[mft_]

(I'm not one of the people you're speaking of with a 128gb M5 but) if you want to run one of the medium-sized open-weights models (Qwen 27b, 35b, Gemma 4 26b, 31b) or larger, you get into an interesting optimisation space.

* yes, you can run it on an older/smaller GPU plus system RAM but performance will suffer

* if you want optimal GPU performance you need the model in VRAM plus context, so 24GB (3090, 4090) or 32GB (5090) cards, plus a system that's reasonable powerful to plug them in to. Ideally you'd have a multiple cards working together but for optimal performance this means either 2x 3090 or nvidia's workstation cards.

* you can go for a 128gb Strix Halo system, but the memory bandwidth isn't great and they're becoming increasingly more expensive (5.5k EUR for HP laptop, 3.9k EUR for GMKtec EVO-X2 mini PC)

* you can go for a 128gb DGX Spark (5k EUR+) which also has unspectacular memory bandwidth or RTX Spark (price unclear but probably not cheaper)

* or go for a Mac with a decent CPU and a good amount of RAM (bandwidth varies by model, but typically a bit better than Strix Halo/DGX Spark and worse than bespoke GPUs.

As usual with such questions, there are of course cheaper paths (if you want to accept the tradeoffs) but Macs are reasonable vs. competition for these workloads.

[edb_123]

I just recently got into experimenting with local LLMs when I had anyway (for non-LLM reasons) built myself a new desktop system with Intel Ultra 270K-Plus and RTX 5080. With 64GB system RAM and 16GB VRAM. Relatively speaking a high-performing and low-to-moderate cost system.

I wasn't really expecting much from these local open weight models neither when it comes to speed or "intelligence", but my preconceptions were quickly put ashame when I got ollama up and running and pulled my first model. I get a consistent 117-128 t/s with Gemma4:26b-a4b without any tuning (just the default settings), which was much faster than I had expected. Can't wait to dive deeper into this, especially with Qwen3.6 models.

Does anyone's have experience adding a 2nd Nvidia GPU of the same generation but different (slower) model in the same system? Will it give a major boost with larger models, or will the slower card just be a bottleneck? I have an unused RTX 5060 Ti 16GB that I'm considering to install alongside the RTX 5080, but it would necessitate removing some other hardware, so I haven't bothered yet.

[skolos]

I'd say adding another 16Gb gpu would be worth it - you'd be able to run larger model/larger context all within gpu's. It would give you more options of what you can run fast. Your current model probably doesn't run completely from GPU (depending on quants I don't think you can squeeze Gemma4:26b into 16Gb vram), so you already have some layers running on gpu and some on cpu. If you add another gpu you might be able to move all layers to vram which should speed up things for you. The layers calculations happen on whatever gpu's it sits, so the layers that are already on your rtx5080 would compute same, but the layers that currently your cpu handles will be computed with faster vram/compute of rtx5060.

[edb_123]

Thanks! I'm seeing a 10/90 split between CPU/GPU with gemma4:26b, so I guess there's at least something to win there by adding the other GPU. And perhaps something to win by connecting the monitor to the iGPU instead to free up VRAM, from what I gather.

Just in case someone should be interested in how a consumer PC setup like this performs, still using only 1x RTX 5080 + 64GB system RAM and Intel Ultra 270K-Plus; I tested Qwen3.6:35b-a3b now (using ollama and default settings) and I'm getting around ~86 t/s. The lowest I've seen so far is 70 t/s. The CPU/GPU split with 35b is 39/61% (with 4K 165 fps monitor connected to 5080, so there's probably some room for optimization here by moving it to the iGPU).

Best thing is that this setup is basically dead silent (it could, hypothetically speaking, be running in my bedroom just fine, and I'm a light sleeper).

[pletnes]

And with a mac, there are no cuda drivers to fiddle with.

[girvo]

But prompt processing is terrible