Considering Selling 5090 for Other Hardware - Thoughts?
retained text
Hey everyone, Here are the specs of my main workstation: * 5090 FE * 14700k * 96GB DDR5 Ram (Two Sticks) Specs of my inference server: * Strix Halo - 128GB (Framework Desktop) * GUFO + Qwen3.8 Flash Next * Embedding Models * STT/TTS Models I also have some other non-AI servers for hosting containers + offsite cloud backup + NAS. When I upgraded my current PC prices were basically MSRP (that's what I paid for the FE and RAM was $400). It was intended to be a very functional work machine for heavy usage + gaming on the side. Now that prices have skyrocketed, I'm seriously considering selling the 5090 and replacing it with a 5070ti + another dedicated machine. The Strix Halo is excellent, but I'm running a business and want my employees to be able to also make use of everything that I've built out infrastructure wise. A single machine is great for me, but you can't reasonably scale it. We do a lot of sensitive client work where we can't use cloud models, so local models are amazing for this. **My goals are:** * Create a more easily "scalable" solution that can grow over time. * Be able to serve at least three concurrent users with 2 or 3 parallel sessions for each (so tool calling works per-user). * Be able to run Qwen3.8 FN equivalent @ 200k context windows. **Sacrifices I'm aware of and willing to make:** * Speed - Obviously this won't come close to cloud infrastructure, but the Strix Halo in its current implementation has been good enough for my usage. 1400 prefill, 50 token average for generation. * Intelligence - We can still use cloud models for internal work, these only replace customer-specific coding and scripting, which require far less complexity. * Pooling hardware - I don't know how much this will impact us in the future, but as of now I'm not sure I require something like the DGX Spark interconnect. **Current Options:** 1. Buy another Strix Halo. I don't actually need to pool the machines together so they can just run independently with custom user profiles. For the price of the 5090 I could actually buy two of these. 2. Buy a 256GB Mac Studio. I might have to cover a cost difference, and this will now force me to upkeep two completely different stacks of infrastructure - but will allow for potentially easier scalability in the future + bigger models. 3. Buy a 5070ti, and re-work my current machine as a server. I already have enough components on hand to build a new PC without a GPU, so I could use the 5090 with 96GB of RAM as a new machine. 4. Buying 4 Radeon R9700s. This also increases my cost because now I have to get a new motherboard, PSU, and a bunch of other stuff. 5. DGX Spark - When I first considered what machine to buy, the spark was about $1000 more than the Halo, which wasn't a justifiable cost for the performance benefit. At the moment it's so outrageously expensive for what it is there's no shot. **Things I won't consider:** 1. Building used. This is for my business and so I want to have a warranty. **A few things I don't know or need to speculate about:** * How good is the 5090 + 96GB RAM with newer software improvements today? It's been a while since I've seriously done any AI work on this. * Is 128GB going to be a problem down the line, and so being able to pool hardware together to run bigger models is a key influence? Any advice is appreciated. My gut is telling me to just get another Strix Halo since I've been using it long enough + have all of the software components dialed in.
document fetched 2026-10-27 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xg4cj4/considering_selling_5090_for_other_hardware