Single Slot GPU's

the problem with A4000s and most single slot GPUs (models With display outputs), is that it doesn’t exhaust out the back of the case. It’s mostly exhausting out the small slots on the cooler shroud near the back of the GPU. So when you have them butted up next to each other, there’s not really anywhere for the hot air to go. 
 

the Titan V, and most other dual slot blower cards actually have venting on the second slot, so it works better, even with them butted up to each other, since usually there’s still a small gab for airflow into the fan. This is even better if you have the cards in a high airflow case or server chassis. 
 

plus the Titan V in particular only uses about 120-140W (same as an A4000) when crunching BRP7 or O3AS so having the dual slot cooler on it is pretty capable, even with restricted airflow. 
 

the only single slot GPUs that would be appropriate in a 7x stacked setup would be Tesla/Instinct data center level cards without display output and good venting out the back of the case. 

_________________________________________________________________________

The AMD W7600 that I run on one of my machines is actually pretty good in that department. It still exhausts some heat to the sides and a blower fan to remove that is recomended, but most of the heat gets out of the case trough the rear panel which is basically a grid with many holes. Yes it would be better without ports but it's surpriseingly good even with the ports.

Although since I work on the machine I'd recomend for GW tasks a card with more than just 8GB of RAM. For crunching only it's fine. I have a RAC of about 900.000 on that machine with some CPU tasks.

Edit: power consumption of the GPU running at 3x is ~110 watts.

i think Tom had trouble properly troubleshooting what the exact issue was. probably bad risers, or bad BIOS settings, or both.

I know he previously used some "EZDIY FAB" brand risers, which are *OK* for gen3, but you really have to QC them yourself to make sure they arent bad. I use/used them and had to return/replace maybe 20-30% of them before I got all good ones. most of the no-name brand risers longer than about 30mm are junk and wont even work at gen3 speeds.

with the Asrock EPYCD8 board, that is not gen4 capable, you have to be careful when using gen4 GPUs with gen3 risers and Rome CPUs. since Rome CPUs support gen4, and when using a gen4 GPU (like rtx 30-series and up) with gen3 risers, if the BIOS is set to auto for link speed, there are no limits in the BIOS to prevent the GPU from trying to run gen4. so you need to make sure the slots are set to gen3 explicitly in the BIOS in this case. Tom shouldnt have that issue with Titan V GPUs though since they will never try to run gen4, as they are limited at the GPU itself. this is technically a BUG in BIOS logic in my opinion, but the easy workaround is to just set all slots to gen3 speeds.

risers are no problem when you set it up correctly. my 8x Titan V system is running the sketchy EZDIY risers, but I QC'd them to ensure they all work, and they aren't that long. most are like 20cm. my 8x V100 are also on risers, though a different kind of setup.

linkup makes good cables from what I saw. but they are more expensive per cable. especially the gen4 rated ones.

I would encourage Tom to try them out, and just eat the cost. it'll be cheaper than a whole new platform. and a lot quieter.

_________________________________________________________________________