I.nfraR.ed

Maybe, just maybe, the CPU is also a factor? ?

I don't think MSI would be able to help. The documents, describing all the SMU features and MSR are under NDA. AMD never released "Kernel and Bios Development Guide" for Zen, while it is available for older generations. I don't have any industry contacts, but even if I manage to sign NDA and get my hands on these documents, it still wouldn't work for me, because "ZenStates" is now open-source. Signing NDA would mean I will have to close it again, which I don't want. So, basically, this leaves me with the only option to reverse-engineer what I can. More about my request. Performance Bias options set different bits in different MSRs. The easiest way to "copy" these and re-use in ZenStates is to read the MSRs from a running system, then compare changes and extract just the bits of interest, which I don't actually know what they are. I only have an uneducated guess. The other option would be to disassemble those binary (.exe) files. Since you have pointed me to the VIII thread thread, I will get the binaries from there and test on my machine instead. I'm using public tools like R-W Everything, CrystalCPUID. Cheers.

Thanks, I don't have a MSI board, only have Crosshair VI Hero. I've tweaked these profiles for ZenStates app, but wanted to see if MSI does something different. That's why I need to read MSR on a running system or somehow get the source code, which ain't gonna happen. Aggressive doesn't boot for me on the CH6H, but I was not aware there are updated exe files. Will take a look, thank you. Having to reverse-engineer everything is very demotivating

All my Sempron CPUs scale equally with cold in terms of HTRef, but I haven't tested many, so that might not be valid for all. I have 2 untested 150's, but can't find time for all the projects I have, so they'll wait. Nowadays, you need a high clocking DDR3 board to be competitive, unless you find a CPU that runs 6.7+ CB15 on DDR2 board. But anyway, you won't find a CPU better than mine ?

Can you do me a favor? AFAIK MSI also has Performance Bias options in bios (that's how it is called on ASUS). If so, could you dump them for me? I can tell you the exact steps - would require several reboots and quick MSR dump for each option.

The stepdowns are here and I've hooked one of them to the Asrock board. Didn't help with freezes when in dual channel and high Vdimm. At ~3.37V real Vdimm I can only see the VGA OPROM message, but the board doesn't POST. Works with one dimm, but not with two and high Vdimm. It's not visible in the photo, but the MOSFET leg is desoldered and bent up, so I can solder the external wire. The green display is what I was using before to monitor Vdimm, but it's now disconnected, thus showing 0.00. So the next idea is to desolder all VDD pins of the DRAM PLL ICS93735 and connect to external 2.5V. If that doesn't work, I give up. PS: The board is probably dead, shorted 2 pins on the PLL

I don't think it's worth it, since you can tweak these things in OS. It's good for gaining some experience, but personally I don't have so much time to waste for this. I waste it on other projects You need to know assembly to successfully implement new functionality. On a side note, I will have MSI KT880 Delta to try as well, should be a little better than the Asrock and its bios is standard AMI, which I could open with AMIBCP and edit - no hidden options though. BTW, does anyone know how to display text correctly in Award Bios Editor? It is some problem with encoding or font and all the setup screen labels show as scrambled symbols. Perhaps it needs to be edited on Windows 98? I can edit it with modbin, but it's less functional.

Many of the registers match, but the primary timings don't. Not much to be gained, though. Everything is already set to fastest values. Slackening the timings and turning some things off might help with FSB, but I'm searching for performance boost.

I usually hot-flash in windows with WinFlash. Have a bios extractor clip, many spare Winbond bios chips and if anything goes wrong I just pop the bad chip out and put the working one. Have a willem programmer as well. Useful for full wipe and re-program, prepare several backup bios chips, etc. But that nanoflasher is much easier to use, since it is just USB.

Starting from KT400A through KT600 and KT880 should be similar. KT400A and KT600 are almost the same (same part numbers), KT880 is an evolution of KT600, but most registers are probably the same. To be honest, I expected a bit better performance from KT880. It might be just the Asrock board/bios, but it's not much faster than single-channel KT600. I believe performance could be improved by tweaking DRAM and V-Link registers, but need the documentation for this. Found this site, so it seems KT333 and KT400 PCR files exist, but downloads are broken and can't find them anywhere else: http://www.georgebreese.com/net/software/ It's probably incompatible anyway, because the part number is different and I need at least KT400A or newer. KT880 was so late in the game that it never got popular. People only bought boards based on it for cost-saving systems.

Unfortunately offsets don't match. Haven't found a tool which can read all the timings. AIDA reads more than others, but if I change a timing with wpcredit cpuz is the only one reflecting it. AIDA always reads 2-2-2-6, even if I set e.g. 2-2-2-9.

Thanks, that would definitely help. Slew rate and drive strengths are exposed in DFI Ultra-B bioses, but I haven't played much with them. In fact haven't played with the DFI board much, since it is much more unstable and picky than Abit. It is important to see the correlation between different bits, for example, one needs to set BYP (bypass) bit to 1 first in order to use the external values for N and P transistors. That's why nothing happened when I tried to set them by had on a running system. I might get my hands on a new tool with all the options included, not just the well-known timings. Depends what can be R/W runtime. Edit: That's the maximum DRAM frequency I can do with BH5@~3.08V in Dual Channel 1T. Higher than 3.1V and the board freaks out. Most of the screens on the desktop got corrupt after some subsequent crashes. Should have saved it on usb earlier, but it is still visible. Interestingly enough, I can't do 250 stable anymore and could not go to 260 for "a screen", no matter the volts (raised or "stock"). Perhaps I did it on a later bios, don't remember. Spent some time figuring out the timings in wpcredit, however I'm able to control the main timings only. Others seem to be read-only b0/d0/f3x56 [1,0] tRP (11-5, 10-4, 01-3, 00-2) [3,2] tRCD (11-5, 10-4, 01-3, 00-2) [5,4] tCL (11-3, 10-2.5, 01-2, 00-1.5 or reserved) - readonly [7,6] tRAS (11-9, 10-8, 01-7, 00-6) Lowest is 10 -> 00010000 -> 2-2-2-6 If someone finds PCR file for KT400/A or datasheet for KT400/600/880 please let me know.

Some update (22 Oct): Got 4 mods now - Vnb/agp, Vsb, Vdimm, Vcore. Mods for NB and VDIMM are "reverse" and require 2 smd resistors to be removed - higher the resistance, higher output voltage. Quite hard to desolder and solder, since it is between slots - you really need to have a steady hand. R-C divider circuits are used for both. Vnb doesn't seem to help much and above 2V I get artifacts, so probably need to test another card. Maybe some ATI would work better. Vsb not sure if helps, but since the controllers are there it might allow me to go higher without dropping IDE. Vdimm, still the same situation. 3.1V seems to be the maximum I can get away with. Anything higher and I think the PLL freaks out, since VDD is derived from Vdimm. So maybe the better plan is to leave Vdimm as it is and disconnect all VDD pins of the PLL, then feed it with constant external 2.5V. It will be hard :(. Perhaps it would be easier to move the whole PLL on external PCB and reconnect just needed pins, but supply VDD pins with external ref voltage from a DC-DC buck stepdown. Board can boot 260+ and work in bios, but is unstable in windows. I can now use TCCD-based memory, but they can't go very high with CR=1T. 2T is needed which kills performance, together with the looser timings required (2.5-3-3-6-12). At 2.4GHz I'm 3 seconds slower than @TerraRaptor and higher frequency would allow me to beat his score (with worse efficiency). I've spend one night trying to install new windows from usb or use acronis to restore a tweaked image without any luck. Finally installed acronis as a windows program and restored from there. The board is PITA to work with.

30-40 celsius load, below 0 idle, I think. Didn't want to push it with higher vcore, otherwise it could probably bench a little higher. Could bench CB15 at 4.4GHz with stock cooler (Wraith Prism) below 1.3V: https://hwbot.org/submission/4255481_i.nfrar.ed_cinebench___r15_ryzen_9_3900x_3479_cb

Regarding CPU detection on socket A systems - it's not black or white. Different bioses might show different CPU name. The CPU string is selected based on the CPUID, VID and FID but not all bioses are updated with support of newer socket A CPUs, especially later models like T-bred Semprons and higher rated Bartons. There are many modded bioses with fixes for CPU strings and everyone implemented them differently, depending on the space available in bios for these strings. On top of that, CPUs can be unlocked in various ways, which sometimes doesn't match any existing factory model. You can even use tools to change that string runtime (Specification). One example: stock bioses have a list of known models and the last string is "Unknown", so if it doesn't match all known models, then the bios will return "Unknown Athlon XP" or something like that, which will then be displayed in cpuz "Specification" field. Most of the modded bioses replace this string with "XP-M", "Mobile", etc. However, the CPUID is read from CPUID MSR and should always be correct. Same for feature set. These are read-only. Even if we assume somehow cpuz is to blame for displaying "Spitfire" core with CPUID of "Morgan", then these scores still don't belong to Duron "Spitfire" category. I've yet to see such an error in detection, but who knows... Given his previous records though, I don't think it's coincidence. Moreover, he used the same cpuz 1.41 for other submissions and it shows the CPU correctly - both Morgan and Spitfire. Btw, Duron 600 is the only Spitfire I'm missing in my collection.

Actually it should be intentional editing, since cpuz would read Morgan if it was the wrong screen. I doubt it is some bios/cpuz quirk. He most probably just copied and pasted Spitfire from other screen. Which should be enough to ban him forever. Notice the SSE instruction set support as well. PS: Reported.

@TerraRaptorThis is XP with some Vista-like theme, but other points are valid. Look at his Model and Ext Model that differ from what is on other guys' screens. CPUID doesn't match Spitfire. It's Duron "Morgan". He either benched with Morgan intentionally or submitted in the wrong category. Either way it should be removed/moved to proper category. Probably this CPU: https://hwbot.org/submission/645138_bwanasoft_superpi___1m_duron_1300_(morgan)_1min_13sec_700ms BTW, unfortunately it's very easy to "fake" a socket A CPU model in the same model family. You can potentially bench with just one golden unlocked cpu and submit in several categories. That's why I usually try to provide at least a photo of the CPU, which ofcourse doesn't mean I've benched with it, but I count on honesty in these oldschool categories. I think there are only good guys benching them nowadays. But nothing stops you to cover all Barton/Thoroughbred categories with just 2 CPUs .

Yes, VTT tracks Vdimm and maintains 1/2 ratio correctly. As soon as I put a dimm in the closest slot to the cpu it freaks out. Without voltmod it is fine. Could do 250 2.5-3-3-6 1T with 2x256MB TCCD Maybe could try to replace the capacitors. PS: I think the ICs chip responsible for DRAM PLL might not like it. I don't see how I can separate them. VDD (several pins) for the ICS93733CF is from Vdimm. Nominal is 2.5V, so Asrock gets away with moderate increase to 2.6-2.7, but I'm hitting the tolerance. Recommended is 2.3V to 2.7V max. I'm way beyond that, closer to the absolute maximum.

Yes, that's correct. AGP frequency is not a problem, the card can work@100+ MHz. Don't know how much more the PCI could handle, but I guess it might not be a problem. The board is a PITA to mod. It is using separate controller for the CPU, can mod it easily (bios limited to +10% which results in 1.775V for the test Sempron 2200+). I can go up to 2V via software in windows though. And there's a FB pin, so this should be an easy mod. The rest (Vagp, Vnb, Vdimm and Vsb) are controlled by a single LM324. Vagp and Vnb are tied together, so incresing Vagp will increase chipset voltage as well. Vsb is 2.7V. Vdimm is problematic. I guess something else is fed with Vdimm and when the DRAM votlage is increased to 3+ the board gets unstable and hardly even POSTs when 2 DIMMs installed. If it boots, then it freezes randomly. The only solution I see is to use separate power for DRAM, e.g. lift the mosfet leg and get power from a DC-DC buck converter, but I will have to order some. "High" setting for Vdimm from bios translates to around 2.64V, but when 2 modules installed the voltage drops to 2.44V. I have tried direct 3.3V rail mod and it behaves the same. Also tried OCZ Booster (which is the same, but source is 5V rail) and still freezing, although I can't really use it for benching on this board, because the slots are very close to each other and it interferes with one of the slots used for dual-channel configuration. On top of that, the mod area is very crammed with all the slots and elements around it. The LM324 chip is between first AGP slot and the expansion card slot. As you might guess 2.44V is not really enough for anything with BH-5 and tight timings. Running different divider than 1:1 means performance loss. So..waiting for the buck converter. Can't figure bios modding out. It's AMI, but all amibcp tools I tried say "bios checksum incorrect" and refuse to load. There's only one good bios for this board (Performance bios M1.00), which has the timings unlocked. It's a "beta" from Asrock.

I have bought the frankenstein board Asrock K7Upgrade-880. The bios is not very tweakable, with most voltages missing (or having something like +10%, normal, high). Will need voltmods for everything. The good things: - Unlocked Sempron 2200+ boots with its native multiplier x9 (and not the highest detected), which will help reaching high frequencies, that are not possible on nForce2. - I can control the multiplier within Windows, which was the original goal - Board works at 250MHz FSB no mods, could go to 260, but it crashed. Memory divider set to 166 right now. - There's S2K drive strength in bios - There are FID jumpers, which might be useful in some cases Will have to check the bios if I can mod it somehow, unlock hidden options, etc. This was the only KT880 board I could find. Came with a Sempron 2600+ and a cooler.

Shouldn't be a problem with strong hardware. It's not a problem for the videocard. If the integrated controllers don't drop at high FSB then external PLL might not be needed. My KT600 stops at ~240MHz with mods, but haven't messed with bios yet.

As far as I can see the max FSB is comparable to nForce2 and that's without the lock. Hooking up external PLL might allow it to go higher. I've seen reviews that it is a little faster clock for clock and might allow higher stable DDR frequency due to a better mem controller. Also has higher bandwidth. As for the diagrams, if you're interested you can make the physical FSB-detection switches with cables soldered at the back of the socket (L12-mod). I have them on the old AN7, but it doesn't help when you have a modded bios. Doesn't hurt to have it, though. I've also made an entire multiplier switch board in an attempt to change the multiplier of a super-locked CPUs, but it didn't work. That's why KT880 would be better. On Nforce2 you're limited by the FSB and with 9x multiplier (in the case with 2200+ Sempron) you can't go much higher, while on VIA chipset 2600+ is possible. If you find a golden chip you can even go to 3GHz with SS/LN2, I guess. I think it would be rather easy to beat @TerraRaptor's wprime scores even on KT600, but since it is single-channel only I couldn't beat his PI 1M score even at 2.6GHz (came close, but still slower). http://fab51.com/cpu/sempron/s10-e.html

In my original bioses I've followed that recommendation and changed all these values. On a side note, the memory timings start from b0/d0/f1 offset 90 (IIRC, at work right now). I've tried to play with some values in the near offsets, without much luck. Will try your values in the evening, but will change the CPU first, since I'm using one of my best XP-M in the moment. Tried to find a similar romsip table in VIA KT600 bioses, again without a positive result. Also checked NF3 and NF4 bioses and although there are some tables, it's not the same as in nforce2 bioses. PS: Btw, I think a better option might be a KT880 board, eventually modding it with external PLL, since it has no PCI/AGP lock. The mem controller is better than nF2 and you can also change multiplier in windows, which is useful for super-locked CPUs. Edit: S2K Bus Disconnect is b0/d0/f0/6F, data set to 1F (or 10). Perhaps we can dig into that direction and the S2K drive strengths are also there and we can control the values runtime. The S2K should be common/similar for all K7 chipsets. But be aware that all documents say it can be dangerous for the CPU and damage it permanently. I remember using S2KCtl application back in the days, to control the bus disconnect feature. CPUId can also control HALT states and divisors. Have to check if something changes accordingly when you modify the romsip table registers from @TerraRaptor's testing. Edit 2: ROMSIP from Nforce4 Ultra-D 702-1 bios (BH-5), located in decomp_blk.bin 000061c0: 65 D0 16 2B 00 00 00 00 00 00 00 00 10 4C FF FF eP.+.........L.. 000061d0: 08 00 03 00 00 00 00 0F 1C 70 E0 81 FF FF FF FF .........p`..... 000061e0: FF FF FF FF FF FF FF FF 00 00 00 00 00 00 00 00 ................ 000061f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006200: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006210: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006220: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006230: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006240: 21 41 24 00 00 E4 16 1D 21 41 25 00 00 E4 16 1D !A$..d..!A%..d.. 00006250: 21 41 25 00 00 E4 16 1D 21 41 26 00 00 E4 16 1D !A%..d..!A&..d.. 00006260: 21 41 23 00 00 E4 27 25 21 41 24 00 00 E4 27 25 !A#..d'%!A$..d'% 00006270: 21 41 25 00 00 E4 26 1D 21 41 26 00 00 E4 26 1D !A%..d&.!A&..d&. 00006280: 21 41 20 00 00 E4 16 1D 21 41 21 00 00 E4 16 1D !A...d..!A!..d.. 00006290: 21 41 21 00 00 E4 16 1D 21 41 22 00 00 E4 16 1D !A!..d..!A"..d.. 000062a0: 21 41 22 00 00 E4 16 1D 21 41 23 00 00 E4 16 1D !A"..d..!A#..d.. 000062b0: 21 41 23 00 00 E4 16 1D 21 41 24 00 00 E4 16 1D !A#..d..!A$..d.. 000063a0: 65 D0 16 2B 00 00 00 00 00 00 00 00 10 4E FF FF eP.+.........N.. 000063b0: 08 00 03 00 00 00 00 0F 1C 70 E0 81 FF FF FF FF .........p`..... 000063c0: FF FF FF FF FF FF FF FF 00 00 00 00 00 00 00 00 ................ 000063d0: 03 80 80 80 4B 04 00 00 03 80 80 80 4B 04 00 00 ....K.......K... 000063e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 000063f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006400: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006410: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00006420: 21 41 24 00 00 E4 16 1D 21 41 25 00 00 E4 16 1D !A$..d..!A%..d.. 00006430: 21 41 25 00 00 E4 16 1D 21 41 26 00 00 E4 16 1D !A%..d..!A&..d.. 00006440: 21 41 23 00 00 E4 27 25 21 41 24 00 00 E4 27 25 !A#..d'%!A$..d'% 00006450: 21 41 25 00 00 E4 26 1D 21 41 26 00 00 E4 26 1D !A%..d&.!A&..d&. 00006460: 21 41 20 00 00 E4 16 1D 21 41 21 00 00 E4 16 1D !A...d..!A!..d.. 00006470: 21 41 21 00 00 E4 16 1D 21 41 22 00 00 E4 16 1D !A!..d..!A"..d.. 00006480: 21 41 22 00 00 E4 16 1D 21 41 23 00 00 E4 16 1D !A"..d..!A#..d.. 00006490: 21 41 23 00 00 E4 16 1D 21 41 24 00 00 E4 16 1D !A#..d..!A$..d.. Very similar to "CPU Interface OFF" [E4] table from TaiPan 0.3, but the last value is adjusted "higher". There's also a difference in the first part of the table, but don't know if it can be used directly in NF2. Same "multiplier" table can be found in nForce3-250GB UT bios. Is decomp_blk.rom just the LHA-decompressor? If so, I guess it's no point of interest?

Doh, thought most of the nForce2 boards use the SilliconImage. Well, my idea still stands. By updating the SilliconImage rom it became much faster and definitely faster that tne IDE->SATA adaptors I use sometimes as well. But if you couldn't manage it, then buy some of these cheap adaptors (very fragile, so get some spares). AFAIK many of these controllers have problems with large drives, which sometimes is solved by updating the firmware in the bios.

Update the sillicon image Sil3112 to 4.4.02 (last available version). I've used it successfully many times on AN7 and always use USB flash drive as an install media. Some boards can't boot from USB though and need some sort of a boot manager mod (I've recently replaced LAN option rom with plop manager on NF7-S). You can update the bios by simply replacing existing RAID rom with the updated version using cbrom. PS: Alternatively, you can give me your bios files and I will quickly mod them. That for the Soltek board, for the promise controller maybe search for a newer version as well. 4402.bin