Author Topic: [MOVED] Hacking NVidia Cards into their Professional Counterparts  (Read 1217786 times)

0 Members and 6 Guests are viewing this topic.

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #750 on: November 25, 2013, 12:34:03 pm »
It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?

Have you checled the 680 photos posted here on the forum that show the location of the resistors? Have you had a look at your card to identify if the resistors are in the same place(s)?
 

Offline mosmo

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #751 on: November 25, 2013, 01:04:39 pm »
It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?

Have you checled the 680 photos posted here on the forum that show the location of the resistors? Have you had a look at your card to identify if the resistors are in the same place(s)?

Not yet, I will do this. I am keen to know if SLI will work or not. Or is it that K series or Quadro don't allow for SLI??
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #752 on: November 25, 2013, 01:15:25 pm »
It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?

Have you checled the 680 photos posted here on the forum that show the location of the resistors? Have you had a look at your card to identify if the resistors are in the same place(s)?

Not yet, I will do this. I am keen to know if SLI will work or not. Or is it that K series or Quadro don't allow for SLI??

I have never tried SLI - I'm pretty sure Grids/Teslas don't support it so the driver doesn't expose it. Then again there are a lot of extra limitations in a VM, I noticed that several options that appear on bare metal don't appear on the VM with the same card (e.g. on XP64 domU the PhysX options don't appear, but on bare metal they do). So I wouldn't pre-emptively expect SLI to work, but if you discover otherwise, please, do report back.
 

Offline mosmo

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #753 on: November 25, 2013, 01:38:48 pm »
It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?

Have you checled the 680 photos posted here on the forum that show the location of the resistors? Have you had a look at your card to identify if the resistors are in the same place(s)?

Not yet, I will do this. I am keen to know if SLI will work or not. Or is it that K series or Quadro don't allow for SLI??

I have never tried SLI - I'm pretty sure Grids/Teslas don't support it so the driver doesn't expose it. Then again there are a lot of extra limitations in a VM, I noticed that several options that appear on bare metal don't appear on the VM with the same card (e.g. on XP64 domU the PhysX options don't appear, but on bare metal they do). So I wouldn't pre-emptively expect SLI to work, but if you discover otherwise, please, do report back.

If I manage to bork both of my 670s I will post it here. If only I had money to burn on the Palit 780ti , I would love to know if this could be converted to the pro equivalent.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #754 on: November 25, 2013, 01:48:26 pm »
It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?

Have you checled the 680 photos posted here on the forum that show the location of the resistors? Have you had a look at your card to identify if the resistors are in the same place(s)?

Not yet, I will do this. I am keen to know if SLI will work or not. Or is it that K series or Quadro don't allow for SLI??

I have never tried SLI - I'm pretty sure Grids/Teslas don't support it so the driver doesn't expose it. Then again there are a lot of extra limitations in a VM, I noticed that several options that appear on bare metal don't appear on the VM with the same card (e.g. on XP64 domU the PhysX options don't appear, but on bare metal they do). So I wouldn't pre-emptively expect SLI to work, but if you discover otherwise, please, do report back.

If I manage to bork both of my 670s I will post it here. If only I had money to burn on the Palit 780ti , I would love to know if this could be converted to the pro equivalent.

I'm more interested to hear if you find that Tesla K10 works with virtualization,

780/Titan aren't fully hacked yet - we are still looking for the resistors that control the 4th nibble. I think I found two most probable candidates, but soldering 0402 components is beyond the limits of my hand steadyness.

Another thing that I've been pondering is that there are clearly a number of other strap resistors, they just aren't the ones that control the device ID. The reason why I'm pondering this is because one of them might be controlling the disabled GL primitives that make Quadros perform much better in applications that rely on them.
« Last Edit: November 25, 2013, 01:50:21 pm by gordan »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #755 on: November 25, 2013, 07:13:51 pm »
Due to popular demand, I have finally found time to write this up. I had to rush it a little, but hopefully you get the gist.
Nvidia GeForce 4xx Fermi Soft Modding Guide

Any questions, please ask away, and I'll update the article to expand on those points where appropriate.
 

Offline mosmo

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #756 on: November 25, 2013, 07:31:32 pm »
Due to popular demand, I have finally found time to write this up. I had to rush it a little, but hopefully you get the gist.
Nvidia GeForce 4xx Fermi Soft Modding Guide

Any questions, please ask away, and I'll update the article to expand on those points where appropriate.

Fantastic! Please let us all know when you have finished your Kepler guide, awesome work.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #757 on: November 26, 2013, 05:01:14 pm »
Due to popular demand, I have finally found time to write this up. I had to rush it a little, but hopefully you get the gist.
Nvidia GeForce 4xx Fermi Soft Modding Guide

Any questions, please ask away, and I'll update the article to expand on those points where appropriate.

Fantastic! Please let us all know when you have finished your Kepler guide, awesome work.

Arguably this thread _IS_ the Kepler modding guide, you know. :)

My main motivation for writing an article about it is to summarize what works and what doesn't, and condense the information so that it isn't necessary to read through 51 (and still ticking up) pages of posts to find all the important info.
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #758 on: November 26, 2013, 07:43:56 pm »
I just wanted to post my experience on average ebay/amazon used price for top notch cards!

Average sales price:

Grid K2  : 1650$ [Geforce 690 new cost $1000 !!]

Tesla K10: $1600 [ same thing as Grid K2 mostly]

Tesla K20: $1600 [k20 is almost like K6000, Quadro K40 is K6000 exactly!]

Quadro 6000: $690 [ older FERMI model that works great with ESXi but doesn't have VGX]

Quadro 4000: $200 [ older fermi model, no ecc]

What I'd like to do is locate possible KEPLER cards that would be a good conversion (sensible to hack)!

Hacking a $1000 Geforce 690 to become a $1650 grid K2 seems a little illogical (risk versus cost)

But hacking a $120 Quadro K600 or NVS into one fourth of a Grid K1 might be a good value!



Just fyi
« Last Edit: November 26, 2013, 10:20:59 pm by mrkrad »
 

Offline Luke212

  • Contributor
  • Posts: 5
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #759 on: November 27, 2013, 02:07:31 am »
Hello,

Thank you for the work in this thread. I am interested in modding a Titan to a TCC enabled device. It will dramatically improve performance in Cuda. The background information is that Titan uses the WDDM display mode driver, which incurs a massive delay every time a kernel is executed. The TCC mode does not suffer this delay because it bypasses the windows WDDM driver.

From what I read, I need to solder a resistor on to the board, flash a K20c bios, and do a hex soft mod? I have very little experience in any of these steps. I will re-read the thread to try an absorb it all, but if there were a guide to go through these steps for beginners, I am sure it would be very popular and greatly appreciated. Thanks again for the hard work here.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #760 on: November 27, 2013, 08:32:37 am »
Thank you for the work in this thread. I am interested in modding a Titan to a TCC enabled device. It will dramatically improve performance in Cuda. The background information is that Titan uses the WDDM display mode driver, which incurs a massive delay every time a kernel is executed. The TCC mode does not suffer this delay because it bypasses the windows WDDM driver.

From what I read, I need to solder a resistor on to the board, flash a K20c bios, and do a hex soft mod? I have very little experience in any of these steps. I will re-read the thread to try an absorb it all, but if there were a guide to go through these steps for beginners, I am sure it would be very popular and greatly appreciated. Thanks again for the hard work here.

The Titan/780 mod isn't ready for public consumption yet. Specifically, we still need to find the resistor pair controlling the 4th nibble. Soft-modding via a BIOS patch does not work on Kepler class GPUs. It will affect the PCI device ID, which will make it appear to work, but the device ID seems to be notified to the GPU via the hard-strap before the EEPROM gets read and the BIOS payload is executed. The driver queries the device ID from the GPU directly, and enables features based on that. To make it worse, if the driver notices the discrepancy between the GPU device ID and PCI device ID it will lock down PCIe speed to 1.1, which will further cripple the performance on I/O heavy tasks.

If you want a TCC capable card and you don't want to have to bother with soldering, get a Fermi class GPU and mod it into a Quadro (or a Tesla), for example a GTX480 and mod it into a Quadro 6000. Instructions are in the Fermi/4xx soft modding guide linked above.

If a Fermi isn't good enough for your number crunching needs, you could go with a GTX 670/680/770 mod into a Tesla K10. This involves only a removal of one resistor, and that resistor is on the back of the PCB, so you wouldn't even have to take off the heatsink. Note that for some kernels a GK104 Kepler is slower than a GF100 Fermi.

My advice would be to wait until the resistor pair controlling the 4th nibble is confirmed. I think I found two good candidates, but my soldering skills aren't good enough for 0402 size components to attempt it.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #761 on: November 27, 2013, 04:24:23 pm »
I just wanted to post my experience on average ebay/amazon used price for top notch cards!

Average sales price:

Grid K2  : 1650$ [Geforce 690 new cost $1000 !!]

Tesla K10: $1600 [ same thing as Grid K2 mostly]

Tesla K20: $1600 [k20 is almost like K6000, Quadro K40 is K6000 exactly!]

Quadro 6000: $690 [ older FERMI model that works great with ESXi but doesn't have VGX]

Quadro 4000: $200 [ older fermi model, no ecc]

What I'd like to do is locate possible KEPLER cards that would be a good conversion (sensible to hack)!

Hacking a $1000 Geforce 690 to become a $1650 grid K2 seems a little illogical (risk versus cost)

But hacking a $120 Quadro K600 or NVS into one fourth of a Grid K1 might be a good value!

You are missing a few points.

1) Grid K1 is based on GK208, which is also what what GeForce 630/635/640 is based on which is still significantly cheaper than the Quadro K600. Why exactly not modify that instead?

The only possible advantage is if you need the few missing GL primitives implemented in hardware/FPGA/whatever that the real Quadros have and GeForces don't, but that's only useful for the kind of people who use the sort of apps whose performance is reflected in the likes of SPECviewperf, and those people aren't the types that will be bothered with saving a few thousand $ on hardware since their company will be paying for it, and unless they are a very tightly run cheap company, their CTO will probably want to hide behind the support contract that is only applicable on supported hardware, which won't include modified GeForce cards.

2) Grid/Tesla cards are all well and good if all you want to do is use them only for number crunching or vSGA, but most of us actually want VGA passthrough / vDGA which you cannot achieve with a real Grid or recent Kepler Tesla cards for the simple reason that they have no video outputs. In this way a 690 is a good option whereas a Grid K2 wouldn't work at all.

3) You are comparing the prices of new GeForce cards with second hand Grid/Tesla/Quadro cards. This is non-sensical since nobody here is likely to be buying a new card just so they can immediately void the warranty on it. Especially when a second hand one that will do just as well can be had with a hefty discount. You really need to think properly about who the target audience here is.

4) Quadro 6000 (and 5000) does in fact work just with vSGA. There have been numerous reports of a demo system at a show a couple of years ago running up to 6 thin clients using HorizonView and vSGA, running Borderlands at 800x600@25fps on each client, all off a single Quadro 6000 in a single ESXi host. A GTX480 modified into a Q6000 would actually fare about 30% better under that load because it has both more shaders and a higher clock speed.
 

Offline echen1024

  • Super Contributor
  • ***
  • Posts: 1660
  • Country: us
  • 15 yo Future EE
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #762 on: November 27, 2013, 04:39:49 pm »
Do you think one of these hacked Quadros could drive IBM's T221 3840x2400 monitors?
I'm not saying we should kill all stupid people. I'm just saying that we should remove all product safety labels and let natural selection do its work.

https://www.youtube.com/user/echen1024
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #763 on: November 27, 2013, 05:00:14 pm »
Do you think one of these hacked Quadros could drive IBM's T221 3840x2400 monitors?

I do indeed have a pair of T221 monitors on my desk, so I can promise you that every Nvidia card with a sufficient number of DVI connectors on the back of it will drive them absolutely fine. Having said that - my Nvidia cards have been driving them just fine before they were modified, too. If you have a DL-DVI adapter for the T221 (DG5/DGP only), you can run it off a single card with 2xDL-DVI.

You'll find a number of articles on the T221 and what works and what doesn't here:
http://www.altechnative.net/tag/t221/
I am writing another update on the T221 for use with Xorg that I'll push out shortly, about defeating xrandr in favour of xinerama (xrandr is a braindead abomination that seems to fundamentally prevent desktop stretching across multiple monitors).

ATI, OTOH, are utterly incapable of handling those monitors. There are no ATI cards newer than HD4xxx series that have more than one DL-DVI port (the other port looks like DL but is only SL). And even on 4xxx series there are major, major issues with drivers more recent than circa 2008 vintage, that make them not work. Worse, even those ancient drivers that do work have an issue where randomly upon a reboot, the monitors come up swapped around (which makes a T221 completely unusable).
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #764 on: November 27, 2013, 05:36:58 pm »
gordan, let me know if you want me to replicate your environment, might be useful to mirror setup (XEN) to see where things diverge..
« Last Edit: November 27, 2013, 06:23:03 pm by mrkrad »
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #765 on: November 27, 2013, 06:24:16 pm »
Do you think one of these hacked Quadros could drive IBM's T221 3840x2400 monitors?

things are bout to change. Let's say big monitors and fruit. soon. that resolution will be everyday ordinary very soon.
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #766 on: November 27, 2013, 06:57:54 pm »
Do you think one of these hacked Quadros could drive IBM's T221 3840x2400 monitors?

things are bout to change. Let's say big monitors and fruit. soon. that resolution will be everyday ordinary very soon.

I'm not going to hold my breath. People have been saying that since 2003 and it has failed to materialize. There is still no other monitor that will do 3840x2400. The closest is the new Asus that does 3840x2160, and that's 31", which is too big IMO - too much looking side to side. 24" is about the limit of what I'd consider nowdays (having had a 30" Dell before the T221s).
 

Offline Luke212

  • Contributor
  • Posts: 5
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #767 on: November 27, 2013, 10:19:02 pm »

I'm not going to hold my breath. People have been saying that since 2003 and it has failed to materialize. There is still no other monitor that will do 3840x2400. The closest is the new Asus that does 3840x2160, and that's 31", which is too big IMO - too much looking side to side. 24" is about the limit of what I'd consider nowdays (having had a 30" Dell before the T221s).

Dell released a 32" 3840x2160 which is the same panel as the Asus. It's quite expensive. I do think we will start to see a shift as the push to 4K televisions is in it's infancy. But it will take a while. I am more interested in VR these days. I think Oculus has made some improvements since the Rift, and it feels to me this is the way to go.  Also thanks for your summary of your Titan hacks. I look forward to progress in this area.

 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #768 on: November 27, 2013, 10:51:47 pm »
Have you noticed the grid K2 and it's USM's pci-id is very close?  I think you might be onto something when you said software.  Perhaps the card is dispatching nearby pci functions to enable the grid vgx licensing?

I'm going to get some dinner and take a close look at the bitmask  pattern of K2-> K2 USM and K1-> K1 USM , perhaps there is a pattern to this madness.

For example for "Grid K2"
- Physical Function PCIid: 10ed:11bf - GK104GL [GRID K2]
- Virtual Function selectable from
PCIid: 10ed:118b - GK104 [GeForce K2 USM]
PCIid: 10ed:118c - GK104 [NVS K2 USM]
PCIid: 10ed: 11b0 - GK104GL [Quadro K2 USM]
PCIid: 10ed:11b1 - GK104GL [Tesla K2 USM]


these look very interesting. I think I have the mobile quadro in a Z1 , GTX 660 is oddly very close by :)

NVIDIA_DEV.11BE = "NVIDIA Quadro K3000M"
NVIDIA_DEV.11BF = "NVIDIA GRID K2"
NVIDIA_DEV.11C0 = "NVIDIA GeForce GTX 660 "
« Last Edit: November 27, 2013, 10:56:34 pm by mrkrad »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #769 on: November 27, 2013, 11:35:45 pm »
Have you noticed the grid K2 and it's USM's pci-id is very close?  I think you might be onto something when you said software.  Perhaps the card is dispatching nearby pci functions to enable the grid vgx licensing?

I'm going to get some dinner and take a close look at the bitmask  pattern of K2-> K2 USM and K1-> K1 USM , perhaps there is a pattern to this madness.

For example for "Grid K2"
- Physical Function PCIid: 10ed:11bf - GK104GL [GRID K2]
- Virtual Function selectable from
PCIid: 10ed:118b - GK104 [GeForce K2 USM]
PCIid: 10ed:118c - GK104 [NVS K2 USM]
PCIid: 10ed: 11b0 - GK104GL [Quadro K2 USM]
PCIid: 10ed:11b1 - GK104GL [Tesla K2 USM]

I have no idea what you are talking about. Can you elaborate? What is "USM"?
 

Offline mrkrad

  • Contributor
  • Posts: 37
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #770 on: November 28, 2013, 12:10:28 am »
best to digest this way:

http://bramwolfs.com/tag/nvidia-vgx/

http://www.nvidia.com/object/user-selectable-machines.html

USM = mode of emulation. Makes you wonder if this is all licensing related?

http://on-demand.gputechconf.com/siggraph/2013/presentation/SG3112-GRID-State-Art-Virtualized-Graphics.pdf

check this out: for windows 7 64bit ->     NVIDIA GRID vGPU Software Release 331 WHQL NVIDIA Recommended    331.82    25.11.2013
Release Highlights:

The release package includes both Windows Display Driver (331.82) and  :-+ GRID vGPU Manager (331.24)  :-+

This software package enables NVIDIA GRID vGPU providing shared hardware based graphics acceleration for Windows VMs running on Citrix XenServer 6.2.

Note: If you do not run NVIDIA GRID vGPU, NVIDIA recommends using driver version 320.86 or higher.

Existing Support:

    Support for OpenGL 4.4
    Support for DirectX11.1

« Last Edit: November 28, 2013, 12:27:28 am by mrkrad »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #771 on: November 28, 2013, 07:34:20 am »
I see - so these are device IDs of different GPUs that the K2 can "emulate" and expose to the domU, which the Nvidia driver knows about. Hmm... I wonder if it would work if a GK104 card were modified to one of those emulated device IDs.

Edit: Something just occured to me - what is the bet that this works similarly to the hard vs. soft strap?
Recap: You have to modify the hard-strap for Kepler cards to work virtualized - soft-strapping doesn't enable the feature, and viewing the device ID using Geeks3D GPU Caps viewer shows the hard strapped ID rather than effective PCI ID. If the hard-strap isn't something virtualizable, virtualization doesn't work.

Ergo, it would stand to reason that the emulated ID is within 5 least significant bits of variance of the hard-strapped ID.
This also explains why my GTS 680 hard-strapped to Grid K2 worked fine when soft-strapped to K5000. I'll be that I could have soft-strapped it to one of the USM IDs and it would have worked just fine.

What this implies is that if you hard-strap to K2, you can then just load a differently strapped BIOS (e.g. via nouveau BIOS loader) before passing to the VM and the VM would see the card according to the soft-straps.

So they licence you for loading a specific BIOS to re-strap the card.

Clearly there is more to it because there is GPU sharing involved as well, so there is some extra abstraction and arbitration going on, but I think this might sort of the the core idea behind it. It also explains why the soft-straps for IDs are still there in the new BIOSes (with new straps in new places) even though we have only seen hard-straps to be effective at achieving virtualization capabilities.
« Last Edit: November 28, 2013, 11:19:57 am by gordan »
 

Offline h1386343

  • Regular Contributor
  • *
  • Posts: 55
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #772 on: November 28, 2013, 10:00:39 am »
All this just to change name in windows without any performance benefit in professional apps, look at this charts!







If anyone interested in SPECviewperf, here the download links:
SPECviewperf 11 32bits - http://spec.it.miami.edu/downloads/gpc/opc/viewperf/SPECviewperf11win32.exe
SPECviewperf 11 64bits - http://spec.it.miami.edu/downloads/gpc/opc/viewperf/SPECviewperf11win64.exe


« Last Edit: November 28, 2013, 10:05:38 am by h1386343 »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #773 on: November 28, 2013, 10:27:28 am »
All this just to change name in windows without any performance benefit in professional apps, look at this charts!

Are you trolling? Or working for Nvidia's damage control team?

No, the modifications in question do not enable the handful of GL primitives missing on non-Quadro cards. The performance is also deliberately crippled in the drivers - those primitives are emulated extremely inefficiently. As the nouveau Linux drivers get more mature (applications which SPECviewperf measures the performance of also run on Linux), there is a good chance the SPECviewperf performance figures will actually be better on a GeForce with nouveau drivers than with Nvidia's binary drivers. Perhaps I'll put a recent Fedora on a machine and test whether this has already, in fact, happened.

What this thread always was about is enabling other features that are purely up to the driver, such as whether a card will work with VGA passthrough/vDGA and vSGA for virtualization purposes. Vast majority of us here don't give half a damn about SPECviewperf figures (but I will still laugh my guts out if it turns out GeForce + nouveau yields better SPEC scores than GeForce + nvidia). What we do care about is decent surround monitor support (gnif's original motivation) and even more so about decent virtualization support - both of which are purely down to the driver deciding to arbitrarily not enable the feature set on GeForce cards because the owner hasn't paid enough for the privilege.

Personally, I think the most awesome outcome would be for the nouveau project to just produce Windows drivers in addition to the Xorg drivers. The second most awesome outcome would be for Nvidia to not disable said features in the drivers.
« Last Edit: November 28, 2013, 05:18:03 pm by gordan »
 

Offline gordan

  • Frequent Contributor
  • **
  • Posts: 277
  • Country: 00
    • Tech Articles
Re: [MOVED] Hacking NVidia Cards into their Professional Counterparts
« Reply #774 on: November 28, 2013, 10:53:56 am »
Update:
As some of you that have been following this thread are aware, I've been having a rather bizzare problem with only DL-DVI working on my modifed GK104 based cards (680 and 690). It just occurred to me that the two cards I have happen to have on thing in common - they are both Gainward cards. Has anyone else successfully managed to get DL-DVI to work in a VM with a Gainward GTX680 Phantom 4GB or Gainward GTX690? Gainward 690s have a few strapping resistors in different places to the EVGA and other 690s, so it is plausible they made some tweaks that cause the problem. Does anyone have either of those cards working with DL-DVI outputs successfully?

On a separate note, I just learned that GTX780Ti has device ID 0x100A. K6000 is 0x103A. We only need to change the 3rd nibble (via oguz286's awesome mod). His mod is to use a 33K resistor to jack up the 3rd nibble from 0 to 2. On my Titan using an 18K resistor instead boosts the ID to 3. So to make a 780Ti to a K6000, simply apply an 18K resistor between VCC and SCLK and voila, job done. For extra points, solder a couple of wires between those pins, solder an 18K resistor on one of them, and a switch to connect them. Break the switch out somewhere accessible (extra extra points for making the switch easily and neatly accessible from the back of the card without obstructing the airflow too badly). Now you can switch between a 780Ti and a K6000 at a flip of a single switch.

Needless to say, the neatness of this means that I'll be acquiring a 780Ti as soon as I've ebayed my Titan.

Edit: Maybe I'll keep my Titan for a little bit longer, as if the PCB is the same (which it almost certainly is), it should be very obvious which resistors are configured differently for the 4th nibble. Watch this space.
« Last Edit: November 28, 2013, 05:20:19 pm by gordan »
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf