Can my system use PE4H's x2 1.0.. in FAQ. You'll need the US$92-shipped PE4L 2.1 available on Feb-20-2012 if want the faster pci-e 2.0 performance on a Sandy Bridge notebook. The linked units below are pci-e 1.1 which is the maximum speed available on Series-5 (1st gen i-cores) or older chipsets.
HIT: Harmonic Inversion Technology. A US distributor of BPlus' PE4H-EC2C/PE4L-EC2C with a cheaper base price, cheaper US$7 UPS shipping, plus are entitled to the Setup 1.x software (valued at US$25) upon your request as part of your purchase. It is recommended US customers purchase from HIT as highlighted.
1 buying the PE4H means can acquire parts later to do a x2 link
Recommend: an extra US$6-shipped DX mini hdmi cable. 2 users requiring a replacement part for the setup to work.
Which one? PE4L or PE4H?
PE4H has a sturdier base, can allow a 15-19V DC jack input and up to x4 connectivity.
PE4L is not as sturdy but is smaller and lighter. It can allow a 12V DC jack input and x1-only connectivty.
Amfeltec has a cheaper x1 1.0 mPCIe kit starting at US$29 here.
2. A video card of your choice
Review next item 2. Which video card do you recommend for best performance on my system? to determine if an ATI or NVIDIA solution will perform better on your setup.
3. A suitable power supply to drive your video card
Can determine how much power you require by counting the number of 6 or 8-pin PCIe connectors on your video card. The total power required to drive it will then be:
75W for the slot 1
+ 75W for each 6-pin PCIe connector
+ 150W for each 8-pin PCIe connector
1: Can be supplied using either the DC or the floppy molex jack. The DC jack differs on the PE4L and PE4H: PE4L takes 12V input, PE4H takes 15-20V (regulated) > 84W. The AC adapter must have a 2.5mm/5.5mm or 2.5mm/5mm barrel end to use the DC jack. The floppy molex input is used with an ATX PSU..
Power can be supplied by using either an ATX PSU or AC adapters. Can also split AC adapter power to supply a PCIe connector as shown here.
Refer also to ithildin's and serialk11r's PSU recomendations for details on how to read the PSU labelling. 430W may actually mean it only supplies 168W for our purposes!!
2. Which video card do you recommend for best performance on my system?This will be determined by your chipset, if additional ports available for x2 linkage and whether you have Intel 4500MHD, HD, HD3000 or HD4000 graphics. The matrix below makes recommendations based on the system chipset with:
- LIME recommend NVIDIA Fermi (GT430 or better) solution for best performance
- RED recommend ATI HD5750 or better solution for best performance
- CYAN choose based on preferred performance bias: DX9 (ATI HD5750 or better) or DX10/DX11 (NVIDIA Fermi). Refer to Performance: ATI versus NVIDIA benchmark results to help you decide. If not sure then suggest a NVIDIA Fermi solution which on balance gives better overall performance and has the option of x1.Opt with the right notebook. Note: if the NVIDIA x1.Opt driver was unlocked to allow primary video cards other than Intel 4500MHD/HD/HD3000/HD4000 plus it engaged with x2 links then I would recommend only NVIDIA solutions.
Left-to-right: slowest-to-fastest with NVIDIA x1.Opt and x2 delivering similiar performance
1 provides full duplex performance equivalent to x1E on Intel chipsets.
2 Optimus tweak engages only when Intel 4500MHD/HD/HD3000/HD4000 primary graphics AND x1 link is detected, improving performance by 20-333%. Provides ability for the internal LCD to render games running on the external GPU.
3 x1E tweak uses a x2 link on port1, 3 or 5 with a single lane for 15-30% more perf (ATI/non-Fermi NVIDIA only).
4 x2E tweak use two lanes of a x4 link, [port1+2] as EC+mPCIe or mPCIe+mPCIe for 30-60% more perf (ATI/non-Fermi NVIDIA only).
5 x2=combining expresscard+mPCIe or mPCIe+mPCIe slots [p1+2], [p3+4], [p5+6] or [p7+8]; x4=[p1+2+3+4] or [p5+6+7+8].
6Series-6/7: Intel setting link width of in the flash descriptor. Requires advanced bios hacking to configure x2 link, cannot be done using Setup 1.x.
? Don't know if non-Intel chipsets can do x2 links.
If going ATI, recommend a HD5xxx/HD6xxx series card over a HD4xxx series due to their more efficient usage of PCIe bandwidth. Benchmarks would predict a HD5670 to outperform a HD4870.
Avoid twin-gpu cards like a HD6970/GTX590 since they require 2x256MB PCIe space windows which might be difficult to accomodate.
Examples based on power consumption requirements- 12V/75W adapter: HD5670, unplugged HD5750, NVIDIA GT430.
- 12v/80w: specific HD5750 units. See comment here, implemented here.
- 12V/120W: GTS450 described here using DC adapter that is shared between notebook and eGPU
- 12V/203W xbox360 PSU: GTX460
- ATX PSU: AMD HD68xx, NVIDIA GTX4xx/GTX5xx
See also Suggested NVIDIA Setups.
3. Is it possible to get an enclosure?
A good thread to see some professional designed enclosures is at redback79's DIY enclosure project. There are some for sale or examples to build on your own. For example:
Joss27's 5.25" SCSI box, Wolf Tohsaka's Livebox Mini, timohour's floppy holder, fr0z3n's carry-it, Fujian stand/enclosure, airzone (1U), Kallogan (192W AC), hlgr, aodioxp, SimoxTav, #1, sprtnbsblplya, Agent9, Lurifax (mini ITX), kizwan (Aero QX2000, mITX), the endless +XB360 PSU, max_headroom and Oddious (1U), lightnica (JuiceBox), pengy_666's 5.25" drive enclosure, 12V/120W "portable", pimped-up (xb360) which is similar to Building an eGPU (CUDA).
4. What software setup do I need to make this work?
Spoiler :
You just need video card drivers for your chosen operating system. This will be available for download from www.amd.com (ATI/AMD) or www.nvidia.com (NVIDIA).
Or custom NVIDIA Optimus, DNA drivers for ATI cards are available too.
DIY eGPU Setup 1.x may also be required if encounter error 12: cannot allocate resources in Device Manager or have specific port requirements such as x1E, x2, x2E link widths or mPCIe antiwhitelisting. Click link for more details.
5. I'm using an ATX power supply to power my graphics card. How do I switch it on?
Spoiler :
PE4H is supplied with a SWEX, a small switched circuitboard used to power on/off the ATX PSU. The alternative is do the paperclip trick to switch on the PSU.
6. What cable connects PE4L/PE4H to EC2C/PM3N and is it replaceable?
Spoiler :
It's a mini_hdmi to mini_hdmi cable. Cheapest replacement being $US6-shipped DX 1.5m cable confirmed to work here.
7. [mPCIe] What can I do to make it easier to connect the mini hdmi cable to the PM3N?
Spoiler :
Could consider a cutout hole on the bottom plastic cover to provide an access point if you otherwise need to remove covers to get to it. OR could permanently attach PM3N internally, threading through a flat mini HDMI cable + adapter. The adapter would go on the PE4H/PE4L end. The bulb of the cable can be dremelled to sit more flush as shown here.
8. Why does my video card fail with an error 12: cannot allocate resources in Device Manager? See Troubleshooting FAQ - Q3.
ATI versus NVIDIA back-to-back benchmark results
NVS4200M vs GTX560Ti@x1.2Opt (i7-2670QM, Tom_ger)
GTX680@x1.2Opt vs x1.1Opt and GTX580 versus GTX680@x1.1Opt (i5-2520M 2.5, arsenal243)
GTX480 versus GTX560Ti@x1.1Opt (i7-620M 2.66, Mikjoa)
GTX460-1GB versus GTX560Ti @x1.1Opt (i5-2520M 2.5, Warwickeo)
GTS450 versus GTX460 @x1.Opt (i7-640M 2.80, docusyn)
HD5750 versus GTX470 on x1, x1.Opt and x2 links (i5-430M 2.26, nando4)
HD5750 versus GTX460 on x1 and x1E links (T9300-2.5, ithildin)
HD5750 versus GTX470 on x1, x1E, x1.Opt and x2 links (T6600-2.2/T2050-1.6/U7600-1.2, nando4)
Scaling Analysis
HD7970: pci-e 3.0 x2, x4, x8, x16 scaling
GTX580: x1.2Opt vs x1.1Opt +30% overclock & (i7-2920XM 2.4, Khenglish)
GTX560Ti: x1.2Opt vs x1.1Opt [internal LCD vs external LCD] (i5-2540M 2.6, nando4)
HD5850: x16, x4, x2, x1 (~i7-2820QM, timohour)
HD5870: x2 versus x1 (i7-720QM 1.6, kizwan)
GTX580+GTX460: x2 vs x1 vs GTX260M (P8700-2.53, mangos47)
GTX460@x1.Opt scaling analysis internal vs external LCD, +25% PCIe, +40% GPU core O/C (L9600@2.90, nando4)
GTX460@x2 scaling analysis +20% cpu, +20% PCIe, ~20% GPU (T9400-2.53, khenglish)
HD5750: x1 versus x1E (T9300-2.5, ithildin)
Examples and Leaderboard
Game screenshots taken by hitting ALT+TAB to go into the windowed mode, then ALT+PRTSC to capture screen. NVIDIA users: please set PhysX in the NVIDIA Control Panel to use the CPU rather than the GPU. Disable the external video card's HDMI audio for best performance.
DX9-centric Performance: RE5 1280x800-variable, 3dmark06 1280x1024, +optional any other current games.
DX10-centric Performance: ensure PhysX is set to CPU in the NVIDIA control panel.
DX11-centric Performance: set 1280x800 resolution and ensure PhysX is set to CPU in the NVIDIA control panel.
Key
RED - a x2 1.0 link using a PE4H linking expresscard+mPCIe or mPCIe+mPCIe. All others are x1 1.0.
ORANGE - appears to have port layout to do a high-performance x2 1.0 setup using either [port1+2], [port3+4], [port5+6], [port7+8]
GREEN - video card that requires only a compact 12v/80w AC adapter for power. No external ATX PSU needed.
All tested at 1280x800 except 3dmark06 which used the default 1280x1024 and those marked *
x1.1: x1 2.0 (5GT/s) capable system running at x1 1.0 (2.5GT/s). Awaiting new pci-e 2.0 DIY eGPU hardware.
x1.Opt: system with Intel 4500MHD, HD, HD3000 or HD4000 iGPU primary graphics using NVIDIA Optimus for higher performance + internal LCD mode.
x2.Opt = x2 link, Intel HD graphics, Optimus driver, but Optimus performance tweak only engages on x1 link as found.
x1E/x2E: apply x1E tweak to ATI cards and older NVIDIA cards on series-4 or older Intel chipsets to get ~full duplex performance
1 Requires DIY eGPU Setup 1.x to overcome 12: cannot allocate resources and/or to set specific port requirements.
2 EC/mPCIe: expresscard/mini_pci-e slot used by a EC2C/PM3N product. Port layout required for x1E/x2/x2E/x4 mode. Intel chipsets' require port1 enabled for other ports to work so port1 may not be a physical mPCIe/EC slot.
Systems incapable of running 4GB+ setups for reasons as explained. These will require a DSDT override to work. - 41st-gen i-core and core2Duo Dell Latitude have TOLUD=3.5GB.
- MSI CRx20
- Lenovo T420 doesn't work unless use bios 1.15 as explained.
- Sony E-Series i-core has an unusual reserved system device. Only a NVIDIA card can fit around it in 32-bit space.
- Dell XPS 1530M has TOLUD=3.5GB as explained.
Systems capable of a higher performance x4 or x2 PCIe links
The Orange and Red systems shown in the above table. Also:
- x4 1.0 (four x1 ports): Dell Inspiron 1520, Vostro 1500, XPS 1640/1645, Lenovo X200.
- x1 2.0 (single port): ExpressCard 2.0 with Series-6 chipsets. Not Series-5 PCIe 2.0 chipsets
- x2 1.0 (two x1 ports): Acer 5920G, 6920G, 6930G, 8930G. BenQ S41. Dell M1530, E4300, E6400?, Studio 1558 (WWAN=port1, wifi=port2), D630, D830, E6410, E5500, Vostro 1310, Fujitsu Amilo xi2555, HP DV9690en, HDX9000, Lenovo T400s?, T510, Y450/Y550, W510, MSI GT627 &, Toshiba A665-3D.
Systems not capable of a higher performance x2 1.0 PCIe link shown in black in the table above, as well as:
Acer 4741G(mPCIe), 5820TG(mPCIe), 8935G, Asus F3JC, G60, N61JQ, Dell Vostro_3400, Inspiron_1545*, HP DV3-2232tx, DV6-2119TX, DV7-3110TX, Lenovo G560, x201s, MSI A620x/Cx620, U200 (mPCIe), Samsung R580, Sony VGNNW29GF, Toshiba A300*, A500-026, A500-031, M305, P500-01g, R700, U500-00k.
* = can do a x1E tweak, relevant to only ATI cards and older NVIDIA cards, since the ExpressCard slot is an odd numbered port with nothing occupying the next port and is a series-4 or older Intel chipset.
04-03-2011 MacBook 17 T7600-2.33 with X1600 + HD6850 implemented with Win7/32
11-10-2010 Mac Mini T7200-2.0 + HD5450 implemented with Win7.
11-05-2010 MBP 17" T2600-2.16 with X1600 + GT240M result with Win7.
9-27-2010 MBP 15" Unibody T7500-2.2, 6GB RAM with 8600M GT+HD4650/HD5770-7LCDs impl. & & & with Win7.
8-11-2010 MBP 15" Unibody T7500-2.2 with 8600M GT + HD4870 implemented with Win7 and MacOS 10.6
For MacOS DIY eGPU implementations see DIY eGPU MacOS Experiences.
| CUDA/OPENCL APPLICATIONS >>> |
Octane Render: 10x faster than CPU-based rendering.
Badaboom: CUDA based video transcoding
| NON-VIDEO PCI-E CARDS >>> |
Creative SB X-Fi Titanium desktop sound card via the ExpressCard slot. See details.
| NON ENGLISH eGPU FORUMS >>> |
Acknowledgement this page has been cosmetically enhanced by the talented justinkw1, SimoxTav and lazyboy. Many thank yous to all the NBR members who have shared their DIY eGPU experiences for others to benefit from.
Last edited by nando4 : Yesterday at 12:31 AM.
|