I have a mixed use home server built in the Sagittarius NAS case, which is limited to GPUs ~240mm long. My Dell RTX 3060 12GB works great for transcoding, but I’d like a little more VRAM to play with models larger than ~8GB, and my specific mobo/case combination don’t let me easily install a second dual slot GPU.
To my understanding, the ASRock R9600D/R9700S Passive and Arc Pro B60/B65 Passive seem to be the only sub-240mm >12GB GPUs available, but they aren’t available from typical retailers and I can’t find any on eBay/AliExpress. Are they available from any lesser-known resellers or etc? I understand they’re passive GPUs, but I would have ~50-60mm available between the case fans and GPU body to whip up a custom fan duct.
Alternatively, does anyone have information on tearing down the Creator versions of the B60, B65, or R9700? The B60 Passive teardown post shows a fan connector and a PCB cutout that would line up with a Creator’s blower fan, and common dimensions between a bunch of Passive/Creator SKUs could suggest that ASRock is reusing one PCB/cooler design across a lot of SKUs.
Passive B60/B65/R9600D/R9700S and Creator B60/B65/B70/R9700 all share the same 112mm width, with the Creators sharing 271mm length/39mm height and the Passives sharing 190mm length (and matching heights within 1-slot and 2-slot Passive SKUs).
If the Creator’s fan and outer shroud can be removed, leaving only the PCB/heatsink in a Passive’s footprint, then it might be possible to “shuck” the Creator SKUs and retrofit them as Passive versions, but I can’t find any teardowns of the Creator models to confirm if the heatsink is part of the shroud, or if the backplate or any midframe could cause issues.
The “correct” answer is to use a larger case like the Silverstone CS382, or split one server into separate storage/compute servers, but the Sagittarius’s ability to fit in IKEA Kallax shelving is quite nice for my small apartment. Swapping to the 64GB Framework Desktop mainboard and accepting slower Strix Halo inference is probably the smartest compromise between size/cost/LLM performance, but I hate to give up my existing motherboard’s onboard IPMI.
Just my 2c but looking for an extremely uncommon footprint card with the expectation that you’ll be able to jerry rig a cooling solution into (what appears to be) a very tightly packed case sounds like (1) an expensive guessing game and (2) unlikely to produce the kind of results you are seeking.
Would it be realistic to expect you to hammer 12gb vram with a larger model while transcoding something? Or maybe I’m grasping this stick from the wrong end. If yes, I would have some real concerns about the temperature in conjunction with your drives.
Were it me I would just bite the bullet and upsize the case, removing all the constraints imposed on the kinds of cards you can install.
I wouldn’t be transcoding and using LLMs at the same time, I just want the ability to toy with larger 16GB-20GB+ models and longer context windows. Sagittarius has two separate cooling chambers for the drives/system, so system temp wouldn’t harm drive temps and vice versa.
It would work, but not worth the cost imo. Cheapest B50 I can find is $599 and gives me only 30% more VRAM, while the ASRock Creator B60 is currently $649 and doubles my VRAM to 24GB.