So ignore the fact that my DGX Sparks are currently naked… I plan to now print or have a new case made up for both.
One is a Gigabyte AI Top Atom and the other an MSI Edgexpert. Both were hitting 82 degrees and throttling.
I decided in my ADHD Infused Ferret wisdom to see if there was some way to improve the cooling, I had seen the ducts but I wanted to see whether the fans inside could be replaced.
What I found was the intakes jammed against the metal casing with the Giagabyte actually having foam around to further decrease the fan efficiency. So I decided to test without the casings and to my surprise, and Geminis surprise, we saw a 16 degree drop on average:
Next stage will be designing a case and added mesh filters. A dremel would also work but I didn’t want to tank the resale value!
"Thermal Status (Post-Cowling Removal):
Temperature: 64°C. This is a 16°C improvement over the previous 80°C peak.
Clocks: Pushing 2444 MHz without any signs of a major thermal dip.
VRAM: KV cache usage is hitting 81.5% for concurrency=10 at 4k depth. We still have plenty of headroom for the 100k depth tests coming up."
I’d be curious the SKUs on the fans and if there are better fans, more efficient higher CFM lower power draw that could replace them, we know MFG are looking for just enough to meet the minimum need vs premium fans. Just did that with the hot plug fans on my MikroTik CRS804-DDQ switch much quiter better CFM and less power draw
Ive literally just done a similar thing - stripping off the case from my GX10 has taken off 15c of max temps. I am planning to create a perspex enclosure with fans on each side, which should hopefully make it run even cooler. Caveat is that WIFI (which I dont use anyway) antenna is on the case, so you lose that… also, I had to switch my bios to auto power up my spark when it gets PSU power, as the power switch is also attached to the case.
wasnt there a test that even the different offbrand gb10 throttle the same way even with better heatsinks ? i think thats a dead end .. the new firmware manages heat pretty decently - run my 8 at full load for like 24/7 without a problem .. often its the brick that did not get the firmware update -and then they act weird - you update there firmeware fully unplug from the socket (important for about 5-10) min and try again .. but test under load not just idle
there is a youtuber that tested i think 6 different sparks and they all perform ± the same even with better thermals initially
I use a pair of smart sockets to power on / off my sparks, coupled with auto-power on setting in BIOS, this way I can fire them up remotely if I want.
As far as airflow is concerned, at least on Asus GX10 - front grille is cosmetic only, it has no air gaps, so those 2 puny fans have to suck up air through the bottom of the case, which probably impacts their effectiveness further.
So probably swapping the fan out to something like
Sanyo Denki 9BMC12P2H01 - 12V, 4-wire PWM
gets you the same size fan to fit in the housing which gives you a fan with the same power draw as the delta but higher torque to overcome static pressure from the copper fins. The deltas are notorious for losing RPMs vs back pressure. Also the Sanyo has a 3 phase drive circuit which gives better balanced drive rotation, and and precision ball bearings vs a drive sleeve higher performance and less noise in the same footprint. Of course it would mean removing the fan from the housing and wiring in the new one either with heat shrink solder or true solder connections. The Sanyo is also rated at 20 CFM vs the delta at 15-16CFM so a 20-25% increase in CFMs
just a thought to help improve the cooling with the same footprint. Would of course need to vet it all out. But it’s almost the same use case I did with my CRS804 Switch, swapped out the Delta fans in how swap housing with same foot print better quality Sanyos, saw a 15-18% temp drop and major noise drop for a pair of 28.00 fans and about 2 hours of my time.
I have a MSI Edgexpert, so interested in your findings. The MSI is supposed to have slightly “better” thermal piping than other variants of the DGX. Have you noticed any differences in operating temps between the two? Also can you see any internal differences re the thermal pipework?
