harkaz

You could boot the system blindly from a USB drive and ensure that a boot script will launch the EC flash commands and shutdown the system upon completion. Once you verify that EC has been properly flashed by checking fans with Fn+1, clear CMOS by pressing Fn+D and keeping it pressedd while you press the power on button. EDIT: The EC Flash fw includes EFI commands (in the *.nsh script file), so use EFI shell.

Eventually, I had a spare heatsink modified by @Developer79. Very satisfied with the result - my CPU, GPU idle temps are even better than the stock heatsink was previously for the same CPU (i9-9900K) and RTX 20xx GPU. It took me some time to add the pads and properly adjust the adapters. VRAM is may not be perfectly cooled at 100% load, which leads to 105 degrees hotspot with aggressive fan curve (100% GPU fan speed above 60 degrees). This must be related to the metal protrusions of the original/modified heatsinks being in different locations than the respective VRAM chips in RTX3080. Using thermal pads cannot fully compensate for this discrepancy. For sustained high VRAM loads, attention is needed. Not sure if the cause for 105 deg hotspot is the VRAM, as the benchmarks did not indicate throttling. No VRAM errors were detected after 1 hour of OCCT testing. Some nvlddmkm errors were fixed by disabling MPO. Furmark results: EDIT: Fixed the hotspot temp issue by reapplying paste and ensuring it covers the entire GPU core before the heatsink is mounted. Now temps reach 87.4/99.8 under sustained 100% load with significantly better performance (avg. Furmark at 131 FPS, score 7829).

For your spare time, an interview that views AI as an 'evolutionary history' event (45:15 and on). It may be that, no because of its qualities per se but because of the way influential people treat it. For a technology that has a risk of quality degradation with increased training on synthetic (non-human generated) data (doi: 10.1038/s41586-024-07566-y) the prediction of its future performance seems rather difficult. Present capabilities cannot predict future capabilities. And of course, it's not reasoning, it's predicting numbers that are matched to characters. It can be useful for finding patterns and making educated guesses based on its training corpus. The decisions of powerful people to treat AI as a deity will however be evolutionary history. And not in a good way.

You have to check the signal quality on the traces with oscillograph. If the equipment is not available you cannot know with high confidence about signal quality causing data loss. You would have to take the financial risk and remove C0, D1 and re-test VRAM. The best option financially would be to get new used GPU.

@Maro97 Most likely all chips that report errors (6/8). Start with the 2 that appear to be physically damaged and report most errors then retest.

Issue fixed. Enable Above 4G decoding and Above 4GB MMIO assignment with modified 128GB BIOS below for P775TM1-G (modification tested on my own config). Sent you a PM for P870DM-G. Additional modifications are needed for RTX3080 driver detection with proper subsys and friendly name display in BIOS Setup. P7xxTM1-0729-9900KS-128G-PCIfixed.7z

Regarding RTX3080 issue with 128GB RAM, I also tried with PCI express for PGI to Gen3 and Force x8 - in order to start with PCIex8 always in P775TM1. Same behavior (no video). Not a PCI express bus issue. Most likely caused by a failure of the UEFI component of the graphics card to load with 128GB DRAM.

EDIT: Issue replicated and fixed on my P775TM1. Refer to later post.

Longevity and reliability of P775TM1 are second to none for laptops. I doubt that any current laptops do match that because of bad thermals and motherboard design. Besides, buying a new laptop with 32+GB DDR5 would be pretty expensive, so upgrading the old one with 64+GB DDR4 does make sense financially.

https://www.bbc.com/news/articles/cq635037g18o "Maybe they actually believe it," Arison said. "Or you could argue they're saying it because it's a great way to gin up more investor support, because the only way to justify these valuations is to actually claim: 'I'm going to take over every industry and I'm going to take over every job, and my AI is going to be doing all that work.'" Let me guess...

https://www.wsj.com/opinion/unregulated-open-weight-ai-is-an-invitation-to-disaster-c16c278f Open weights are bad.. really bad. Due to paywall, however, I was unable to find out if they started a pandemic after all.

Intel ME must be disabled only to address up to 128 GB of RAM - with other modification that dsanke has added. If you enable Intel ME, then up to 64GB of RAM can be addressed.

All latest versions of dsanke BIOS and EC work properly.

https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/ It's now official. 13B USD seems a little bit excessive for a git repo. Of course, it is a very large repo, but still. Apart from that, the llama.cpp inference engine is being maintained by employees of HuggingFace. We'll see how that affects open weight LLMs in the next 18-24 months, but I am not optimist about the future of open weights - if open-weight LLM development continues at this pace, OpenAI and Anthropic and probably the US economy will face serious issues: AI "safety regulations" are just a legal pretext to control the pace of open LLM development. (See also: https://www.reuters.com/world/china/bill-gates-alarmed-by-ai-has-policy-ideas-he-wants-discuss-with-chinas-xi-2026-08-26/ )

@Selim1905, @SuperMG Thanks for your suggestions.