If you can afford it or somehow find a used unit, you can go Epyc for 12 channels.
8/12 channel DDR5 will beat DGX Spark in inference/decode even without a GPU of any kind, as it’s memory bandwidth bound, and the Spark tops out at ~240gb/s real world.
With some optimisation and maths, it’s entirely plausible to ach
You are paying an extraordinary amount of money for the convenience of a super small unit, with still mediocre software support, but at least a community. Expect to be crawling through forum posts regularly, as SM121/Spark has many quirks and ecosystem issues still.
Please don’t pay another 70-80% gross margins on top of already inflated DRAM prices unless you need. The Spark IS really nice if you want to test out ConnectX or if you really need something small and compact and quiet.
Also consider: used Adas or even Ampere NVIDIA workstation GPUs can come with a lot of VRAM and be “reasonable”, with CUDA.
It's a dense model, not MoE like e.g. Qwen 35b or Gemma 4 26B A4B. On a Spark it will be memory bandwidth limited
I haven't tried yet (working on it) but back of the napkin estimate puts it at around 15tok/s even after converting to NVFP4. Prefill would be much higher though. That 15tok/sec is pretty typical for dense models of this size:
NVFP4 Q/K/V/O and MLP projections: ~13 GB/token
BF16 attention gates: ~3 GB/token
BF16 LM head: ~2.5 GB/token
Total: ~18.9 GB/token
At 273 GB/s, that gives a bandwidth-only ceiling of about 14.5 tok/s; actual performance would be lower.
That's pretty respectable.
Still working on optimizing and cleaning up before I push it.