[OFFICIAL ISSUE] 8800GT (new PCB) report delle temperature

[Computerboy]

Allora la soluzione è stata trovata su Guru3d: il link alla discussione (in inglese, è questo [Solo gli utenti registrati possono vedere questo collegamento. ]).

It looks like we're close to getting direct readings from G9x thermal diodes. I need some user with G92 based card and normal BIOS (which has thermal monitoring working) to perform the following:

1) Record current core temperature
2) Run "RivaTuner.exe /rr20008 /rr20400", which will give you a message box containing output of two GPU registers.

The formula used by ASUS SmartDoctor for such cards is

T = (-13115 + raw_diode_data) / 18.7 + 1, where raw_diode_data is lower 14 bits of Reg 20008.
So it gives (-13115 + 0x3665) / 18.7 + 1 = 44C

However, currently we're absolutely not sure if it gives the real and correct temperature. G80 and older thermal diodes have non-fixed formulas, using diode inaccuracy specific to each diode sample, and I'm afraid that ASUS SmartDoctor programmers just skipped this code to simplify their job. And the dumps/temperatures posted by users really shows that the equation used by ASUS is a bit different to the ForceWare's one.
To find out the correct equation I'll need to install any G92 card with working thermal monitoring in my rig, but currently I have no G92 based samples at all. So I'm trying to find one and maybe I'll be able to get it and implement direct G92 diode support in NVThermalDiode plugin in 2.08[ndr: Rivatuner 2.08, nuova versione].

Si è scoperto infatti che il problema di lettura delle temp risiede in ambiente bios: il parametro di lettura arriva dal diodo interno del core ma manca un algoritmo preparato dagli ingegneri diel G92 che sia in grado di approssimare con certezza e convertire il gradi Celsius questo valore in esadecimale.

L'unico programma in grado di fare un calcolo approssimativo e con una stima di errore di +10/-10° è stato ASUS SmartDoctor versione 5.08, tuttavia l'algoritmo utilizzato dagli ingegneri asus è molto artificioso: in base alla lettura che potete trovare anche voi utilizzando la riga di comando

Run "RivaTuner.exe /rr20008 /rr20400", which will give you a message box containing output of two GPU registers

è possibile infatti notare la lettura del diodo in valore esadecimale. I tecnici asus, sulla base di questo valore e SULLE MISURAZIONI avvenute esternamente al core via sonda termica hanno calcolato un valore di offset (spostamento) e una costante: la formula creata, suggerita da Unwinder, non fa altro che "convertire" abbastanza empiricamente un valore esadecimale facendo in modo di rendere verosimile il risultato, cioè di farlo coincidere con la temperatura misurata. Tale artificio permette di monitorare abbastanza verosimilmente la temperatura almeno da Rivatuner.

Looking at the way ASUS implement thier temperature calibration formula in SmartDoctor code (i.e. T = (offset + raw_diode_data) / divider), I may say almost with 100% warranty that the formula doesn't come from NVIDIA and ASUS have approximated it using raw diode readings and comparing it with external sensor readings. Most likely they recorded raw diode data for 0C monitored by their external sensor and this way received offset then recorded raw diode data for let's say 50C, applied previously calculated offset and this way calculated the divider. This fully explains why they use offset / divider approach instead of offset / gain formula, traditionally used by NVIDIA driver on the rest G92 with properly functioning thermal monitoring.
Unfortunately, as I've said above, the diodes are NOT ideal and calibrated differently, so creating approximated formula using one diode can give good results on the same card but lie on other.

Il seguente problema non è assolutamente presente nelle "vecchie" versioni della 8800Gt, poichè nella rev.2 è stato accorciato (e anche di un bel po' il PCB) con le seguenti cause:

I've invested some more time into analyzing the G92 thermal monitoring codepath in the driver and G92 thermal table in G92 BIOS images and I guess we can close the issue, because the situation is more than clear for me now. The facts:

1) Due to some unknown reason thermal table of all G92 BIOS images (i.e. special data area in BIOS, which calibrates thermal logic of the card) contains "Auto detect external thermal sensor" flag, which tells the ForceWare that it should try to use supported external thermal sensor instead of internal one. Reference designG92 boards use ADT7473 external sensor / PWM controller to automatically adjust fan speed, and this sensor's readings are used by the ForceWare to report GPU temperature on any G92.

2) Some vendors including ASUS are altering cooling systems then trying to make PCB cheaper by removing unwanted ADT7473 controller from reference PCB and unfortunately this way remove the only sensor supported by ForceWare on these cards too.

3) NVIDIA G92 BIOS images completely lack G92 internal diode calibration info in the thermal table, so it won't be that easy to edit the BIOS to force the driver to use the internal sensor instead of external one. Enabling diode is not the only thing to be done in BIOS - diode calibration must be also provided to driver to allow it to convert raw data to temperatures.
Due to the same reason it won't be that easy to provide G92 diode support in NVThermalDiode plugin. I've no ideas why NVIDIA decided to avoid internal sensor usage on G92, probably it is broken at hardware level. If external sensor is used on G92 erroneously, then you can hope for fix in future VGA BIOS updates and ForceWare drivers.
5) ASUS SmartDoctor's diode calibration formula had been approximated for sure by ASUS engineers and unfortunately it is absolutely not the best way, which can easily give +/-10C inaccuracy comparing to real diode temperatures.

Alcuni venditori tra cui ASUS, PALIT, TWINTECH, GIGABYTE per limitare i costi di produzioni e per rendere quindi la board più appetibile sul mercato, hanno preferito accorciare il PCB della scheda, eliminando quindi l'ADT7473 controller, l'unico controller esterno alla gpu che sulle schede reference funge da termal monitor. Così facendo hanno tolto l'unico sensore nativamente supportato dai Forceware, per cui è svelato l'arcano delle temperature 0°.

RT does nothing to calculate / correct temperature when default data provider is selected. In this case monitoring is performed by means of NVIDIA driver, driver selects desired sensor (according to BIOS), calibrates reading (according to BIOS thermal table again) and reports resulting temperature.

Rivatuner non può quindi calcolare la lettura, poichè già in ambiente bios non arriva da quel controller da cui i Forceware si aspettano una lettura.

{<<Probably many of you noticed that initially I requested G92 report with _two_ registers (20008 and 20400), but it fact only one of them (20008) contained trustworthy temperature related data on 8800GT series>>}.
According to my analysis of driver's internal thermal monitoring implementation internal sensor of G92 (or more precisely internal sensor of all NVIDIA GPUs newer than G84 A2) can report both raw data (as the previous thermal diodes included in NV4x, G7x and G80) which can be calibrated by software algorithm in the driver and ready to use temperature calibrated with special hardware logic. However, the second register (assumingly containing calibrated temperature) was set to 0 on all G92 chips so I thought that I simply misunderstood the driver's code. Now I see that I've not misunderstood it and new chips really contain hardware calibration logic and unfortunately it is not working on G92 at all. I vote for hardware bug in G92 internal thermal diode.

In aiuto è quindi arrivato il fix ideato e scritto da Igor Levicki:

Ok guys...

A bit of a good news for everyone who has NVIDIA G92 based video card -- mainly 8800 GT 512 MB models which do not have the reference cooling solution and lack ADT7473 chip.
Until Unwinder releases RivaTuner 2.07 and an updated NVThermalDiode plugin, you can download, install and use G92 plugin I wrote from my website downloads section -- [Solo gli utenti registrati possono vedere questo collegamento. ].
Be sure to read the disclaimer though.

Ricapitolando: finchè Unwired non rilascerà Rivatuner 2.07 è necessario supportare la lettura via software delle temperature attraverso un driver .dll che ricolleghi la lettura di RT non al chip ADT7473 (mancante su queste boards!) ma al diodo interno già presente dalla serie G84 (vedi sopra).
Potete scaricare la seguente G92.dll a questo indirizzo:[Solo gli utenti registrati possono vedere questo collegamento. ]
Anche a questo link:[attachment=0:gw3edgqw]G92.zip[/attachment:gw3edgqw]