cancel
Showing results for 
Search instead for 
Did you mean: 

R6E modded bios 4001 with ucode 02007108

Int8bldr
Level 12

@restsugavan , I think you modded the R6E BIOS 4001 with the latest published ucode for 7980XE:  cpu50654_platB7_ver02007108_2023-07-23_PRD_FE927D02.bin and also the trusted platform? right?

1. Are you happy with it your modded 4001?

2. Is your modded 4001 (ucode: 02007108) better and faster than ASUS official 4001 (that has the older ucode: 2007006) 

3. any other benefits to go modded with 4001?

In any case, I would be very happy if you could post this modded 4001 in this forum somewhere. Original thread is now closed so probably you need can just reply to this thread. 

I am generally very happy with the official ASUS 4001 which is probably the best R6E bios ever! Thank you ASUS for your commitment to the X299 platform and this community! I found that almost all of performance that was lost with all the security patches in ucodes over the last 4-5 years are mitigated with the 4001/uc:2007006, so that I am now back to early 2020 performance and much more secure! Excellent! 

Thank you @restsugavan !

759 Views
1 ACCEPTED SOLUTION

Accepted Solutions

restsugavan
Level 14

Hi @Int8bldr  😎  Here is 4001 modded BIOS specific only Skylake X CPUID 50654 with 02007108 microcode mate. I used testing it with Windows Insider Canary build without problem. Only on AVX2/AVX512 penalty performance hit that can't be solve by 02007108 microcode is only weak point.

Download Link :: https://drive.google.com/file/d/1sJPw3lzeLRGeWbyOl-On6OlXxqqEB7Pm/view?usp=sharing

Caution :: The Modded BIOS support only Skylake SP/X/W CPUID 50654  for internally test with Windows Insider Preview Canary Edion under developer agreement with Intel and Microsoft by developer only. So there are no support in case of failed boot or system damage at all. Trying by your own risk. Intel and Microsoft didn't responding for any error instability malfunction all case for any users experience.

The Downfall Vunerable still nightmare for newier architect from Intel too. Thus Lion Cove and Skymont microarchitect had been release without Hyperthreading were from these root cause indeed.  Due Hyperthreading function always call " GATHER Instruction " to perform FMA3 /AVX2/AES/AVX512/VNNI instruction processing on Intel complex FPU design.  Even Redwood Cove on Core Ultra 100 Series and Granite Rapid AP/W/XSP will be lost there peak FPU performance due " Downfall " vulnerable patch according to lastest Intel document below.

สกรีนช็อต 2024-10-08 101355.png

 All affected Intel CPU list on " Downfall " Vulnerable lastest update 2024/09/30

https://www.intel.com/content/www/us/en/developer/topic-technology/software-security-guidance/proces...

By contrast AMD FPU platform split FADD and FMUL each pipeline were luckily automatic " Downfall " mitigation by its design. Look like AMD were behind " Downfall " attackers to shooting out Intel most strength point on it CPU design aka " Intel FPU's FMA3 Unit "

Xeon-E5-2600-V3-AVX2-FMA (1).jpg

In the past Intel winning over AMD on Intel FMA3 that use 3-clock latency to perform FMA instruction AMD loss with there FMA4 that use more than Intel 1 clock latency.  Today AMD strike back with " Downfall " vulnerables that cause Intel FPU performance penalty hit by 30-50%. So It hard to believe that AMD wasn't stand behind or be a partner of " Downfall " attacks.

On Lion Cove and Skymont Design , Intel had learning more their loss on " Downfall " The FPU of both Lion Cove Core and Skymont core had been change and disable Hyperthraeding to avoid " GATHER " Instruction call.

Lion Cove FPU had been redesign to prevent all folks above from AMD and " Downfall " teams attack again. Intel was split there FPU into ZEN styles. If it has been attacking by " Downfall " team AMD CPU will be impact too. It was the answer from Intel Oregon team shortly.

So no more " GATHER " instruction need and no more " Downfall " on Lion Cove at these point. Thus

" No performance Penalty hit anymore No microcode work around No more complaint "

2024-06-04_8-51-24.jpg

 Skymont also complied same target reduce using " GATHER " Instruction by split it FPU design like Lion Cove too.2024-06-04_8-50-34.jpg

 That's all folks today. Thank you ASUS ROG Forum teams and every X299 friends who still running today.

 

W11 25H2 27754.1000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 566.14 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

View solution in original post

9 REPLIES 9

restsugavan
Level 14

Hi @Int8bldr  😍

1. Yes! I'm happy with Intel engineer Skylake X team best trying. They can't bring back AVX2/AVX512 Performance hit on " Downfall " mitigation however they had reduce Core to Core latency on mesh interconnect node as best as they can do. So when we utilized the 020007108 microcode with 4001 BIOS. You'll see the core scalable ratio were increse from 02007006. 

GNRPEROK1.jpg

 Setup method " Load Optimized Default " , XMP Overclock Profile DDR4 3200 16-18-18-36 , ASUS Multicore Enhancement AUTO "

2. The modded BIOS contain recently module as update as can be for Skylake X (CPUID 50654 Specifically )

R6E4001MODDEDA6.jpg

 3. Look like Intel engineer trying there best on Core-to-Core latency deductive. 020007108 microcode provide better multicore scalable enhancement. But those 020007006 and 020007108 didn't bring back Skylake X AVX2/AVX2 Performance yet. Skylake X loss around 31.42% on AVX512 and 48.88% on AVX2 Performance when testing with CPU-Z 19.0.1.64 (AVX2 Beta) and CPU-Z 19.0.1.64(AVX512 Beta) .  Thus when testing with Geekbench 6 Cinebench R23/R24 ( used of heavily AVX2 instruction our Skylake X / Cascade Lake X look like the drunk boy inferior even old school ARM64 ) From best static I'd collect. 

CPUZSCOREA1.jpg

 

When you compare the three picture above ( W11 25H2 / W11 24H2 / W10 Last Dev build.) you'll notice that the multicore scalable on newier microcode always better on CPU-Z 17.01.64 that mixing X87/MMX/SSE4.2 instrucions measurement. The 02007108 can be archived 22.00+ scalable ratio which higher than old school microcode. In contrast , on AVX2/AVX512 performance 02007108 and 02007006 microcode perform worse when compare with older microcode.

 

 

 

W11 25H2 27754.1000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 566.14 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

@restsugavan ,

Thank you! Very interesting and valuable analysis! I have the same experience with 4001/02007006. But I have not done all the detailed testing that you have. I've also found that in order to bring back AVX512 performance closer to pre " Downfall ", you need to overclock quite a bit and play with AVX512 and AVX2 offsets much more. Have also found that the "weak" cores (I mean the 5 "least" favored cores) have to be dialed back in OC core rations quite a bit 42 ->39 even down to 37 for the weakest - not good (It is only 5 cores but still). The topmost favored cores can be OCed as before the rest need to be dialed back x1. But it seems like the weak cores get hit much harder in 4001/02007006 - they overheat very quickly with higher frequence, and I need to use a negative voltage offset to dial down the voltage auto scaling for the weak cores (they seem to over-voltage at 3.7ghz, causing excessive temps). I too cannot reach pre-"downfall" performance especially in AVX 512 workloads. Still 4001/02007006 is a good attempt at fixing "downfall". With all the security fixes included so that is still very impressive work from the ASUS BIOS dev team and Intel's uCode dev team. I've tested most new BIOS/uCodes the last 4 years, but decided every time to stay on my older BIOS/old uCode that gave me the best trade off (reasonably secure but performant). 4001 is the first one I'm considering to move to as production. 

With 4001/02007006, I get a cpu-z Multi Thread Ratio of around: 20.5 average of 10 runs in clean testing state (minimum other processes going on) testing over two days - it is pretty stable at 20-21 for years - for BIOS 3105 was also 20.5. So an uplift to MTR = 22 with 4001/020007108 would be welcome!

Tested cinnebench 15 and 20 similar outcome to yours...

Seems like uCode 020007108 is better.

Would you mind to please post your modded 4001/020007108 bios here? I do not have it.

Thank you @restsugavan 

 

Int8bldr
Level 12

as you show in your screen shots the problem with 4001/02007006 and 4001/02007108 is how unacceptably bad avx2 and avx-512 performance degradation is. At first testing, which I did, 4001 is the best release to date in mixed workloads and CPU-z benchmarking looks good. But the AVX-2/-512 numbers are just too bad. I cannot work with that. After reverting back to 3605 I am again having a STABLE CPU with COOLER running cores (the bad cores in particular) with HIGHER performance numbers - similar to your windows 10 bios 3403/6B05 CPU-Z numbers above.  I think I will not upgrade to 4001/02007006. I need stability, performant and cool running cores in avx-2/-512 first priority for me. Hypothetical security concerns are secondary.

Thank you!

restsugavan
Level 14

Hi @Int8bldr  😎  Here is 4001 modded BIOS specific only Skylake X CPUID 50654 with 02007108 microcode mate. I used testing it with Windows Insider Canary build without problem. Only on AVX2/AVX512 penalty performance hit that can't be solve by 02007108 microcode is only weak point.

Download Link :: https://drive.google.com/file/d/1sJPw3lzeLRGeWbyOl-On6OlXxqqEB7Pm/view?usp=sharing

Caution :: The Modded BIOS support only Skylake SP/X/W CPUID 50654  for internally test with Windows Insider Preview Canary Edion under developer agreement with Intel and Microsoft by developer only. So there are no support in case of failed boot or system damage at all. Trying by your own risk. Intel and Microsoft didn't responding for any error instability malfunction all case for any users experience.

The Downfall Vunerable still nightmare for newier architect from Intel too. Thus Lion Cove and Skymont microarchitect had been release without Hyperthreading were from these root cause indeed.  Due Hyperthreading function always call " GATHER Instruction " to perform FMA3 /AVX2/AES/AVX512/VNNI instruction processing on Intel complex FPU design.  Even Redwood Cove on Core Ultra 100 Series and Granite Rapid AP/W/XSP will be lost there peak FPU performance due " Downfall " vulnerable patch according to lastest Intel document below.

สกรีนช็อต 2024-10-08 101355.png

 All affected Intel CPU list on " Downfall " Vulnerable lastest update 2024/09/30

https://www.intel.com/content/www/us/en/developer/topic-technology/software-security-guidance/proces...

By contrast AMD FPU platform split FADD and FMUL each pipeline were luckily automatic " Downfall " mitigation by its design. Look like AMD were behind " Downfall " attackers to shooting out Intel most strength point on it CPU design aka " Intel FPU's FMA3 Unit "

Xeon-E5-2600-V3-AVX2-FMA (1).jpg

In the past Intel winning over AMD on Intel FMA3 that use 3-clock latency to perform FMA instruction AMD loss with there FMA4 that use more than Intel 1 clock latency.  Today AMD strike back with " Downfall " vulnerables that cause Intel FPU performance penalty hit by 30-50%. So It hard to believe that AMD wasn't stand behind or be a partner of " Downfall " attacks.

On Lion Cove and Skymont Design , Intel had learning more their loss on " Downfall " The FPU of both Lion Cove Core and Skymont core had been change and disable Hyperthraeding to avoid " GATHER " Instruction call.

Lion Cove FPU had been redesign to prevent all folks above from AMD and " Downfall " teams attack again. Intel was split there FPU into ZEN styles. If it has been attacking by " Downfall " team AMD CPU will be impact too. It was the answer from Intel Oregon team shortly.

So no more " GATHER " instruction need and no more " Downfall " on Lion Cove at these point. Thus

" No performance Penalty hit anymore No microcode work around No more complaint "

2024-06-04_8-51-24.jpg

 Skymont also complied same target reduce using " GATHER " Instruction by split it FPU design like Lion Cove too.2024-06-04_8-50-34.jpg

 That's all folks today. Thank you ASUS ROG Forum teams and every X299 friends who still running today.

 

W11 25H2 27754.1000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 566.14 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

@restsugavan , thank you for the modded BIOS and the analysis!

I think I now understand what intel means with gather instructions being the issue with Downfall. It also sounds like Intel have no way of fixing it in ucode, since the hyperthreading and gather behavior is too core to the FMA and hard wired in silicon.  With "downfall", the core engine of AVX2 and AVX-512 (which is the FMA) is nuked. And the core value proposition of using AVX2 and/or AVX-512 and Intel for heavy compute (e.g. AI) is also nuked. It's unacceptable to lose 30%-48% in performance with these ucode fixes since gather is so frequently used.

Sadly, it also sounds like this also affect Granite Rapids? not good...

However, to execute a successful "Downfall" attack the attacker needs to have local physical access to your machine so you can actually solve "downfall" with physical security. 

So, to me personally, it will mean that I will accept the risk of some idiot gaining local access on my machine (I mean what is the risk in reality) to place a "downfall" virus on my pc. I will therefor revert back to older performant uCodes. I cannot accept 30%-48% drop in performance in AVX2/AVX-512. It is not working for me.

Also, and equally important, in my testing with the new 4001 bios I saw weaker cores getting overheated at OC clock frequence that were fine with older BIOS/uCodes but now I have to down clock weaker cores from multiplier 41 to 39 and even 35 for the weakest core. And even after that, I still had very high temps (using cinebench 15,20,24 and quick cpu-z benchmarking/test tool). That is also not working me. I'm concerned, I might damage the CPU if I continue with my OC setting. I am not sure what is safe for weaker cores anymore. It's like I have to go through and redo all OC testing (I did years ago) to find the new best OC settings by core. Seems it's better for me to go back to  pre-"Downfall" BIOS/uCode with known, tested and proven safe OC settings (years to running stable and cool).

Thank you for your information and insights! They are very valuable!

 

Int8bldr
Level 12

Ok I did some more testing, this time on a 10980XE BIOS 1808/ latest ucode: 3707

and the new BIOS/uCode is catastrophic!

In my test I am training a neural net for 30 EPOCs and measuring the time.

the model makes heavy use of AVX-512 since it is (or at least was) a very effective way of crunching numbers fast (FMA is core to this).

old BIOS: 1403/uCode: 3003: model finish in 1minute 17seconds

new BIOS: 1808/ucode: 3707: model finish in 2minute 58seconds

Summary:

BIOS:1808/ucode: 3707 takes 131% more time to complete the same task (all else being equal) or

BIOS:1808/ucode: 3707 only achieve 43% of the BIOS:1403/uCode: 3003 performances

Conclusion: hard NO to "upgrade". I am considering downgrading to even earlier BIOS and uCodes.

Basically, the  uCodes from the last few years are unacceptably bad for me in my cases which are heavily dependent on effective AVX-512 - it's so very important.

 

PanosXidis24
Level 11

i told everyone here the best choice is grap latest bios and modding the oldest cpu microcode

@PanosXidis24 

Do you forget to tell everyone the trick you've use here .😋

Rename " Intel microcode dynamic link file to AMD microcode dynamic link file "  and rename " AMD microcode dynamic link files to Intel microcode dynamic link files " first . To avoid Windows 11 24H2 or higher get blue screen at boot before test. 

Do you forget to tell everyone the trick you've use here .😋

rename mcupdate_genuineintel.dll to mcupdate_authenticamd.dll

rename mcupdate_authenticamd.dll to mcupdate_genuineintel.dll

😋 Aren't you ?

https://www.reddit.com/r/overclocking/comments/1d28a8y/windows_11_24h2_and_mcupdate_microcode/

W11 25H2 27754.1000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 566.14 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

PanosXidis24
Level 11

@restsugavan  no im not post on reddit