DGX Spark vs RTX 5090: usable memory, bandwidth, and catalog fit

GeForce RTX 5090 32GB (desktop) is discrete VRAM. NVIDIA DGX Spark (GB10, 128 GB unified) is unified memory. Discrete VRAM and unified memory are not additive and are not the same pool. Net budget is catalog capacity minus the catalog default reserve: 30.00 GiB for RTX 5090 (32.00 GiB capacity, 2.00 GiB reserve) and 112.00 GiB for DGX Spark (128.00 GiB capacity, 16.00 GiB reserve). Memory bandwidth: RTX 5090 not in the catalog; DGX Spark not in the catalog. The columns are not ranked.

No decode ceiling is shown. Bandwidth is not in the catalog, and a missing bandwidth is not a memory verdict.

Catalog memory

NVIDIA DGX Spark (GB10, 128 GB unified)GeForce RTX 5090 32GB (desktop)
Memory kindunified-memory unified memorydiscrete-vram discrete VRAM
Marketed label128 GB LPDDR5x coherent unified system memory32 GB GDDR7
Capacity137,438,953,472 bytes = 128.00 GiB34,359,738,368 bytes = 32.00 GiB
Catalog default reserve17,179,869,184 bytes = 16.00 GiB2,147,483,648 bytes = 2.00 GiB
Net budget (capacity minus reserve)120,259,084,288 bytes = 112.00 GiB32,212,254,720 bytes = 30.00 GiB
Memory bandwidthnot in the catalognot in the catalog

Software listed on the hardware record

Same models on both devices

File-backed quant only, context 8192, 1 sequence. The runtime is that quant's preferred engine. The range and tier come from the calculator. Rough means the judge marked it rough. A missing quant is left blank and is not replaced.

ModelQuantRuntimeDGX SparkRTX 5090
Qwen2.5-7B-InstructGGUF Q4_K_M
gguf-q4_k_m
llama.cpp4.92 GiB–5.42 GiB · comfortable comfortable4.92 GiB–5.42 GiB · comfortable comfortable
Qwen3-8BGGUF Q4_K_M
gguf-q4_k_m
llama.cpp5.93 GiB–6.48 GiB · comfortable comfortable5.93 GiB–6.48 GiB · comfortable comfortable
Qwen3.8-27BGGUF Q4_K_M (mradermacher)
gguf-q4_k_m
llama.cpp16.43 GiB–19.74 GiB · comfortable · rough comfortablerough16.43 GiB–19.74 GiB · comfortable · rough comfortablerough