I have seen this a few weeks ago only problem is it don't support gtx 980 ti just wish the author can edit the file to support this as I would try this again just to test the gtx 980 ti for core issues while the E1 and F0 bank temporary disabled
About 980ti: Unfortunately, Maxwell2 GPUs are not supported. The "devinit script" that is edited by utility do present in a same form in all cards from Maxwell1 to Turing.
However, only GTX750Ti Maxwell1 boots fine after modding it. When such edited VBIOS is flashed onto the Maxwell2 cards - they just go black screen.
It seems that starting with Maxwell2 the "devinit script" part of VBIOS is somehow verified.
If you are ready to some "flash original VBIOS back" quest - you can even perform such experiment yourself.
Save your 980ti VBIOS, run the utility and use the "Open original VBIOS file" button to generate VBIOS variants
It will create several VBIOS files in the subfolder (
similar to this example .zip for GTX Titan 6GB). Then use the nvflash.exe from details subdirectory to flash DisableE variant - that nvflash.exe is capable to flash modified VBIOS on Maxwell2 GPUs.
Then reboot - I expect that 980ti will not show any picture after such modding. And note: depending on the motherboard it may be some quest to flash original VBIOS on it again - the MB need the capability to force use builtin iGPU boot with monitor plugged in even when the buggy 980ti is attached too. Some advices and typical Motherboard parameters are described in
troubleshooting section of the utility guide.
Unrelated, but does OP know a way to unlock "disabled" CUDA Cores like Dell/Alienware this
The count of CUDA-cores physically present on a die is identical for all chips from some family (like GA104). However, for several reasons (including the non-working cores) some of them are disabled - this is performed by disabling a whole TPC. The disabling is done in two steps:
- Some TPCs are disabled at factory, at the same moment when the chip DeviceId is assigned. So, as far as I understand, the cores disabled at factory should be identical for all GPUs having same first half of device ID (10DE 24DD for mentioned Alienware M15 R5). Also, typically each device ID corresponds to a physically visible GPU marking, like GA104-770-A1 for 10DE 24DD. This marking is effectively a part number, having identical DeviceId and identical count of factory-disabled TPCs.
While the count of disabled TPCs is identical, and each TPC having identical number of Cuda-cores - for some GPUs the performance may slightly differ depending on how they are disabled! I verified that on the Pascal family of GPUs, but I think that it may be a case for later GPUs too. The difference is caused by the fact that while having identical sum number of TPCs, the enabled ones may have different distribution between GPCs - sometimes GPCs have nearly identical number of TPCs enabled, and sometimes some GPCs are all TPCs enabled, while others have only single TPC active. The performance difference is not critical, but it was easily noticed 2-4% while comparing the crypto-mining performance of identical P102-100 GPUs running on identical fixed clocks. I suppose it may be caused by overloading/underloading of some internal GPU buses.
- And some more TPCs (or even a whole GPC) can be disabled at the VBIOS level (for a single boot). This was the case for the "94.04.42.00.97" VBIOS; the subvendor ID - second half of device ID (1028 0A97) is set at VBIOS level too.
As far as I know, VBIOS-level disabling is quite rare for Nvidia, and only such cores can be re-enabled.
I had zero practice with laptops, so the text below is just some theory, without being proved by my own practice.
Unfortunately I don't know a way to edit existing 30xx series VBIOS and getting it accepted by the GPU. But generally, flashing the VBIOS with same device ID (10DE 24DD) should be possible (with some modded nvflash or a hardware programmer), and, say 20%-to 50% of laptop VBIOSes with same IDs in average are compatible. The other 50-80% are not,
so before such experimenting with laptop GPU - you should be ready for flashing saved original VBIOS back with a hardware programmer. And there maybe some other laptop-specific issues I'm not familiar with - so as always - flashing non-factory VBIOSes is at your own risk.
So, if you see that some laptop with 10DE 24DD XXXX XXXX GPU having less Cuda-cores than another laptop with 10DE 24DD XXXX XXXX GPU (XXXX may differ) you may try to flash another VBIOS on the laptop with lower Cuda-cores. The M15 R5 has exactly that case - with "94.04.42.00.97" VBIOS it showed 4608 Cuda-cores
https://www.techpowerup.com/img/YOLmOeQrWxb2A38y.jpg while the
GA104 description on Techpowerup tells that GA104-770-A1 should have 5120 cores.
All available VBIOS variants for a specified DevicedID can be searched on Techpowerup by a manually created URLs: (this is a really useful feature, great thanks to Techpowerup!)
Unfortunately, the VBIOS page doesn't provide info that allows to estimate the effective count CUDA cores and the effective compatibility. So some researchers just flash then one by one to determine the actual behavior.