Two GPUs are not one bigger GPU, and unified memory is not extra RAM stacked on VRAM

Independent replicas, tensor parallel, and DGX Spark unified memory are different topologies.

Updated 2026-09-19. Figures use calculator 0.1.0 unless the article says otherwise.

Two 24 GB cards are not a 48 GB card. This calculator refuses to treat deviceCount = 2 plus “single device” as a merged pool.

If you run independent replicas, each replica still needs a full copy of the weights. Total VRAM across cards is not the budget for one copy. If you need tensor or pipeline parallel, this version does not invent a per-GPU split. It reports that a sharding plan is required, following the spirit of vLLM’s parallelism docs.

Unified memory is a shared CPU/GPU pool. NVIDIA DGX Spark is documented as 128 GB LPDDR5x coherent unified system memory, not 128 GB of dedicated VRAM. This catalog uses 128 GiB with a 16 GiB default reserve for OS and other processes. Host RAM is not added on top. Two Spark units are not one transparent 256 GB device.

Mixtral-8x7B-Instruct-v0.1 BF16 safetensors total_size is 93,405,585,408 bytes (~87.0 GiB) even though about 12.9B parameters are active per token. The default estimate keeps all experts resident. That is why a 24 GB card is insufficient for Mixtral BF16 regardless of “active parameter” marketing.

Calculator 0.1.0, dataset 2026-09-19.1. No tokens/s figures are derived for Spark or multi-GPU here.

Open the calculator