Start with the workload, not the ceiling

On Nov. 18, NVIDIA said compact RTX Spark desktops from ASUS, Dell and MSI were available, with configurations offering up to 128GB of unified memory. S1 That puts dedicated local-AI desktops on the market; it does not settle what workload a machine can sustain.

Unified memory is a shared pool: model weights compete with the operating system, runtime, context state and other open apps. “Up to 128GB” is therefore a system ceiling, not a promise that a chosen model, long context and creative application all fit at once—or run at an acceptable speed. Capacity is a starting constraint, not a performance result.

Treat the headline fit as a claim to test

ASUS says the ProArt GR1X is available with 64GB or 128GB of unified memory and advertises support for 120-billion-parameter models with up to one million tokens of context. S2 Those are manufacturer claims, not an independently measured workload report. The release does not name a model checkpoint, numerical format or runtime, provide throughput, or explain whether the largest model and context figures were demonstrated together.

That gap matters because parameter count alone does not tell a buyer the memory footprint or response time of a particular model build. “It loads” is not the same as “it answers fast enough” once the intended context, tools and applications are running. Treat every maximum as a hypothesis for the exact configuration you plan to buy.

Make the pilot answer one question

Before procurement, write down the real task and freeze its variables: model and checkpoint, quantization, expected context, concurrent agents, open applications and the response time the user will tolerate. Run that workload on the shipping 64GB or 128GB configuration, not a loosely comparable demo system.

Record peak memory use, sustained prompt and generation throughput, wall power, fan noise and recovery after an interrupted run or restart. Ask for measurements from the closed retail chassis under a long session, not just peak compute or a short open-bench demonstration. Verify the required model runtime, drivers, storage and applications on the exact Windows configuration as well.

If the whole stack stays within your latency, thermal and power limits, the desktop may earn its place as a dedicated local host. If it only clears a model-size headline, the missing evidence is the product decision: measure whether the workload fits and remains usable before paying for the ceiling.

Sources and editorial method

This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.

Common questions

Does 128GB of unified memory mean a model gets all 128GB?

No. The operating system, runtime, context and other applications also use the shared pool, so test the full workload on the intended configuration.

What should a buyer measure in a pilot?

Measure peak memory, sustained throughput, wall power, fan noise and recovery behavior while running the intended model, context and applications.