DIY eGPU experiences

[mobius1aic]

It's pretty damn amazing how far you guys have come with these projects. *Kudos*

[ViciousXUSMC]

I was asked to bench my G73 for this thread.

I will add my results however it must be stated for my conscious to be clear that the benchmarks being used are ridiculous.

First of all 3dmark06 is too old now, it is easy cpu bottlenecked by any modern high end and even upper mid ranged video cards. Its also not going to play well with a quad core like whats in the G73 so I will hit a cpu limit way way way before I hit a gpu limit so you will be looking at the score thinking "wow thats low" but you wont be looking at the true score for the gpu and that is what this thread is all about.

Likewise the RE5 benchmark has a variable and a fixed. Fixed is the same in every run this is the one you would want to use to compare multiple systems, while variable is... well variable.

It lets AI and other components randomize the benchmark so the scores can fluctuate between each run. But by far the largest problem is the requested resolution for this test 1200x800? who games at that resolution?

Again this is going to be incredibly easy for the gpu and cause a massive cpu bottleneck and give invalid scores and false representations of a high end system.

In 2006 these would have been ok but now in 2010 technology has moved way beyond these low standards.

If you want to do this and do it right bench something like 3dmark09 Vantage, and then post ONLY the gpu score rather than the combined score so your comparing the gpu power and not the cpu + gpu power.

Just had to say that because too many ill informed eyes may look upon the information presented here and not understand what they are really looking at.

[key001]

G73 w/ 720 gets ~12400 3dmarks06 with 10.3 cat
can't do a comparison between 5870m and 5750 cuz there is no expresscard in it....

[User Retired 2]

x1E/x2E tweak mode on Intel chipsets gain 15-30% performance on odd numbered ports using ATI cards or older NVidia cards (GTX4xx is not affected)

NOTICE: ATI or older NVidia cards using Intel notebook chipsets' mPCIe/expresscard x1 1.0 are underperforming by 15-30%. NVidia's latest GTX4xx cards are not affected as shown here. Affected Systems: Intel ICH7/8/9M I/O-equipped chipsets below. Video cards tested: ATi HD5750 below and nVidia 9800GT. GTX4xx cards are NOT affected nor are AMD/NVidia chipsets so this problem is because of ATI/NVidia x1 1.0 pci-e handshaking implementation. series-5 chipsets do not see any performance improvement with the tweak below as documented in the DV4-2000+HD5750 implementation. RECOMMEND: escalate the issue to ATI/NVidia to get a solution. The problem has also been logged on Intel's community support page here and ATI's forum here.

Using ATI cards or older NVidia cards on Intel I/O chipsets' x1 1.0 pci-e ports has handicapped performance requiring setting it to x2 width to get a 15-30% performance improvement. This workaround applies only if the port used for the eGPU occupies an odd numbered port. It would appear that the these older ATI/NVidia cards may be running half-duplex, requiring setting it to x2 width to get the full x1 1.0 bandwidth.

The Nvidia MCP79 chipset based Dell XPS M1340 benchmarks here blitzes all the Intel machines at x1 1.0 speed. It gets full x1 1.0 bandwidth.

The tweak discovered to enable full x1 1.0 speed on a Intel pci-e port requires the mPCIe or expresscard to be an odd numbered port with no device on the next sequential port, eg: port1/port3/port5 with no device on port2/port4/port6. I'm calling this x1E mode, the "E" standing for Intel-Enhanced.

x2E being similar but requires a port1+port2 setup, with either lane3 or lane4 or both having NO physical device attached to them. port1 is set to x4 mode, running with only the first 2 lanes. x2E also offering 15-30% faster performance than x2 alone.

NOTE: if x1E is enabled on port3 or port5, the effect is the same as when x2 is engaged on them - port1 or port3 respectively will also be switched to x2 which in turn render the port2 or/and port4 non-operational.

Background

With the arrival of a PE4H V2 I attempted a x2/x4 link with the DV4, port layout here, hoping that Intel had made a mistake somewhere and I could link say port1+port3 as a dumbed down x4 link working in x2 1.0 mode. It didn't work.

During the testing however a very unusual performance increase was observed. I had set port1 into x2 width and was seeing ~20% better performance. That is *without* connecting a second lane. The 3dmark06 values don't look much different, but delve deeper and see the DV4's SM3.0 value of 3dmark06 has increased from 4344 to 5247 and is *higher* than the i7-620M+HD5750 x1 1.0 implementation's 5085. SM3.0 correlates directly to 3D gaming performance. The RE5 benchmarks improve considerably, dmcv4 dx9 sees a 30% improvement!! In case you were wondering, does setting port1 to be an x4 port with only the single lane connected see any further performance? The answer is no.. it does not.

x2 versus x1E versus x1 link speed, 2GB HD4870 versus 1GB HD5750

See also x1E versus x1 benchmarking and here and here (nVidia).

System	RAM	GPU	3dmark		RE5 (*=dx10)		dmcv4.scene4		PCIScript^2	Ports	OS
			06	vant.g	var	fixed	dx9	dx10
T2050-1.6 HP DV2000^1	3.0	HD5750@x2	6735	5579 +	38.5*	18.3*	60.1	55.4	Setup 1.x	ICH7M	Win7/32
		HD4870.2GB@x2	6710	5499	40.2*	18.1*	58.7	51.9		ICH7M
P8600-2.4 HP Pav_DV4	4.0	HD5750@x1E	10151 &	5215	64.0 &	35.0 &	66.3 &	41.8	-	ICH9M	Win7/64RC1
T6600-2.2 Dell Insp_1440	4.0		9551 &	5004	60.9 &	30.8 &	66.7 &	42.1 &	-	ICH9M	Win7/32
		HD4870.2GB@x1E	9188	5486	57.0	31.2	51.4	43.6		-
T2050-1.6 HP DV2000^1	3.0	HD5750@x1E	6694	4903 +	34.6*	17.0*	59.9	43.7	-	ICH7M	Win7/32
P8600-2.4 HP Pav_DV4	4.0	HD5750@x1	9461 &	4805	52.8 &	26.0 &	41.7 &	36.4 &	-	ICH9M	Win7/64RC1
T6600-2.2 Dell Insp_1440	4.0		8997 &	4432	51.6 &	25.7 &	44.7 &	38.0 &	-	ICH9M	Win7/32
		HD4870.2GB@x1	7913	4576	42.4	22.7	34.9	32.1	-	-
T2050-1.6 HP DV2000^1	3.0	HD5750@x1	6560	4219 +	33.2*	16.2*	43.8	36.7	-	ICH7M	Win7/32

&=overclocked HD5750 from 700/1150 to 790/1430
^1 port3 required mPCIe socket to be soldered onto systemboard's pads marked MINI2
^2 only first instance of PCIScript shown, duplicates per system have been removed.
Bold - benchmarking showing x2's superior performance even with a greatly inferior CPU

About HD5750 versus HD4870

We see the HD5xxx series has made more inroads at minimising pci-e bandwidth since the HD5750 outperforms the HD4870 in virtually all cases, particularly on low x1 1.0 bandwith setups. This contradicts what we'd expect. x16 2.0 benchmarks results indicate a HD4870 to be onpar to a HD5770 as shown here.

It is not clear then if 2GB VRAM does have any bandwidth sparing effects as I have suggested in my posts. Would need a direct comparison of the same series card (eg: 1GB HD4870 versus 2GB HD4870 or 1GB HD5850 versus 2GB HD5850) to see how much extra performance 2GB VRAM provides.

Overall I would advise against a HD4xxx card in favor of a more pci-e bandwidth efficient HD5xxx card.

How can I switch my port into x1E mode?

This requires an accessible odd numbered mPCIe or expresscard port with no device connected to a next even port as otherwise it won't work [port1 for DV4, port5 for Inspiron 1440]. If you have limited options, with perhaps a LAN card or some other device occupying a next even port, consider finding the datasheet to figure how to disable it. Or else if can identify Vcc power pin to the chip, it could be possible to do a cut the pin to disconnect it from being on the PCI bus. I really wanted to see if x1E could work on port3 of the DV4 since it's the more convenient expresscard slot, but the LAN card on port4 prevented that. I attempted to write 00 data to the LAN card's PCI space to try to disable it the device was still active. So x1E is only possible on the mPCIe port1 on the DV4.

The no-brainer solution is to use the DIY eGPU Setup USB Bootdisk to switch your port to x2 via the menus, or using 'iport x2 [port]' at the commandline.

Spoiler below details what 'iport x2' does which can be implemented using the older DIY ViDock bootdisk or, if you really want to not use the 'iport' command then can do this manual way using the new DIY eGPU Setup 1.x

Spoiler :

1. Identify RCBA address of your system using . setpci -s 0:1f.0 f0.l Disregard the trailing 1 treating it as zero. Eg: if setpci returns 0xFED18001 then RCBA is 0xFED18000. This can be done using Baredit in Windows as shown here. 2. If your bios has a next even *enabled* port, (eg: port2/4/6), do steps 3-4 otherwise go to step 5.

3. Add RCBA+0x3418 (that's hex.. can use calc.exe in scientific mode). That's were the function disable register is. 4. Set bit 17/19/21 to 1 to disable port 2/4/6 as visually shown here. 5. Add RCBA+0x224 (that's hex.. can use calc.exe in scientific mode). That's were the port configuration is stored. 6. Add appropriate details below to your DIY ViDock bootdisk PCIScript. :: Disable next even port port2/port4/port6 if bios has it enabeld pt MEM write 4 [RCBA+3418h] [change bit 17/19/21 to 1] ::choose port1 (+ non occupied port2) to be x2 high priority port pt MEM write 4 [RCBA+224h] 0x81 ::choose port3 (+ non occupied port4) to be x2 high priority port ::This also sets port1 to be a x2 port [ no way to prevent this! ] pt MEM write 4 [RCBA+224h] 0xA2 ::choose port5 (+ non occupied port6) to be x2 high priority port pt MEM write 4 [RCBA+224h] 0xC4 ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: :: Examples: create a x1E link. A 20% faster x1.0 link :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: :: example #1: HP_Pavilion DV4 using mPCIe port 1 :: 'setpci -s 0:1f.0 f0.l' returns RCBA=0xFED1C001 pt MEM write 4 0xFED1c224 0x81 :: example #2: HP_Pavilion DV2000 using mPCIe port 3 :: 'setpci -s 0:1f.0 f0.l' returns RCBA=0xFED1C001 :: Disable port2 (as port1 becomes x2 port) and port4 pt MEM write 4 0xFED1f418 0x3A0061 pt MEM write 4 0xFED1c224 0xA2 :: example #3: Dell Inspiron 1440 using expresscard port5 :: 'setpci -s 0:1f.0 f0.l' returns RCBA=0xFED18001 pt MEM write 4 0xFED18224 0xC4

What about users that have a PNP system without needing a DIY ViDock bootdisk?

This could only apply if you intend to setup port3 or port5 as a x1E link. port1 is the "root port". Disabling it affects all other devices off port1-port6, so it's not a good idea to use this method on port1. Use the DIY ViDock bootdisk instead to set a x1E port1.

The following worked on the Inspiron 1440 but caused a freeze on the DV4. The DV4 is using the off-limits port1.

1. Disable the port you want to set as x1E and any associated even ports as explained in steps 2-4 above using Baredit like shown here
2. Do a device manager scan or use 'devcon rescan'
3. Change desired port into x2 width by writing one of the following to RCBA+224h using Baredit.: 0x81(port1), 0xA2(port3) or 0xC4(port5).
4. Connect your EC2C or PM3N and power your video card.
5. Enable the port by reversing step 1. When you click "write" in baredit the port is enabled, now in x1E mode.
6. If Win7 doesn't automatically detect the port, do a device manager scan or do 'devcon rescan'

Only this sequence of steps resulted in a stable setup. The sequence above can be scripted. The baredit command can be 'saved', which saves it as \bar_save\bar_edit.ini. Running bar-edit from that directory then executes the commands in bar_edit.ini without user intervention.

Quirks

A standby/resume of the Inspiron 1440 resulted in port5 reverting back to x1 width. DV4 maintained the port state after a standby/resume. Suggest check your port with Everest after a resume to see if your bios is quirked this way.

[joseph_lin]

Updates of my VPC-Z11 Core i7 notebook.

The new RE5 benchmark is in the following link.
http://zwhraq.bay.livefilestore.com/...14_1458212.BMP

I can't use Ati Control Center as it keeps complaining no video card is found, other than that, it work fine, that said, no overclocking possible. I tried other overclocking tools, none of them detects ATI 5750!!

[key001]

Anyway.. I'm doing more investigation to see what ports could possible be made to work in this setup, so watch this space. For the time being, if you have an Intel-chipset, an accessible port1 as mPCie or expresscard, port2 *unoccupied*, then could get the extra 20% performance by booting up the DIY ViDock bootdisk and typing 'pt MEM write [RCBA+224h] 0x1' to switch port1 into a x2 port using only 1 lane to liberate 20% of performance. Detail as per ICH8M or newer datasheet

Shouldn't your onboard videocard be on port 2, northbridge?

I can't use Ati Control Center as it keeps complaining no video card is found, other than that, it work fine, that said, no overclocking possible. I tried other overclocking tools, none of them detects ATI 5750!!

Did you try Ati tray tools?

[jamesbond007]

i have a question. I remember in the first page it was mentioned that you configure the port 1 for x4 mode right. Why can't you do this with the pe4h to get a x4 speed on one port instead of ganging up two port to get x2

[wilse]

Intel chipsets have handicapped x1 1.0 ports

WIth the arrival of a PE4H V2 I attempted a x2/x4 link with the DV4, port layout here, hoping that Intel had made a mistake somewhere and I could link say port1+port3 as a dumbed down x4 link working in x2 1.0 mode. It didn't work.

During the testing however a very unusual performance increase was observed which made no sense at all. I had set port1 into x2 mode and got ~20% better performance. That is *without* connecting a second lane.

HP DV4, P8600-2.4 + HD5750, stock clocks + overclocked to 800/1430 benchmark results. I used the mPCIe port1 to do this. The expresscard slot is port6. Note the SM3.0 value of 3dmark06 of 5247 which is *higher* than the i7-620M+HD5750 implementation's 5085.

I've made a point of using PE4L for the testing below so it's clear it's only x1 1.0.

PE4L: Port set as x1 testing (max 3dmark=9854, RE5=55.1, RE5-fixed=26.3)


]

PE4L: Testing with x1 1.0 port1 set as a dummy x2 port (max 3dmark06=10587, RE5=68, RE5-fixed=36.0).

This is NOT x2 1.0. It's just full duplex/bandwidth x1 1.0 which is ~20% faster than the handicapped x1 1.0 in the above tests.




What is going on? It would appear then that either Intel divide the bandwidth amongst multiple ports OR their x1 port is half-duplex, requiring setting it to x2 mode, to get the full x1 performance.

This may explain too why the Nvidia MCP79 chipset based Dell XPS M1340 benchmarks on the first post blitzes all the Intel machines. Likely it gets full x1 1.0 bandwidth. It would be worthwhile asking Intel exactly how their ICHxM x1 1.0 pci-e lines conform to pci-e 1.1 spec of 2.5Gbps?

Anyway.. I'm doing more investigation to see what ports could possible be made to work in this setup, so watch this space. For the time being, if you have an Intel-chipset, an accessible port1 as mPCie or expresscard, port2 *unoccupied*, then could get the extra 20% performance by booting up the DIY ViDock bootdisk and typing 'pt MEM write [RCBA+224h] 0x1' to switch port1 into a x2 port using only 1 lane to liberate 20% of performance. Detail as per ICH8M or newer datasheet here.

Anyone have an insider in Intel that could provide some meaningful documentation on these ports, and how to configure them for maximum x1 1.0 performance?

nando - where or how did you get the pe4h v2? currently only v1.0 is orderable at hwtools.net

[avlan]

Does it means a possibility , if I have port #1 not used, I can disable it to speed-up ExpressCard's port#4?

[chung_gun]

Does it means a possibility , if I have port #1 not used, I can disable it to speed-up ExpressCard's port#4?

20% speedup is applied to odd number ports(1,3,5...) connected to external video card. the other needed condition is just one adjacent port(2,4,6,..)be empty.
so, you cannot speed up port 4.

and.. i think the disabling port #1 would result in disabling all pci express ports. because port #1 is a root port of pci express ports.

...