Very interesting experiments! I had always assumed 3dfx would bin their avenger chips for the various models, but wasn't sure because of the voltage and cooling system differences. It definitely makes sense though and great to see this confirmation. I also wonder how much week-to-week manufacturing differences would play a role here too. Overclocking beyond 166MHz is a bit of a different story though and in my experience its usually the memory holding the cards back. Its unfortunate that you can't clock the core/memory separately like you can with the Voodoo Banshee. I have a Voodoo 3 3500 that I hope to do some experiments on one of these days.
A Voodoo 3 3500 is still missing in my collection. Some day I'll get one! Regarding the binning, I believe that they definitely binned their chips. It would not be good business if you buy a Voodoo 3 2000 and overclock it to 3500 level with their own overclocking utility. Especially that 3dfx turned the greed on when they closed the door to their previous board partners.
If you spread a bit of thermal paste on the heatsink, it should be visible under the thermal camera because it's not reflective, and should also not affect the heatsink efficiency. Lower temperatures also allow you to run lower voltages. If you can reduce the temperature from 70C+ to 40C with an active cooler, you might get a few milivolts of improvement. Reducing the voltage ripple by increasing capacitance might also allow you to run lower voltages by another few milivolts (an oscilloscope is needed for this). Keep it up! I really enjoy seeing people trying to get the most out of old hardware :)
Dude I still have a voodoo 3000, and the CDs. Man I remember how revolutionary that tech was back then. I also have my old gf3. Total nostalgia seeing 3dmark 2k 2k1, I still have those programs saved in my archive hdd. Lol.
Chipbinning was roughly done by the location of the chip on the wafer. In the middle, the structures were best (the location of every layer above the last one), the more you go to the outer locations, the worse it gets. That's because of the precision while exposing the silicone and the repeatability of this process. Outer chips had slightly more shifts between each layer, so the component values where more off the theoretical values.
@@bitsundbolts Yes, nowadays the methods are not as rough as back then. Now they use f.e. light scattering (shining light at a specific angle and checking the reflection with a camera) and various far more superior methods. Take a look at "wafer map" / "substrate mapping", "optical wafer defect inspection" or "wafer probing". Sometimes you can even connect to the chips directly via pogo pins (probes) and check their performance, but not at cpu-level, simply to much pins... But the rule of thumb still applies that wafer more in the middle are better than the outer ones. BTW I think you noticed without further mentioning, but i really like your channel and content! Keep on your work! I do similar stuff myself, but, despite of a handful of short videos without editing, never found time to create and polish content like you do. Thumbs up!
I guess this is evidence there was genuine binning of Voodoo3 chips back in those days, there was always the suspicion that the 2000/3000 difference in silicon was entirely artificial but it seems not in every case. Many 2000's seemed to have no issue being run at 166MHz though, it was almost accepted wisdom that every 2000 could be run at that speed without issue when that probably wasn't strictly true.
I am testing my second Voodoo 3 2000 now. I think it was probably rare to have one of those cards that can't run at 166 MHz. If there was binning in place, those probably would have been kept for Voodoo 3 1000s. The only problem is the cooling. I'll try to find out which of my 2000 is the better one - and I'll turn it into something special 😉
If I'm not mistaken the AGP 3000 card is running at a 66MHz host bus while the PCI 2000 card only has a 33MHz connection to the host. I could see this being why the voltage needed to keep the system from locking up is much higher.
I have every flavour of Voodoo3 except the motherboard integrated one, one of which is in a permanent retro build & I run them all with fans either on or next to the heatsink because they all run hot, especially inside period correct cases & they always did.
Voodoo3 PCI supports both 3.3/5V PCI signaling (which is why there are two "keys" on the PCI connector). The motherboard you are using has 5V PCI (which is pretty standard). The AGP Voodoo3 connector is keyed for 3.3v. A Voodoo3 chip in PCI 5V mode probably uses a different buffering/re-driving mechanism than in 3.3v/AGP mode. What you ran into was probably the difference in sensitivity of those two bus voltage modes. That's my guess anyway.
My STB V2000 works overlocked to V3 3500 183mhz with a small radiator + cooler (no other changes in bios/Windows except mhz). Only for tests ofc. Standard V2000 works for me ideally :) Great times, love Voodoo!
Great video, for more accurate core temperature readings I would suggest measuring the PCB behind the chip - due to the package type of the Voodoo3 chips most of the heat goes into the PCB and not the heatsink. On well used cards that spot can be discoloured from years of it running pretty hot!
I will put a fan on those cards. For now, I just want to know how they were sold to the end user. I absolutely agree that they are very hot running cards and a fan should be put on them.
It's also possible to put a little bit of black electrical tape on the heatsink in order to be picked up by the thermal camera. Great video, can't wait to see what shenanigans you get up to next!
i am pretty sure you could go more efficient on the V3-2000 (higher clocks or lower voltage) if you would replace its cooler with something more beefy. lowering the temperature will increase efficiency.
Yes, that could work, but it noticed that I could measure the temperature at the back of the card as well. Maybe there are a few degrees difference, but it should be good enough.
Had they done this testing, they might have been able to offer *mobile* voodoo chips. That would have been a much better option than being full stack manufacturer as they tried to do.
I know there are utilities that can read the internal temperature measurement of the die, and tools that ran in the background to log the temperature measurements to a CSV file. However I don't remember if these old cards had the thermocouple and circuitry to read it built into the die. Perhaps it's worth investigating?
I have never found any information about thermal sensors in any Voodoo graphic chips. I think those came later, but would have been great to have those available back then!
BEWARE: artefacts are a poor way of verifying that the GPU is functioning correctly. Vector/polygon calculation errors will begin way before any typical artefacts become visible. OCCTPT is a benchmark that tests for such errors. Also "Artefact Locator" & Video Memory Stress Test" (both by Misha Cherkes) can be useful.
The memory is powered by a 3.3 volt supply, which actually comes from the voltage regulator with the heatsink. I don't think there is an issue since the core is already receiving lower power.
Maybe if I ever get my 1000 fixed up and if I were to slap a heatsink on it (1000s had no heatsink), I wonder if I could get it to 2000 or maybe 3000 speeds
@@bitsundbolts it’s a full feature core with 16mb SGRAM, I think some came with 8mb. There were also the velocity cards with only a single TMU and I knew they only had 8mb
To be honest I did not think this channel will grow up so nicely. Probably with 3dfx chip force !
Simply one of the best 3dfx voodoo channels online. Subscribed for more stuff like this
Excellent video. A fan will not only reduce the temperature but will also increase the maximum stable frequency by around 5-7 mhz.
After I am done testing my second Voodoo 3 2000, I'll get into changing cooling solutions.
Very interesting experiments! I had always assumed 3dfx would bin their avenger chips for the various models, but wasn't sure because of the voltage and cooling system differences. It definitely makes sense though and great to see this confirmation. I also wonder how much week-to-week manufacturing differences would play a role here too. Overclocking beyond 166MHz is a bit of a different story though and in my experience its usually the memory holding the cards back. Its unfortunate that you can't clock the core/memory separately like you can with the Voodoo Banshee. I have a Voodoo 3 3500 that I hope to do some experiments on one of these days.
A Voodoo 3 3500 is still missing in my collection. Some day I'll get one!
Regarding the binning, I believe that they definitely binned their chips. It would not be good business if you buy a Voodoo 3 2000 and overclock it to 3500 level with their own overclocking utility. Especially that 3dfx turned the greed on when they closed the door to their previous board partners.
If you spread a bit of thermal paste on the heatsink, it should be visible under the thermal camera because it's not reflective, and should also not affect the heatsink efficiency. Lower temperatures also allow you to run lower voltages. If you can reduce the temperature from 70C+ to 40C with an active cooler, you might get a few milivolts of improvement. Reducing the voltage ripple by increasing capacitance might also allow you to run lower voltages by another few milivolts (an oscilloscope is needed for this).
Keep it up! I really enjoy seeing people trying to get the most out of old hardware :)
3dfx are the best videos of your channel. I watch them all
Dude I still have a voodoo 3000, and the CDs. Man I remember how revolutionary that tech was back then. I also have my old gf3. Total nostalgia seeing 3dmark 2k 2k1, I still have those programs saved in my archive hdd. Lol.
Chipbinning was roughly done by the location of the chip on the wafer. In the middle, the structures were best (the location of every layer above the last one), the more you go to the outer locations, the worse it gets. That's because of the precision while exposing the silicone and the repeatability of this process. Outer chips had slightly more shifts between each layer, so the component values where more off the theoretical values.
Very interesting! Thanks for sharing! That obviously makes the binning process a lot easier! It won't be 100% accurate though. So, Silicon Lottery :)
@@bitsundbolts Yes, nowadays the methods are not as rough as back then. Now they use f.e. light scattering (shining light at a specific angle and checking the reflection with a camera) and various far more superior methods. Take a look at "wafer map" / "substrate mapping", "optical wafer defect inspection" or "wafer probing". Sometimes you can even connect to the chips directly via pogo pins (probes) and check their performance, but not at cpu-level, simply to much pins... But the rule of thumb still applies that wafer more in the middle are better than the outer ones.
BTW I think you noticed without further mentioning, but i really like your channel and content! Keep on your work!
I do similar stuff myself, but, despite of a handful of short videos without editing, never found time to create and polish content like you do. Thumbs up!
I guess this is evidence there was genuine binning of Voodoo3 chips back in those days, there was always the suspicion that the 2000/3000 difference in silicon was entirely artificial but it seems not in every case. Many 2000's seemed to have no issue being run at 166MHz though, it was almost accepted wisdom that every 2000 could be run at that speed without issue when that probably wasn't strictly true.
I am testing my second Voodoo 3 2000 now. I think it was probably rare to have one of those cards that can't run at 166 MHz. If there was binning in place, those probably would have been kept for Voodoo 3 1000s. The only problem is the cooling.
I'll try to find out which of my 2000 is the better one - and I'll turn it into something special 😉
@@bitsundbolts BuB the best 3dfx youtube channel
If I'm not mistaken the AGP 3000 card is running at a 66MHz host bus while the PCI 2000 card only has a 33MHz connection to the host. I could see this being why the voltage needed to keep the system from locking up is much higher.
I came for Voodoo3 video to learn something new, I am not disappointed !
Voodoo 3 2000 was my very first 3D GFX card. Road Rash, Destruction Derby 2 (?) and Scorched Planet (?) were included. It was magic. Puse awesomeness.
I have every flavour of Voodoo3 except the motherboard integrated one, one of which is in a permanent retro build & I run them all with fans either on or next to the heatsink because they all run hot, especially inside period correct cases & they always did.
Voodoo3 PCI supports both 3.3/5V PCI signaling (which is why there are two "keys" on the PCI connector). The motherboard you are using has 5V PCI (which is pretty standard). The AGP Voodoo3 connector is keyed for 3.3v. A Voodoo3 chip in PCI 5V mode probably uses a different buffering/re-driving mechanism than in 3.3v/AGP mode. What you ran into was probably the difference in sensitivity of those two bus voltage modes. That's my guess anyway.
I love 3Dfx Voodoo videos of yours. Very interesting to me. Subscribed for more content like this. Thank you
My STB V2000 works overlocked to V3 3500 183mhz with a small radiator + cooler (no other changes in bios/Windows except mhz). Only for tests ofc. Standard V2000 works for me ideally :) Great times, love Voodoo!
Great video, for more accurate core temperature readings I would suggest measuring the PCB behind the chip - due to the package type of the Voodoo3 chips most of the heat goes into the PCB and not the heatsink. On well used cards that spot can be discoloured from years of it running pretty hot!
Yes, I noticed that it is better to take the temperature on the back of the PCB, especially when a fan is blowing over the heatsink.
I am really surprised you run them without a fan. I installed fan immediately to both voodoo 3 cards I have for sake of longevity when I got them.
I will put a fan on those cards. For now, I just want to know how they were sold to the end user. I absolutely agree that they are very hot running cards and a fan should be put on them.
Ex-3dfx engineers are proud of you
Thank you!
Very informative. I love 3Dfx hardware. These videos are most of interest to me. I have 3dfx cards collections. Love these cards.
It's also possible to put a little bit of black electrical tape on the heatsink in order to be picked up by the thermal camera. Great video, can't wait to see what shenanigans you get up to next!
i am pretty sure you could go more efficient on the V3-2000 (higher clocks or lower voltage) if you would replace its cooler with something more beefy. lowering the temperature will increase efficiency.
3dfx is the best retro cards
nice. I'd be interested to see some results with maximum underclocking and undervolting. How much cooler will the V2000 be at, say, 100 Mhz and 2.0 V?
Ha! Great idea!
just use some mat painters tape on the heatsink, then it will heat up the tape and the cam will not have the wrong reflection
Yes, that could work, but it noticed that I could measure the temperature at the back of the card as well. Maybe there are a few degrees difference, but it should be good enough.
Marker paint won't help with reflection?
May be timings are involved. Try flashing bios of voodoo 3 2000 and do voltage mod for lower voltage.
Interesting idea. The crash of the Voodoo 3 3000 did not feel natural. Will put your idea on the project list. Thank you!
Had they done this testing, they might have been able to offer *mobile* voodoo chips. That would have been a much better option than being full stack manufacturer as they tried to do.
I know there are utilities that can read the internal temperature measurement of the die, and tools that ran in the background to log the temperature measurements to a CSV file. However I don't remember if these old cards had the thermocouple and circuitry to read it built into the die. Perhaps it's worth investigating?
I have never found any information about thermal sensors in any Voodoo graphic chips. I think those came later, but would have been great to have those available back then!
u are very productive recently! keep up, go out sometimes! ;)
BEWARE: artefacts are a poor way of verifying that the GPU is functioning correctly. Vector/polygon calculation errors will begin way before any typical artefacts become visible.
OCCTPT is a benchmark that tests for such errors. Also "Artefact Locator" & Video Memory Stress Test" (both by Misha Cherkes) can be useful.
Very interesting!
just put a little spot of thermopaste on aluminium cooler and measure temperature ;)
I'm wondering if the RAM voltage needs to be adjusted too as they may be getting out of sync with the core.
The memory is powered by a 3.3 volt supply, which actually comes from the voltage regulator with the heatsink. I don't think there is an issue since the core is already receiving lower power.
@@bitsundbolts Oh yea, but I was just thinking, lower core voltage and the timings of the core to the memory..
Half Life best Played on 3Dfx and Aureal Vortex 2
Facts
Good serie
Maybe if I ever get my 1000 fixed up and if I were to slap a heatsink on it (1000s had no heatsink), I wonder if I could get it to 2000 or maybe 3000 speeds
I could imagine getting it to 2000 level. Reaching 3000 may be a bit of a challenge. Isn't a Voodoo 3 1000 also only equipped with 8MB?
@@bitsundbolts it’s a full feature core with 16mb SGRAM, I think some came with 8mb. There were also the velocity cards with only a single TMU and I knew they only had 8mb
I didn't know they had so many variants at the low-end. I guess they really tried to have a Voodoo for every pocket.
I wonder, if you could lower the voltages more, if you add the lowest ESR caps available on the market.
One video each day? Slow down a bit...
Yeah, it was a lot this week...
omg, just use one 120mm fan
3dfx so fast it's kinda ridiculous!