Hi NVIDIA team,
I am considering purchasing an NVIDIA RTX Spark 128GB Windows system specifically for local ComfyUI workloads.
I need confirmation about one exact configuration:
MiniMax H3 FULL/native:
- minimax_h3_fl2va_bf16 (~66.3 GB)
- Qwen3-VL-32B BF16 text encoder (~51.5 GB)
- native non-quantized video VAE
- native audio VAE
- no INT8
- no FP8
- no NVFP4
- no GGUF
Questions:
-
Has NVIDIA tested or validated this exact FULL H3 configuration on ONE RTX Spark 128GB under Windows 11 ARM?
-
If yes, what was the peak unified/CUDA memory usage?
-
What is the practical maximum CUDA/PyTorch memory available to one process on a 128GB RTX Spark Windows system?
Can one workload reliably use 100GB, 110GB, 115GB or 120GB? -
Is staged loading/offloading in ComfyUI expected to make this FULL configuration feasible, or is the available Windows memory budget too small?
I am specifically asking about the full native models above, not quantized versions.
Thank you.