Two GPUs are not one bigger GPU: multi-GPU and unified memory explained

No, two GPUs do not add up into one bigger GPU. Two 24 GB cards do not create a single 48 GB VRAM pool. If you run two copies of a model, each card needs the whole model. If you split one model across both, each card still has to fit its share plus its own KV cache and communication buffers. And NVIDIA DGX Spark’s 128 GB is unified memory shared with the operating system, not 128 GB of dedicated VRAM.

Three ways to use more than one GPU

Two copies (replicas)

Each GPU loads the full model. You get more throughput, for example two users at once, but the biggest model you can run is still whatever fits on one card.

Example: Qwen2.5-72B-Instruct at GGUF Q4_K_M is 44,010,465,472 bytes (about 41 GiB) of weights. An RTX 4090 has 22 GiB usable after the default reserve, so one card won’t hold it, and two cards running copies won’t either: each copy is still 41 GiB.

Tensor parallel (split each layer)

Each layer’s weights and attention heads are sliced across GPUs (tp=2 in vLLM). This is how you run one model that doesn’t fit on one card. Fit is still checked per card after the split, and capacities are never simply added. The calculator only estimates tensor parallel for engines that document it (vLLM, SGLang and llama.cpp); other joint layouts are reported as deployment estimate incomplete rather than as a pooled total.

A fast link between cards helps tensor parallel, but whether NVLink is required depends on the runtime, the model and the workload. vLLM’s own guidance is on vLLM parallelism and scaling. Worked numbers for two RTX 4090s are in two RTX 4090s and tensor parallel.

Pipeline parallel (split by layers)

Consecutive layers go to different GPUs. This tolerates slower links between cards, at the cost of idle time while one GPU waits for the other. Each card still has to fit its layers. The calculator does not invent a layer split you didn’t set up.

DGX Spark: unified memory

Spark’s memory is one pool that the CPU, GPU and operating system share. It is not “128 GB of VRAM plus host RAM”, so don’t count system RAM again on top. Product details: NVIDIA DGX Spark (catalog source, fetched 2026-09-20). Per-model fits are on the DGX Spark hardware page.

Several Sparks

Two Sparks are not one 256 GB device, and four are not one 512 GB device. NVIDIA says multiple DGX Spark systems can be connected for larger workloads (DGX Spark product page, checked 2026-09-20), but connecting machines is not the same as pooling their memory, and it is not a verified way to split one model across them. Until there is one, the calculator reports two-node, 2+2 and four-node layouts as deployment estimate incomplete instead of a likely fit.

When you set up a multi-Spark scenario in the calculator, keep these apart: independent copies, how many devices each copy uses, and whether a joint split is actually supported.

Mixture-of-experts models

Mixtral-8x7B-Instruct-v0.1 in BF16 safetensors is 93,405,585,408 bytes (about 87 GiB), even though only about 12.9B parameters are used per token. Every expert stays loaded, so a 24 GB card won’t run Mixtral BF16 no matter how small the active parameter count looks. Details in why MoE models load every expert.

Numbers are estimates from model configs and published file sizes, not measurements. How we calculate.

Numbers from calculator 0.3.0, model data 2026-09-20.1. Sources last checked 2026-09-21.