Author Topic: [MOVED] Hacking NVidia Cards into their Professional Counterparts  (Read 1644983 times)

0 Members and 3 Guests are viewing this topic.

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #650 on: November 03, 2013, 07:03:45 am »
Anyone have bios link to K10 or Grid K2 (or k1) please.

 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #651 on: November 03, 2013, 09:01:05 am »
Interesting. So GTX780Ti is overdue to hit the shelves. Full shader count (like the K6000), but half the VRAM of the Titan at 3GB, and 1/3 cheaper than the Titan. I wonder if DP will be crippled.

Seems we need to figure out where in the BIOS the VRAM size is stored and in what format. I'm going to be quite displeased if the Quadrified Titan doesn't work for VGA passthrough without a K6000 BIOS flashed onto it, and that will only work with the VRAM size adjustment.

Has anybody got a copy of a K20X BIOS handy? That should "just work" on a Titan with the strap mod.

Anyone have bios link to K10 or Grid K2 (or k1) please.

Why do you need it? GTX680 works just fine as a GK2 with it's original BIOS. I'd be surprised if GT635 didn't also work as a GK1 with it's original BIOS.

The BIOS requirements we are talking about here seems to be a new thing on GK110 based cards.
« Last Edit: November 03, 2013, 09:15:36 am by gordan »
 

Offline johnjoe

  • Contributor
  • Posts: 18
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #652 on: November 03, 2013, 09:29:20 am »
Interesting. So GTX780Ti is overdue to hit the shelves. Full shader count (like the K6000), but half the VRAM of the Titan at 3GB, and 1/3 cheaper than the Titan. I wonder if DP will be crippled.

Seems we need to figure out where in the BIOS the VRAM size is stored and in what format. I'm going to be quite displeased if the Quadrified Titan doesn't work for VGA passthrough without a K6000 BIOS flashed onto it, and that will only work with the VRAM size adjustment.

Has anybody got a copy of a K20X BIOS handy? That should "just work" on a Titan with the strap mod.

Anyone have bios link to K10 or Grid K2 (or k1) please.

Why do you need it? GTX680 works just fine as a GK2 with it's original BIOS. I'd be surprised if GT635 didn't also work as a GK1 with it's original BIOS.

The BIOS requirements we are talking about here seems to be a new thing on GK110 based cards.

I have a bios for k20xm and i modified the tesla driver. I was now able to install driver. I will test and post results.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #653 on: November 03, 2013, 09:55:30 am »
What is a K20XM? I cannot find any official reference to it. Do you have a download link for that BIOS?

And what exactly do you mean by "modified the tesla driver"? Modified how and to do what?
 

Offline johnjoe

  • Contributor
  • Posts: 18
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #654 on: November 03, 2013, 10:14:10 am »
K20xm should be the same like k20x from the point of performance. I didn't found any differences. I looked at the inf file of nvidia tesla driver and found that no 1020 id is listed, means no k20x. I changed the predefined 1021 (k20xm) to 1020 and i was able to install driver. But gpuz for example report only 512mb vram. And cuda sim programs still not recognize the card.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #655 on: November 03, 2013, 10:25:57 am »
That's probably because there is an issue due to the modified driver. Change the strap in the Tesla BIOS to make it 0x1021 for K20xm, then do a clean install of the unmodified driver. That might help.
 

Offline johnjoe

  • Contributor
  • Posts: 18
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #656 on: November 03, 2013, 10:29:45 am »
I will test. How to change straps to 1021?
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #657 on: November 03, 2013, 11:32:49 am »
Interesting. So GTX780Ti is overdue to hit the shelves. Full shader count (like the K6000), but half the VRAM of the Titan at 3GB, and 1/3 cheaper than the Titan. I wonder if DP will be crippled.

I'm sure it will be crippled the same as the GTX780.

The BIOS requirements we are talking about here seems to be a new thing on GK110 based cards.

I'm pretty sure that only because of the modification to Tesla cards (TCC mode code is probably not available in GeForce BIOSes).

What is a K20XM? I cannot find any official reference to it. Do you have a download link for that BIOS?

The "M" versions are passively cooled cards. That's the only difference (maybe also lower clock speeds).

K20xm should be the same like k20x from the point of performance. I didn't found any differences. I looked at the inf file of nvidia tesla driver and found that no 1020 id is listed, means no k20x. I changed the predefined 1021 (k20xm) to 1020 and i was able to install driver. But gpuz for example report only 512mb vram. And cuda sim programs still not recognize the card.

512MB VRAM? Wow, that is strange. I still don't think there is much difference in Tesla BIOSes. If you just get your device ID correct, then it should behave like that card so you probably do not have to use the K20Xm BIOS.

I will test. How to change straps to 1021?

I should be changing the 0x00000080 value at 0x20 to 0x00040080 (with 0x1C still at 0xFFC3FF7F).
« Last Edit: November 03, 2013, 11:36:33 am by oguz286 »
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #658 on: November 03, 2013, 12:34:07 pm »
OT: firepro S7000 S8000 S9000 S10000 V7800P V9000P [v7800p? cheap!] for esxi 5.5!!

Any potential with these models to have shared vgpu *more than 1 user* and cross-flash/hack AMD?

also special edition 780 will have 6gb and 12GB. which maybe gets rid of whole memory issue problems stored on rom
« Last Edit: November 03, 2013, 12:36:50 pm by mrkrad »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #659 on: November 03, 2013, 01:45:25 pm »
I wouldn't take rumours for granted - don't expect a GTX780Ti with more than 3GB of RAM until you actually see it. There are also still a number of questions about the 780Ti that remain to be answered, not least of which is whether the DP performance is crippled unlike on the Titan.

BIOS cross-flashing on ATI cards was doable around the HD4xxx era, don't know what the deal with it is now. But if the VGA passthrough experience with ATI is anything to go by, I wouldn't begin the experiment with much optimism.
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #660 on: November 03, 2013, 07:09:22 pm »
@johndoe: I just read that it was GPU-Z that reported 512MB  |O GPU-Z cannot be trusted. It says that my GTX780 has 48 cores (my primary card is a GT610 for testing purposes, and THAT card has 48 cores).

Use the nvidia-smi tool to read out the real values. And of course you should also run the deviceQuery CUDA sample and report its output.
After some more research, I figured out some things. The amount of VRAM is determined by a couple of things: the bus-width, the number of RAM chips, and the size of each RAM chip. The GTX780 has 12 chips under the cooler, and zero on the back. A Titan has 12 chips under the cooler, and 12 on the back (please confirm this), and the Tesla BIOS is configured to select 24 chips, whereas I have 12. That's why it doesn't work and that why it should probably work for you.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #661 on: November 03, 2013, 07:33:49 pm »
@johndoe: I just read that it was GPU-Z that reported 512MB  |O GPU-Z cannot be trusted. It says that my GTX780 has 48 cores (my primary card is a GT610 for testing purposes, and THAT card has 48 cores).

I have to say I have never seen GPU-Z get it wrong before. There is a drop-down box where you select which card you want GPU-Z to report on.
Typically GPU-Z gets it badly wrong only when the driver is force-installed and doesn't initialize properly.

Use the nvidia-smi tool to read out the real values. And of course you should also run the deviceQuery CUDA sample and report its output.
After some more research, I figured out some things. The amount of VRAM is determined by a couple of things: the bus-width, the number of RAM chips, and the size of each RAM chip. The GTX780 has 12 chips under the cooler, and zero on the back. A Titan has 12 chips under the cooler, and 12 on the back (please confirm this), and the Tesla BIOS is configured to select 24 chips, whereas I have 12. That's why it doesn't work and that why it should probably work for you.

Interesting observation. Does that help gain some insight into how the memory size is encoded in the BIOS?
 

Offline johnjoe

  • Contributor
  • Posts: 18
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #662 on: November 03, 2013, 08:04:59 pm »
So, i used now bios of k20c and titan, both with id of k20x, modified the driver because all k20 cards are set to the same driver section. gpu-z is maybe. I added the id of k20x. The results of devicequery.exe with titan bios are in the picture. Results of cuda-z are also added. Nevertheless, the sim software does not recognize the card. Last thing I will do is to use k20xm id with my k20xm bios.
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #663 on: November 03, 2013, 09:55:20 pm »
It will not work if you use a modified Titan BIOS, so you will have to use a Tesla BIOS. If your primary card is the GT9500 and the Titan is the secondary card, then it should work. I also don't see why you need to add the K20X device ID to the driver installer. Are you using an older driver? Just use the newest GeForce driver.

The only difference I can see is that you replaced the 25K resistor with a 40K resistor. I did not remove the 25k resistor, but I added a 33K resistor between SCLK and Vcc on the EEPROM. My card does work (except for memory issues), so that is the only thing that I can think of. Maybe your solution changed some other straps without you knowing it.
 

Offline johnjoe

  • Contributor
  • Posts: 18
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #664 on: November 03, 2013, 10:28:59 pm »
Ok, i understood now. Maybe you are right and I changed add. straps. In your case, you only added a 33k res between vcc and sclk without changing anything else? And you took normal Geforce driver instead of Tesla driver from nvidia? I can only try tomorrow with resistors but will report. What you mean with secondary card, not connected to screen? it is not connected.
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #665 on: November 03, 2013, 11:21:23 pm »
Yes, the only change is the 33K resistor between SCLK and Vcc.

With secondary card I mean that your GT9500 is in the first PCI express slot (primary) and your Titan is in the second PCI express slot (secondary). That is crucial, because else it will not work (at least it didn't work for me).
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #666 on: November 03, 2013, 11:53:24 pm »
GeForce driver running a Tesla? Really? That's not supposed to work because the GeForce driver .inf doesn't contain any Quadro/Tesla/Grid device IDs. Unless this is some kind of a unified pre-release driver.

I thought we already established there is no need to replace the nibble 3 resistor - you can just leave it off, and fix the potential slight flapping issue by setting the 5th bit in the soft-strap appropriately.

It is normal that with a Tesla BIOS the machine won't POST on that card - Teslas don't work as standard VGA adapters, they normally show up as 3D adapters. Having said that, on the 690, the primary GPU shows up as VGA, the secondary as 3D, and that posts on ports attached to either, so the chances are that the Tesla BIOS just disables all the video outputs (since Tesla cards don't have any). I imaging flashing a modified Grid BIOS to a GTX680/770/690 would also produce similar results.

Having said that, I seem to recall I found that the device type (i.e. VGA or 3D) is set by a bit in the secondary strap - but I don't know for sure where the secondary strap is in the UEFI setup (several possible candidates IIRC) or whether the one on the main BIOS payload is the effective one. But if you can find it, you could potentially make the Tesla BIOS work with normal VGA enabled, but that would be a bit of a bizzare use-case (going headless after booting).
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #667 on: November 04, 2013, 12:10:12 am »
GeForce driver running a Tesla? Really? That's not supposed to work because the GeForce driver .inf doesn't contain any Quadro/Tesla/Grid device IDs. Unless this is some kind of a unified pre-release driver.

I thought we already established there is no need to replace the nibble 3 resistor - you can just leave it off, and fix the potential slight flapping issue by setting the 5th bit in the soft-strap appropriately.

It is normal that with a Tesla BIOS the machine won't POST on that card - Teslas don't work as standard VGA adapters, they normally show up as 3D adapters. Having said that, on the 690, the primary GPU shows up as VGA, the secondary as 3D, and that posts on ports attached to either, so the chances are that the Tesla BIOS just disables all the video outputs (since Tesla cards don't have any). I imaging flashing a modified Grid BIOS to a GTX680/770/690 would also produce similar results.

Having said that, I seem to recall I found that the device type (i.e. VGA or 3D) is set by a bit in the secondary strap - but I don't know for sure where the secondary strap is in the UEFI setup (several possible candidates IIRC) or whether the one on the main BIOS payload is the effective one. But if you can find it, you could potentially make the Tesla BIOS work with normal VGA enabled, but that would be a bit of a bizzare use-case (going headless after booting).

How do you want to change the device ID from 0x1004/0x1005 to 0x1020 when you can only go as high as 0x101F with just the soft straps?

Yes you can use the GeForce driver with Tesla cards. At work we have Teslas and we never installed Tesla drivers, so I don't know why it wouldn't work. We use the Tesla C2050 which has two DVI outputs, so when you use the Tesla with the standard WDDM driver, it is basically a slower GeForce card.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #668 on: November 04, 2013, 08:33:05 am »
How do you want to change the device ID from 0x1004/0x1005 to 0x1020 when you can only go as high as 0x101F with just the soft straps?

Perhaps I wasn't clear enough, apologies. I wasn't suggesting a purely soft-mod. I was saying that if you remove the resistor and don't replace it, the value of the nibble goes to the top of it's settable range, but there is often an instability in the 5th bit. For example, on a GTX680, if you remove it, it will go to A or B, and often flip between the two when you reboot. You can compensate for this by soft-strapping the 5th bit high (to make it B) or low (to make it A). Since removing an 0402 is slightly easier than replacing it, it makes for a smaller, easier to apply hard-mod, and the rest can be soft-modded.

Yes you can use the GeForce driver with Tesla cards. At work we have Teslas and we never installed Tesla drivers, so I don't know why it wouldn't work. We use the Tesla C2050 which has two DVI outputs, so when you use the Tesla with the standard WDDM driver, it is basically a slower GeForce card.

That surprises me - I didn't think there was that overlap in the Windows drivers. I hadn't expected the device IDs for any Tesla/Quadro/Grid cards to be listed in the .inf. Certainly when you go to the Nvidia site and select that you want a driver for the Tesla, it will point you at the Tesla/Quadro/Grid download rather than the GeForce one. But hey, if it works for you... :)
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #669 on: November 04, 2013, 10:07:35 am »
How do you want to change the device ID from 0x1004/0x1005 to 0x1020 when you can only go as high as 0x101F with just the soft straps?

Perhaps I wasn't clear enough, apologies. I wasn't suggesting a purely soft-mod. I was saying that if you remove the resistor and don't replace it, the value of the nibble goes to the top of it's settable range, but there is often an instability in the 5th bit. For example, on a GTX680, if you remove it, it will go to A or B, and often flip between the two when you reboot. You can compensate for this by soft-strapping the 5th bit high (to make it B) or low (to make it A). Since removing an 0402 is slightly easier than replacing it, it makes for a smaller, easier to apply hard-mod, and the rest can be soft-modded.

Ah, that's what you meant. Now I understand. Still, I don't know if that would work in this case. It's true that removing a 0402 resistor is easier than soldering it (good luck with that :P), but my solution is non-destructive. I can always remove the 33K resistor (1206 btw) and the card will be a GTX780 again, whereas adding the 25K resistor back on the board would be harder :)
There is also another difference. The 25K resistor is a pull-down resistor, and I added a pull-up resistor. The lines that go to the GPU most likely have an ADC that measures the voltage and based on that it sets certain straps. I had also removed the 28.5K resistor and let me tell you, that one had to be precise! It will not boot with a value of 28K or 29K :(

Yes you can use the GeForce driver with Tesla cards. At work we have Teslas and we never installed Tesla drivers, so I don't know why it wouldn't work. We use the Tesla C2050 which has two DVI outputs, so when you use the Tesla with the standard WDDM driver, it is basically a slower GeForce card.

That surprises me - I didn't think there was that overlap in the Windows drivers. I hadn't expected the device IDs for any Tesla/Quadro/Grid cards to be listed in the .inf. Certainly when you go to the Nvidia site and select that you want a driver for the Tesla, it will point you at the Tesla/Quadro/Grid download rather than the GeForce one. But hey, if it works for you... :)

I guess the only difference between the drivers is that the Tesla drivers do not have the device IDs of the GeForce cards in the inf file. Haven't checked it though.
« Last Edit: November 04, 2013, 11:29:08 am by oguz286 »
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #670 on: November 04, 2013, 12:21:40 pm »
what I was saying. the K10 tesla is the quadro grid K2 - features like software-ecc and vgx support on both. Just strap them and go.

finding the kepler that is closest to the K10/grid k2 should not be that hard since they are all the same core chip with a tweak to the soft straps .

That crazy Quadroplex Quadro 6000 was renamed to be the #2 card. Strapped it back to #1 position and all was good.

Now that optix 3 supports regular nvidia cards (OSX premier adobe) the need for tweaked geforce cards  may change a lot soon. The bet I'm going for is finding the cheaper Quadro/Tesla to make into the better (quadro/tesla). The GK104 has many many choices to work with. GK107 too.

I'm hunting VGX since it is better than API-intercept. I think I agree with Gordan. VGX is more like "VT-d" for video cards, alone it won't do squat but it will work with nvidia's VGPU API to accelerate much faster than pure binary translation mode (API intercept)
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #671 on: November 04, 2013, 12:48:57 pm »
Ah, that's what you meant. Now I understand. Still, I don't know if that would work in this case. It's true that removing a 0402 resistor is easier than soldering it (good luck with that :P), but my solution is non-destructive. I can always remove the 33K resistor (1206 btw) and the card will be a GTX780 again, whereas adding the 25K resistor back on the board would be harder :)

Can you post a photo of your mod showing exactly where you put the 33K resistor?

And why do you think removing the 3rd nibble resistor wouldn't work in this case? It works on the GTX680 and GTX690 I have. You mean it might upset the voltage somewhere else and cause unrelated straps to end up with wrong values? Surely adding a pull-down resistor in addition to the pull-up would run the same risk of doing that.

Note: I'm not disputing that dealing with a 1206 is far preferable than dealing with an 0402, especially without specialist tools. :)

There is also another difference. The 25K resistor is a pull-down resistor, and I added a pull-up resistor. The lines that go to the GPU most likely have an ADC that measures the voltage and based on that it sets certain straps. I had also removed the 28.5K resistor and let me tell you, that one had to be precise! It will not boot with a value of 28K or 29K :(

Which 28.5K resistor? What was it for?

what I was saying. the K10 tesla is the quadro grid K2 - features like software-ecc and vgx support on both. Just strap them and go.

...

I'm hunting VGX since it is better than API-intercept. I think I agree with Gordan. VGX is more like "VT-d" for video cards, alone it won't do squat but it will work with nvidia's VGPU API to accelerate much faster than pure binary translation mode (API intercept)

I'm not sure I follow what this would be for.
1) Why do you need ECC for graphics rendering and video stream encoding?
2) I'm pretty sure VGX requires no special features at all. vDGA is just straight PCI passthrough a-la Xen. vSGA just makes the GPU act as a co-processor. I'm going to try putting together an ESXi test bed machine with a spare motherboard I have which I _think_ has a non-broken IOMMU with the required features for ESXi PCI passthrough, and try and get my Gridified 690 working on it. If that works, it would prove you need no special features on the GPU itself to make it work.

It will also be good to hear back from foxdie when he has had a chance to test the Quadrified GTX470 with vSGA. If that works, the chances are that other modified cards will, too.
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #672 on: November 04, 2013, 01:10:16 pm »
Ah, that's what you meant. Now I understand. Still, I don't know if that would work in this case. It's true that removing a 0402 resistor is easier than soldering it (good luck with that :P), but my solution is non-destructive. I can always remove the 33K resistor (1206 btw) and the card will be a GTX780 again, whereas adding the 25K resistor back on the board would be harder :)

Can you post a photo of your mod showing exactly where you put the 33K resistor?

And why do you think removing the 3rd nibble resistor wouldn't work in this case? It works on the GTX680 and GTX690 I have. You mean it might upset the voltage somewhere else and cause unrelated straps to end up with wrong values? Surely adding a pull-down resistor in addition to the pull-up would run the same risk of doing that.

Note: I'm not disputing that dealing with a 1206 is far preferable than dealing with an 0402, especially without specialist tools. :)

I have the card in my computer now with the cooler on it. When and if I take it apart again, I'll take a photo. But it's really just two very thin wires soldered on the Vcc and the SCLK pins of the EEPROM going to a 1206 33K resistor.

The card has a pull-down resistor by default, and I added a pull-up resistor. If the other end of the SCLK line has an ADC, then having no resistor on it will cause a floating pin which is not something you want with analog electronics. A pull-down, pull-up, or both will make everything more stable electric wise (even though you override that value with the soft-straps).

There is also another difference. The 25K resistor is a pull-down resistor, and I added a pull-up resistor. The lines that go to the GPU most likely have an ADC that measures the voltage and based on that it sets certain straps. I had also removed the 28.5K resistor and let me tell you, that one had to be precise! It will not boot with a value of 28K or 29K :(

Which 28.5K resistor? What was it for?
It is a pull-down resistor on the SO pin of the EEPROM. I measured around 2.3M resistance when I removed the 28.5K resistor which is actually a 30K resistor when measured outside the circuit. Because it needs to be exactly that value, I used a high-precision pot because I couldn't put back the 0402 resistor. It is on the schematic on my site and on the photo I posted a couple of pages back.

what I was saying. the K10 tesla is the quadro grid K2 - features like software-ecc and vgx support on both. Just strap them and go.

...

I'm hunting VGX since it is better than API-intercept. I think I agree with Gordan. VGX is more like "VT-d" for video cards, alone it won't do squat but it will work with nvidia's VGPU API to accelerate much faster than pure binary translation mode (API intercept)

I'm not sure I follow what this would be for.
1) Why do you need ECC for graphics rendering and video stream encoding?
2) I'm pretty sure VGX requires no special features at all. vDGA is just straight PCI passthrough a-la Xen. vSGA just makes the GPU act as a co-processor. I'm going to try putting together an ESXi test bed machine with a spare motherboard I have which I _think_ has a non-broken IOMMU with the required features for ESXi PCI passthrough, and try and get my Gridified 690 working on it. If that works, it would prove you need no special features on the GPU itself to make it work.

It will also be good to hear back from foxdie when he has had a chance to test the Quadrified GTX470 with vSGA. If that works, the chances are that other modified cards will, too.

Not to be a jerk, but wouldn't it be wiser if the conversation about virtualized GPUs would take place in a separate topic? I'm interested in it as well, but the last pages of this topic has been nothing but questions about if card X can be modified or about virtualization. You know, just to make it less cluttered. :)
« Last Edit: November 04, 2013, 01:45:23 pm by oguz286 »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #673 on: November 04, 2013, 04:30:49 pm »
I have the card in my computer now with the cooler on it. When and if I take it apart again, I'll take a photo. But it's really just two very thin wires soldered on the Vcc and the SCLK pins of the EEPROM going to a 1206 33K resistor.

Oh, I see. So just short the EEPROM Vcc to SCLK with a 33KO resistor?
Won't this potentially upset other things?

The card has a pull-down resistor by default, and I added a pull-up resistor. If the other end of the SCLK line has an ADC, then having no resistor on it will cause a floating pin which is not something you want with analog electronics. A pull-down, pull-up, or both will make everything more stable electric wise (even though you override that value with the soft-straps).

Indeed, that that sounds like a good idea. I might check if the same approach might work on the GTX690, if I manage to steady my hands enough to solder some 24.8K resistors back on.

Not to be a jerk, but wouldn't it be wiser if the conversation about virtualized GPUs would take place in a separate topic? I'm interested in it as well, but the last pages of this topic has been nothing but questions about if card X can be modified or about virtualization. You know, just to make it less cluttered. :)

Agreed - let's keep this purely down to the modding part. There are at least two other threads here for what to use the modded cards for.
 

Offline oguz286

  • Contributor
  • Posts: 39
  • Country: nl
  • Make, break, hack, tweak
    • GuzTech
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #674 on: November 06, 2013, 12:21:56 pm »
I have the card in my computer now with the cooler on it. When and if I take it apart again, I'll take a photo. But it's really just two very thin wires soldered on the Vcc and the SCLK pins of the EEPROM going to a 1206 33K resistor.

Oh, I see. So just short the EEPROM Vcc to SCLK with a 33KO resistor?
Won't this potentially upset other things?

Yes, just a resistor between Vcc and SCLK. Seeing as I'm the only one that has a working Tesla (except for memory size issues), I'm guessing that my mod does not change anything else. I would like to try it with a Titan, but that thing is expensive. :( That's why I'm waiting for johndoe to apply my mod see if it works then. Btw, did you get a Titan?
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf