I don't see why testing at lower resolutions and quality settings would put more load on the GPU.

Just look at ATITool as an example.

The Serpent Rider wrote:

Pro tips: UT99 - bad benchmark to heat up GPU, due to being extremely CPU heavy game. 2x AA and high resolution - limitations of […]
Show full quote

Pro tips:
UT99 - bad benchmark to heat up GPU, due to being extremely CPU heavy game.
2x AA and high resolution - limitations of the ROP performance and memory bandwidht are not optimal for full GPU load.

Conclusion: needs better CPU and GPU heavy game at low resolution without AA.

Valid tips for faster GPUs, but in this case I think CPU limit is not a problem (it was running at 10 - 15 fps at most).
Well, VSA-100 doesn't have much else besides ROPs and memory controller. 😀 At least I can be sure at this settings there is definitely no CPU limit.

Anyway, now I have temperatures for Voodoo5 as well:

V5 5500 PCI with original heatsinks and fans = 71 °C (primary GPU) and 70 °C (secondary GPU)
V5 5500 AGP with original heatsinks and Sunon Maglev fans = 62 °C (primary GPU) and 61 °C (secondary GPU)

havli wrote:

Anyway, now I have temperatures for Voodoo5 as well:

V5 5500 PCI with original heatsinks and fans = 71 °C (primary GPU) and 70 °C (secondary GPU)
V5 5500 AGP with original heatsinks and Sunon Maglev fans = 62 °C (primary GPU) and 61 °C (secondary GPU)

Nice!

Very significant improvement just with the fan. Seems well worth at least replacing the fan. The card I'd purchased had had the original heatsinks removed and replaced with waterblocks. I removed those and installed Enzotech SLF-1 Ultras. I do not have measurements for the original, of course.

Look who got their own Voodoo4 4500? 🤣

I'll be contributing some of my own numbers here shortly, after I buy that replacement fan you linked havli. This will be my open-air test bench that I'll be conducting the tests.

This will make every game/benchmark entirely GPU-limited, and should push the V4 to its limits. That should help a bit. 😀

EDIT: Just bought the fan. Should be arriving at the latest next week. Wooh!

As have I - so thanks for that tip, BTW. My Voodoo 5 fans were sounding a little crusty. I'm glad there's an easy-to-source alternative.

It should make a difference if the V5-AGP is 3.3V or the newer 1.5V ?
I have been told that a Noctua NF-A4x10 5V can also fit as a "more silent" replacement?

dr.zeissler wrote:

It should make a difference if the V5-AGP is 3.3V or the newer 1.5V ?

There's no performance differences between AGP cards keyed for 3.3V, 1.5V, or either, AFAIK, so if there were two different versions of the V5 with different voltage keys (which, again, AFAIK, I don't believe there are), there wouldn't be any difference between them except their compatibility with motherboards.

dr.zeissler wrote:

I have been told that a Noctua NF-A4x10 5V can also fit as a "more silent" replacement?

Looks like a nice fan, but I think why that isn't a potential replacement is because of the fan's thickness. It's most likely too thick to be used with the screws used to attach the stock fans on the V4 and V5. The only way you would be able to use such a thick fan is if you found longer screws of the same diameter and threading as the stock screws. I would need havli to verify these statements, but I think I'm for the most part correct.

EDIT (1/29/2019) - I'm most likely wrong. Read below for clarification.

which, again, AFAIK, I don't believe there are

1.5v engineering sample does exist.

the_ultra_code wrote:

Yo

I made a video about safely removing the heat sinks. I did lots of research online before attempting this, and it worked great.

https://www.youtube.com/watch?v=5hB4jOCHLHc&t=2s

mzry wrote:

the_ultra_code wrote:

Yo

I made a video about safely removing the heat sinks. I did lots of research online before attempting this, and it worked great.

https://www.youtube.com/watch?v=5hB4jOCHLHc&t=2s

No way! You're the guy who did that video! I saw that video before. I didn't know you had a presence here on VOGONS.

I might attempt that someday, but for now I'll just stick to something a little less risky.

By the way, out of all of the cards you have popped the heatsinks off, what is the rate of failure, if any? 1:10, 1:50?

the_ultra_code wrote:

Look who got their own Voodoo4 4500? 🤣

I'll be contributing some of my own numbers here shortly, after I buy that replacement fan you linked havli.

This [test bench of mine] will make every game/benchmark entirely GPU-limited, and should push the V4 to its limits. That should help a bit. 😀

So, I finally got around to doing some temperature tests on my Voodoo4. I got in the mail a cheap thermal couple thermometer with K-type couples to test with, and this is what I conducted the tests with:

I placed the first couple (T1) on the back of the PCB right "underneath" the GPU die, and the second one (T2) on one of the memory modules, also one the back of the PCB (because I attached some heatsinks on the RAM chips themselves, 'cause why not? ; factor that fact into the results).

This is how the card looked for testing. Like I said, I had attached some heatsinks on the RAM chips using Sekisui #5760 thermal tape, and have cleaned the entirety of the stock fan and made sure the main heatsink was clean. The thermal couples were attached to the GPU also using the same thermal tape, "stacked" as such - PCB > thermal couple > thermal tape.

So, here are the results, followed by an explanation on how I did the tests. I used 3DMark 2000 (all tests; 1280x1024 @ max settings; on loop for 30 minutes) for placing a load on the GPU for these tests. Under Windows 98SE, the power plan was set to "Always On", and neither the screen nor the SSD were allowed to "sleep".

Room temperature (as measured by a thermal couple) - 21.8 C

*Temps in Celsius _ _ | Behind GPU die (T1) | Behind a RAM chip (T2) |
--------------------------------------------------------------------------------
Idle, stock clocks_ _ _| _ _ _ _ _ _ _ _ 49.3 | _ _ _ _ _ _ _ _ _ _ 29.2 |
--------------------------------------------------------------------------------
Load, stock clocks _ _ | _ _ _ _ _ _ _ _ _ 55 | _ _ _ _ _ _ _ _ _ _ 33.8 |
--------------------------------------------------------------------------------
Idle, OCed (186MHz)_ | _ _ _ _ _ _ _ _ 55.1 | _ _ _ _ _ _ _ _ _ _ 29.7 |
---------------------------------------------------------------------------------
Load, OCed (186Mhz) | _ _ _ _ _ _ _ _ 57.7 | _ _ _ _ _ _ _ _ _ _ _ 32 |
---------------------------------------------------------------------------------

I first read the room temperature by letting a thermal couple sit for maybe 30 minutes. I then proceeded to attach the thermal couples to my Voodoo4, and gather the above temperature data. For "idle" temps, they were gathered after 30 minutes from a cold boot when the card was at or around room temperature, and then right afterwards I ran 3DMark 2000 on loop to gather the "load" temps.

The 186MHz overclock (for both the core and memory, since the Voodoo OCing settings in the driver software only allow you to adjust the clocks for both at the same time in unison) is most likely the highest, most stable OC I am able to achieve for the card. 187MHz was the actual highest stable OC where 3DMark 2000 was able to be completed in its entirety, with 188MHz and beyond displayed "snowflaking" during the tests, and resulted in 2000 locking up the entire system in doing so. The reason why I OCed my card is because a) I'm curious to see what clocks my specific card can achieve, so that later in the future if I plan to throw this card in a system, I can choose a more mild OC for long-term use (175-180MHz comes to mind), and b) so that I could see how hot the card would get with the stock cooling solution.

I tried to account for as many variables as possible, but I can't account for everything. It might be safe to assume an error margin of 5-10C, in light of the results that halvi was able to procure. Still, these temperatures are well within reason and totally fine for the health of the VSA-100 chip on Voodoo4/5s (or any GPU for that matter).

I will say that the fact that the GPU was as hot at idle with a 20MHz OC as it was under load with stock clocks, and only 2-3C hotter during load with the OC then it was during load with stock clocks surprises me. Assuming that I conducted my tests accurately, I would theorize that with stock clocks, there most likely is some sort of energy-saving feature or something similar to that that allows the GPU to run cooler while at idle.

What didn't see such weird behavior was the memory, which showed similar temps at idle and under load with and without the overclock. I guess that, unless you push the RAM to an extreme degree, it's really not going to deviate from stock behavior.

So, to answer my question, the V4 and by extension V5 runs at perfectly reasonable temperatures with the stock heatsink + fan. As halvi has showed, it does help to replace the stock solution with a custom cooler, but it is really not necessary. A simple fan replacement will suffice for better results.

Which brings me to my situation on replacing the stock fan on my V4.

halvi, that Sunon fan you recommended... it's not what I thought.

It's a two-wire fan. It doesn't come with the correct header needed to plug into the card. Also, I noticed that the fan lets the screw "through more" than the stock fan does. That means that, without washer or glue, that fan will be loose on top of the heatsink.

Now, the good news is that you can buy the required fan header needed to connect the fan to the card from this ebay listing (I actually bought 2-3 of them, just in case). Also, I found this similar yet thicker Sunon 40x40x10mm fan on Digikey that I'm waiting on to come from the factory (there should be enough length to the screws that are used to hold the stock fan to accommodate the additional 4mm of height, but I'll find out once I get the fan in the next month or so; finding a 40x40x6mm fan that is similar to the stock one in terms of how far it allows the screws to go through is like impossible - none are really made anymore). Once I get a new replacement fan to my liking, I'll also conduct the same testing to see how much the temperature improves.

As I have said before, the more data, the merrier. 😀

Finally, a correction. dr.zeissler, that Noctua fan you suggested would be a fine replacement fan (in fact, I zip-tied that same exact fan on my Voodoo3 3000 to help cool it). Just that that would have to plug into the motherboard, and not into the card itself. Guess you could mod it to connect to the card, but, I mean, would you want to? 😜

EDIT - Oh and more thing. I just bought a Voodoo5, so I'll also conduct the same testing on first GPU die (closest to the VGA port) and report that data here too. 😀

Yeah, I am aware the Sunon fan doesn't have connector to fit V4/5. However I never considered this to be a problem - I just connect the fan directly to PSU (5V) and no problem. I also found out there is maybe 1mm of empty space between the fan and heatsink but function wise this should make no difference.

Using 40x40x10mm fan is possible of course but perhaps it will require longer screws. Also another solution could be to make converter from the stock VSA-100 heatsink to bigger fan like 50x50 or 60x60. Should be fairly easy to 3D print it.

Does the V4/5 care at all about the RPM sensor?

havli wrote:

Yeah, I am aware the Sunon fan doesn't have connector to fit V4/5. However I never considered this to be a problem - I just connect the fan directly to PSU (5V) and no problem. I also found out there is maybe 1mm of empty space between the fan and heatsink but function wise this should make no difference.

Using 40x40x10mm fan is possible of course but perhaps it will require longer screws. Also another solution could be to make converter from the stock VSA-100 heatsink to bigger fan like 50x50 or 60x60. Should be fairly easy to 3D print it.

Noted

SirNickity wrote:

Does the V4/5 care at all about the RPM sensor?

That is a great question. I might be able to answer that shortly, now that I have the connectors needed to plug in these replacement fans into the cards themselves.

Any who, I finally tested my new Voodoo5 5500, and it works! I cleaned the fans and heatsinks, and finished up cable management in my PIII-S 1.4GHz machine, which is the most powerful system I have that can physically take a V5, what with its universal AGP slot. The CPU isn't as powerful as the Core 2 Duo that I used in my earlier testing, but it will have to suffice for testing purposes. I'll do some temperature tests on my V5 here shortly, maybe even tomorrow. 😀

Looking forward to it. The fans on my V5 aren't dead yet, but they're probably not long for this world, so I snatched up a couple of those 6mm Sunons while I was ordering parts anyway. Been too preoccupied revitalizing some old PSUs to get to it. I'm really curious to see whether it's a real improvement inside a crowded case.

SirNickity wrote:

Looking forward to it. The fans on my V5 aren't dead yet, but they're probably not long for this world, so I snatched up a couple of those 6mm Sunons while I was ordering parts anyway. Been too preoccupied revitalizing some old PSUs to get to it. I'm really curious to see whether it's a real improvement inside a crowded case.

Mind you, I am only going to be doing side-panel-off testing, or else it A) would be very hard to get my thermocouples onto the card itself and B) would not be as comparable to my previous testing. However, this case has three fans in it, so even though the case is a little tight with all of the ribbon cables everywhere, that V5 would still be getting plenty of fresh air. 🤣

Right now, I'm waiting on a pair of new "surface" thermocouples, which Steve from Gamers Nexus has said works better than the rounded-tipped ones I have been using. Of course I'll see how much better they are, so that my previous V4 testing is still comparable.

I'm back.

So, I got my new surface thermocouples a few days ago, and shortly thereafter I began conducting thermal testing on my Voodoo5 5500. In these tests, I used my 1.4GHz Pentium III-S rig as my test system, as that is the most powerful system I own that has support for 3.3V AGP cards with its universal AGP slot. Also during the testing, I had the side panel off for ease of testing.

So, up first was to determine the accuracy difference between my bulb and my surface thermocouples. For reference, I'm using these thermocouples that I bought off of Amazon:

To test the difference, I attached both my bulb and surface thermocouples to the back of the PCB under GPU die #1, and measured the temperature of that die after 30 minutes of 3DMark2000, with idle time in between testing both thermocouples. I did use the pre-attached thermal tape that came with the surface thermocouple for this testing.

Bulb vs. Surface Thermocouple

Room temperature - 19.7 C
---------------------------------
|_ _ _ _ |_ _ Bulb | Surface |
|Celsius |_ _ 50.3|_ _ _ _59|
---------------------------------

As you can see, there was a *massive* difference between the readings between the two types of thermocouples. This effectively makes my prior testing with my Voodoo4 inaccurate by at the very least 4-5C I would say, and reinforces that surface thermocouples are truly the best for this kind of testing.

The next bit of testing I did was more for later reference. I wanted to see how my Sekisui #5760 thermal tape compared to the film-like thermal tape that came pre-attached to the surface thermocouples, since, if I'm going to use these thermocouples more than once, I'm most likely going to use new thermal tape each time so that the thermocouple is most strongly attached to the surface I'm measuring. I again used GPU die #1 as the part of the card I was testing with, but this time, I also measured temperatures after 30 minutes of being idle in the desktop.

Pre-attached Thermal Tape vs Sekisui Thermal Tape

------------------------------------
|*In Celsius | Pre-at. | Sekisui |
|Room Temp|_ _ _22|_ _ _21.7|
|Idle_ _ _ _ |_ _ 55.6|_ _ 53.6|
|Load_ _ _ _|_ _ 60.9|_ _ _ 58|
-----------------------------------

From the results, it does appear that the pre-attached film-like thermal tape that came with the thermocouple was a little better at transferring heat to the thermocouple itself than the Sekisui by 2-3C. However, it's good to know the difference here-out, as I did use Sekisui thermal tape for the rest of the testing. So, for the rest of the testing that I use Sekisui thermal tape, tack on another 2C to the results. For future testing, I'll try to secure some of that what I'm going to call thermal film, as that seems to do the best job at heat transfer. If someone can tell me what exactly that stuff is and show me were I can buy it, that would be awesome. 😀

Finally, after all of that confirmation testing, I got to actually testing the thermals of the V5. Like with my V4 testing, I measured temperatures from a cold boot after 30 minutes of sitting idle at the desktop, and then 30 minutes of 3DMark2000, and I used Sekisui thermal tape to interface the thermocouples to the card. And for the overclocked results, I ran the V5 at 181MHz core+memory combined (the highest stable OC I could achieve - anything higher would lock up the system, even though there was no graphical artifacts during testing).

First up was GPU die temps. I attached one thermocouple to the back of GPU die #1, and another to GPU die #2.

GPU Temperatures

Room temperature - 21.7C
-------------------------------------------
|_ _ _ _ _ _|GPU die #1|GPU die #2|
|Stock Idle |_ _ _ _53.6|_ _ _ _50.7|
|Stock Load|_ _ _ _ 58|_ _ _ _ 57.2|
|OCed Idle |_ _ _ _55.4|_ _ _ _53.3|
|OCed Load|_ _ _ _58.4|_ _ _ _57.7|
-------------------------------------------

These are actually good numbers, and are much closer to havli's results than my previous V4 testing was. 60C under load is completely acceptable for modern GPUs, and I would strongly believe the same for this almost 2-decade-old GPU.

These results, though, go counter to the results I noticed in my previous V4 testing - that stock load temps equaled OCed idle temps. I assume that either room temperature changed significantly testing, or something was up with the thermocouple. From what I have heard, the V4/5, or any GPUs of that era, sometime after, and especially before, never had any sort of turboing behavior. I'll retest my V4 in this same system again with the same surface thermocouples to see if I can get similar behavior, but I doubt I will.

Next was memory temps. Unlike with my V4 testing, though, all of the memory chips are bare. I tested one memory chip, placing one thermocouple on the chip itself, and another on the back of the PCB behind the chip.

Memory Temperatures

Room Temperature - 21.3C
-------------------------------------------
|_ _ _ _ _ _|On Memory|Behind Memory|
|Stock Idle |_ _ _ _ _36|_ _ _ _ _ _ 35.3|
|Stock Load|_ _ _ _ _39|_ _ _ _ _ _ _ 36|
|OCed Idle |_ _ _ _37.7|_ _ _ _ _ _36.7|
|OCed Load|_ _ _ _39.2|_ _ _ _ _ _36.1|
-------------------------------------------

Now, I'm no expert, but I think 41C is a completely safe operating temperature for memory. These numbers also back up the original behavior I noticed in my V4 testing - memory temperature is pretty stable from stock to overclocked.

I would like to say again that I tried to account for as many variables as possible, but I can't account for everything. It might be safe to assume with this testing a small error margin of at most 2-3C after taking into account the 2C difference between the thermal film that came with the thermocouples and my Sekisui thermal tape.

So, to sum things up (again), the Voodoo4/5s operate at perfectly reasonable temperatures. Nothing to be concerned about with the stock cooling solution.

whew Now that was a lot of work, and I'm still not done. 🤣

Like I mentioned above, I'll be reconducting my V4 temperature tests shortly. I need to buy some new surface thermocouples, because one of mine snapped just after I finished testing and I went to remove the thermal tape. I'll also be testing different ways to attach thermal tape to the thermocouple to see which is the best for getting accurate temperatures.

Again, if anyone else can share any of their own data, that would be awesome.

That was great, man. Thanks for going to the trouble.

I really don't know anything about common GPU testing procedure, so I'm going to ask a dumb question: Why measure the PCB behind the GPU? Fiberglass thermal conductivity isn't exactly fantastic, right? Has this shown to be the hottest or closest to die temp? I can think of two things that might contribute to this: 1) Heat rises, and there's very little to cool this area, so it could become a hot spot; and 2) The heatsink temp will be affected by the active cooling from the fans. Am I close?

SirNickity wrote:

That was great, man. Thanks for going to the trouble.

I really don't know anything about common GPU testing procedure, so I'm going to ask a dumb question: Why measure the PCB behind the GPU? Fiberglass thermal conductivity isn't exactly fantastic, right? Has this shown to be the hottest or closest to die temp? I can think of two things that might contribute to this: 1) Heat rises, and there's very little to cool this area, so it could become a hot spot; and 2) The heatsink temp will be affected by the active cooling from the fans. Am I close?

Trust me, I'm not an expert either. Just try to do things that make logical sense. 😜

As for the first three-ish questions: The back of the PCB behind the GPU die on Voodoo3-5s is the easiest way to get the closest-to-actual-die temperatures AFAIK. However, I do believe you are correct when you said:

SirNickity wrote:

Fiberglass thermal conductivity isn't exactly fantastic, right?

This article comes to the conclusion that a PCB is a horrible conductor of heat. Obviously, the only way to get the best realistic measurement of how hot the GPUs are getting is to put a thermocouple on one of the GPU dies themselves. Problem is, there's a heatsink that's been thermal epoxied to the GPU die itself! 😠

I think the only way someone would be able to get the most accurate measurement of how hot a GPU die gets on a Voodoo3-5 with the stock cooling solution would be to remove the stock heatsink, completely clean off all of the stock thermal epoxy off of both the heatsink and the die, and then reapply a new application of thermal epoxy close in specifications to the original between the heatsink and die, and have a thermocouple directly touching the die during that application. Of course, I don't think anyone (me included) would want to do that, but if someone wanted to do that, it would be a lot more informative than whatever we are doing.

Now, I can try to put a thermocouple in-between the heatsink and die somehow, but I believe (if I can do that) it would be along the edge of the die, and might not get the most accurate reading. I'm willing to try that with my Voodoo4 with my next batch of tests to see if that nets any higher temps that gathering reading from behind the die would, though.

As for your explanations as to why the back of the PCB behind the GPU die gets hot (reinterpreting your second reason to read "the GPU die temperature will be affected by the active cooling", as the temperature of the heatsink isn't important), I believe yes, you are right. Those are some good reasons why the back of the PCB gets hot, and could be hotter than the GPU die itself (assuming that the die's plastic housing and thermal epoxy are doing a good enough job of transferring heat away from the die itself and to the heatsinks).