Looks like some TNTs (ELSA SGR as an example) use 5V rail to convert it to 3.3v for memory and/or core. Probably to address early AT AGP motherboards blowing up from excess load on 3.3v line, but by sacrificing efficiency.

havli wrote on 2026-03-29, 11:00:

Here we go: […]
Show full quote

Lostdotfish wrote on 2026-03-22, 14:33:

I'd be really interested in seeing some numbers for early Nvidia cards. Riva 128, TNT, TNT 2, 256 etc

Here we go:

Riva 128.......................................100/100 MHz........4 MB SGR.........Idle = 3,4 W.........3DMark99 = 4,9 W............ Unreal = 4,2 W........(most power from 3.3V rail)
Riva 128 ZX..................................100/100 MHz........8 MB SGR.........Idle = 3,8 W.........3DMark99 = 5,3 W ............Unreal = 5,0 W........(most power from 3.3V rail)
Riva TNT SDR.................................90/110 MHz.......16 MB SDR.........Idle = 7,1 W.........3DMark99 = 9,8 W............ Unreal = 9,8 W........(most power from 3.3V rail)
Riva TNT SGR.................................90/110 MHz.......16 MB SGR........Idle = 11,3 W........3DMark99 = 15,2 W.......... Unreal = 15,4 W........(most power from 5V rail)
Riva TNT SGR OC..........................120/135 MHz.......16 MB SGR........Idle = 13,7 W........3DMark99 = 18,6 W......... Unreal = 18,8 W........(most power from 5V rail)
Riva TNT2 Vanta...........................100/125 MHz......16 MB SDR..........Idle = 4,8 W.........3DMark99 = 6,7 W.......... Unreal = 6,6 W........(most power from 3.3V rail)
Riva TNT2 m64.............................125/125 MHz......32 MB SDR..........Idle = 7,6 W........3DMark99 = 10,0 W.......... Unreal = crash........(most power from 3.3V rail)
Riva TNT2......................................125/150 MHz.....32 MB SDR.........Idle = 7,7 W.........3DMark99 = 11,1 W........... Unreal = 11,2 W........(most power from 3.3V rail)
Riva TNT2 Ultra.............................150/183 MHz......32 MB SDR........Idle = 10,0 W........3DMark99 = 13,8 W.......... Unreal = 13,6 W........(most power from 3.3V rail)
GeForce 256 SDR...........................120/166 MHz......32 MB SDR........Idle = 15,8 W........3DMark99 = 25,9 W......... Unreal = 24,9 W........(most power from 3.3V rail)
Quadro SDR (NV10, 64 MB)...........135/166 MHz......64 MB SDR........Idle = 17,2 W........3DMark99 = 26,6 W........ Unreal = 28,4 W........(most power from 3.3V rail)
GeForce2 MX..................................175/166 MHz......32 MB SDR.........Idle = 8,5 W........3DMark99 = 12,6 W.......... Unreal = 12,4 W........(most power from 3.3V rail)
GeForce2 GTS.................................200/333 MHz......32 MB DDR.......Idle = 14,7 W........3DMark99 = 19,1 W.......... Unreal = 19,3 W........(most power from 3.3V & 5V rail)
GeForce2 Ti....................................250/400 MHz.....64 MB DDR.......Idle = 15,5 W........3DMark99 = 21,5 W.......... Unreal = 21,1 W........(most power from 3.3V & 5V rail)

I tried a small sanity check against those numbers:

DangerManfred measured a Geforce 2 GTS: https://tehsiggi.github.io/agp-power-monitor/ … geforce_2gts_32

That looks very comparable to your measurement. So I think your values work well. I guess you're running the software with the default calibration mulitiplicator right? Seems it is sufficient!

The Serpent Rider wrote on 2026-03-29, 14:00:

Looks like some TNTs (ELSA SGR as an example) use 5V rail to convert it to 3.3v for memory and/or core. Probably to address early AT AGP motherboards blowing up from excess load on 3.3v line, but by sacrificing efficiency.

And since the regulation is completely linear it seems, that'll be quite the efficiency loss (only 66% conversion efficiency, rest 34% is heat).
Not sure what they all power from that.. but comparing the consumption of the SGR vs. SDR in 3Dmark99 -> the SDR consume around 64% of the SGR. Coincidence? 😉

The Serpent Rider wrote on 2026-03-29, 14:00:

Looks like some TNTs (ELSA SGR as an example) use 5V rail to convert it to 3.3v for memory and/or core. Probably to address early AT AGP motherboards blowing up from excess load on 3.3v line, but by sacrificing efficiency.

Yes, I used this exact SGR card for measurement. And the SDR version is also ELSA - this one https://commons.wikimedia.org/wiki/File:ELSA_ … Riva_TNT%29.jpg It seems both memory and core must be powered directly by 3.3V rail from the slot as there are no voltage regulators present.

Also from the datasheet is seems SGRAM consumes more power. Most likely both factor contribute to higher consumption.
SGRAM https://tvsat.com.pl/PDF/K/km4132g512_sam.pdf
SDRAM https://www.digchip.com/datasheets/parts/data … 20CT-G8-pdf.php

tehsiggi wrote on 2026-03-29, 15:51:

I tried a small sanity check against those numbers:

DangerManfred measured a Geforce 2 GTS: https://tehsiggi.github.io/agp-power-monitor/ … geforce_2gts_32

That looks very comparable to your measurement. So I think your values work well. I guess you're running the software with the default calibration mulitiplicator right? Seems it is sufficient!

Yeah, the default correction factor of 1.0275. There is error of 0.25W on the 12V rail, but that is not significant. Interesting fact - these old GPU don't use 12V from AGP at all. Notable exceptions are Voodoo3 3500TV, Rage 128 Pro AIW and Quadro 2 Pro (all of them take about 4W from 12V rail).

Yeah, you really see at which time AGP was born. 3.3V and 5V were pretty dominant, 12V was somewhat on an "auxiliary" voltage at best, hence only having a singular pin in the whole slot.

The quarter watt on the 12V is basically noise. I never went ahead and put a "zero out" function in place, where you turn the system on without a GPU, zero out and then put the card in. If I ever find time..