(Working) Tesla K40m in a 32Bit BAR Motherboard.

[Poet129]

@ShrimpBrime

Issue 3 is addressed in windows 10 build 20190.

Still working on having to reflash the v-bios.

[Guest]

8 minutes ago, Poet129 said:

@ShrimpBrime

Issue 3 is addressed in windows 10 build 20190.

Still working on having to reflash the v-bios.

That's awesome. 

[Poet129]

On 4/25/2021 at 11:59 AM, ShrimpBrime said:

That's awesome. 

More awesomeness... not sure what I did tbh, but for whatever reason it survived a reboot. The only change I recall making was enabling memory integrity in windows defender.

Edit:

It has also survived a shutdown.

[Guest]

9 minutes ago, Poet129 said:

More awesomeness... not sure what I did tbh, but for whatever reason it survived a reboot. The only change I recall making was enabling memory integrity in windows defender.

Edit:

It has also survived a shutdown.

That's Brilliant. This thread is bookmarked for later  

[Poet129]

16 minutes ago, ShrimpBrime said:

That's Brilliant. This thread is bookmarked for later  

How do say this... I stand corrected... The driver shows up without any errors after a reboot / shutdown however it doesn't function... I suppose this is still a step in the right direction however, disappointing.

[Guest]

41 minutes ago, Poet129 said:

How do say this... I stand corrected... The driver shows up without any errors after a reboot / shutdown however it doesn't function... I suppose this is still a step in the right direction however, disappointing.

Well you're making headway, but no disappointment. I couldn't figure it out at all  

[Poet129]

On 4/26/2021 at 10:23 PM, ShrimpBrime said:

Well you're making headway, but no disappointment. I couldn't figure it out at all  

So the next reasonable step was to pull the v-bios from the card after getting it to work. This confused me a little... So the v-bios with the card enabled in device manager is completely different at the end than the v-bios pulled with the card disabled in device manager. So then I checked for consistency... I'm finding none even just writing to the disk and then verifying is giving different results at this point. I'm lost. Also, sorry I haven't posted in a while I sprained my wrist in a track and field meet Tuesday, feeling a little better now.

Edit:

I've found over 7 different v-bios' I think there is something wrong.

[Guest]

53 minutes ago, Poet129 said:

So the next reasonable step was to pull the v-bios from the card after getting it to work. This confused me a little... So the v-bios with the card enabled in device manager is completely different at the end than the v-bios pulled with the card disabled in device manager. So then I checked for consistency... I'm finding none even just writing to the disk and then verifying is giving different results at this point. I'm lost. Also, sorry I haven't posted in a while I sprained my wrist in a track and field meet Tuesday, feeling a little better now.

Edit:

I've found over 7 different v-bios' I think there is something wrong.

Lol. Sounds like one hell of a battle. 

 

I'd imagine it should work even on the stock bios. It's the board you're trying to trick, not the video card

[Poet129]

9 minutes ago, ShrimpBrime said:

Lol. Sounds like one hell of a battle. 

 

I'd imagine it should work even on the stock bios. It's the board you're trying to trick, not the video card

The original plan was to trick the board into thinking it could use 64 bit BAR, but what I've managed to do is more based on tricking the card into thinking it is the c variant of the card rather than the m variant of the K40. This may be wrong, but I think what is happening is it is getting some of the information mixed up when reading the card's v-bios due to the fact that it is only able to communicate with the first 32 bits, so some of the 64 bits of the BAR may be interfering with the v-bios getting read. This is a huge stretch, but because I have no clue what's going on this makes the most sense.

[Guest]

23 minutes ago, Poet129 said:

The original plan was to trick the board into thinking it could use 64 bit BAR, but what I've managed to do is more based on tricking the card into thinking it is the c variant of the card rather than the m variant of the K40. This may be wrong, but I think what is happening is it is getting some of the information mixed up when reading the card's v-bios due to the fact that it is only able to communicate with the first 32 bits, so some of the 64 bits of the BAR may be interfering with the v-bios getting read. This is a huge stretch, but because I have no clue what's going on this makes the most sense.

Is the memory the same make on the M and C varients? Timings ect?

[Poet129]

5 minutes ago, ShrimpBrime said:

Is the memory the same make on the M and C varients? Timings ect?

No clue... But this is how I've gotten the card to work thus far and there aren't any graphical errors or random crashes.

[Guest]

1 minute ago, Poet129 said:

No clue... But this is how I've gotten the card to work thus far and there aren't any graphical errors or random crashes.

If the memory is different, the card can become magical

yes still operate in some cases, but with some unknown side effects. 

[Poet129]

1 minute ago, ShrimpBrime said:

If the memory is different, the card can become magical

yes still operate in some cases, but with some unknown side effects. 

They appear to be the same on techpowerup.com in gpu specs... something interesting to note though is that the DirectX versions for the cards are different. The K40c has 11.0, while the K40m has 11.1. When the card is functioning it shows with 11.0 in gpuz.

[Poet129]

Nothing crazy... fixed overclocking issues.

[Poet129]

Further developments... for some reason before I reflash the v-bios back and forth it shows as PCIE 3.0, which this board only has 2.0, more interesting is that it goes down to 2.0 after it starts functioning. Perhaps the flashing of the v-bios is just making it recheck what pcie speed it can run at... then it turns it down to match the board... maybe, still working on a permanent fix nothing yet.

[Guest]

2 hours ago, Poet129 said:

Further developments... for some reason before I reflash the v-bios back and forth it shows as PCIE 3.0, which this board only has 2.0, more interesting is that it goes down to 2.0 after it starts functioning. Perhaps the flashing of the v-bios is just making it recheck what pcie speed it can run at... then it turns it down to match the board... maybe, still working on a permanent fix nothing yet.

That sounds normal to me for some reason....

[tdkid]

as far as i can tell. the EVGA resizable bar would only be for the EVGA z590 motherboard and 30 series GPUs because it was only released recently. EVGA does do a livestream on twitch on Tues., Thurs., and Friday so you can ask them directly.

[Poet129]

Nothing new to add, except some performance metrics...

8K World of Warcraft:

1050ti - 1 FPS

Tesla K40m - 10 FPS

This appears to be due to running out of memory on the 1050ti.

[Guest]

5 minutes ago, Poet129 said:

Nothing new to add, except some performance metrics...

8K World of Warcraft:

1050ti - 1 FPS

Tesla K40m - 10 FPS

This appears to be due to running out of memory on the 1050ti.

Yeah, that would do it lol. K40 10fps, not terrible. Hows 1080 lookin in comparison?

[Poet129]

22 hours ago, ShrimpBrime said:

Yeah, that would do it lol. K40 10fps, not terrible. Hows 1080 lookin in comparison?

No 1080 here, sorry.

[Poet129]

Nothing new for a permanent fix, however I made a program that can "SLI" two GPUs in DirectX 11. (Note this won't effect any other applications, I should say use two GPUs together for a single application that is C++ and uses DirectX 11.)

Here are my results:

 

1280 Objects:
1050ti + K40m: 1178 FPS
1050ti:                941 FPS
K40m:                 625 FPS
512 Objects:
1050ti + K40m: 1823 FPS
1050ti:               1772 FPS
K40m:                 930 FPS

 

Here is the Source.

[Guest]

6 hours ago, Poet129 said:

Nothing new for a permanent fix, however I made a program that can "SLI" two GPUs in DirectX 11. (Note this won't effect any other applications, I should say use two GPUs together for a single application that is C++ and uses DirectX 11.)

Here are my results:

 

1280 Objects:
1050ti + K40m: 1178 FPS
1050ti:                941 FPS
K40m:                 625 FPS
512 Objects:
1050ti + K40m: 1823 FPS
1050ti:               1772 FPS
K40m:                 930 FPS

 

Here is the Source.

That's just spectacular effort on your part. Just the fact you got it working partially is amazing. 

 

The SLI with a 1050ti shows decent results too. 

 

Are you using NVInspector to enable the SLI or just the NV control panel?

[Poet129]

1 hour ago, ShrimpBrime said:

Are you using NVInspector to enable the SLI or just the NV control panel?

Neither, that is why it won't work for other applications, which is why I provided my source code. That way once I made the "SLI" better it could be implemented by other DirectX 11 users, however that will likely take a while before it is adopted.

1 hour ago, ShrimpBrime said:

That's just spectacular effort on your part. Just the fact you got it working partially is amazing. 

 

The SLI with a 1050ti shows decent results too. 

The results are actually going to better I'll post an update when I get them, I did an update really late last night so I didn't get to run the results, however I did with the 1050ti and a emulated card from the CPU.

Which gave me this:

 

1280 Objects:

1050ti + CPU: ~1300 FPS

1050ti:             ~1000 FPS

CPU:                   ~40 FPS

 

I'm not sure exactly what happened there, but I'll take it. I'm working on the retest results for the 1050ti + K40m. I'll post back the results when I get them.

[Guest]

1 hour ago, Poet129 said:

Neither, that is why it won't work for other applications, which is why I provided my source code. That way once I made the "SLI" better it could be implemented by other DirectX 11 users, however that will likely take a while before it is adopted.

The results are actually going to better I'll post an update when I get them, I did an update really late last night so I didn't get to run the results, however I did with the 1050ti and a emulated card from the CPU.

Which gave me this:

 

1280 Objects:

1050ti + CPU: ~1300 FPS

1050ti:             ~1000 FPS

CPU:                   ~40 FPS

 

I'm not sure exactly what happened there, but I'll take it. I'm working on the retest results for the 1050ti + K40m. I'll post back the results when I get them.

Give inspector a try, see if you can adapt any settings to improve the SLI performance.

[Poet129]

It has been a while since I posted here... just an update, no progress on maintaining gpu availability across restarts, however I've now tested with multiple nvidia drivers and Windows 11, and they all work fine.