Why Unified Memory Fits Local MoE Models

Many of the strongest open models you can download today are mixture-of-experts designs. They keep a large set of specialist networks on disk and in memory, then use only a few of them for each word they generate. Compute per token stays closer to a mid-size dense model. Memory still has to cover every expert.

Unified-memory deskside systems put CPU and GPU on one shared pool. For MoE work, that pool is often why people choose platforms such as NVIDIA DGX Spark or the CORSAIR AI Workstation 300. The full expert set can sit in one place instead of stopping at a mid-size GPU card.

Those models load many specialist networks and use only a few for each word they generate. Model pages call that a mixture of experts. NVIDIA Nemotron 3.5 Bolt has 74 billion parameters in total. About 9 billion of them run for each word.

Only some of the model runs for each word, but the whole model still has to be in memory. Unified memory, where the processor and the graphics use the same memory, is one way to hold it. The conversation and the app still need room on top. For a separate graphics card, see How much GPU memory you need for local AI.

On a classic discrete-GPU workstation, GPU memory and system RAM are separate. A large MoE file that exceeds the card has to spill, quantize harder, or move to a bigger GPU.

On a unified-memory system, the CPU and the accelerator share one coherent pool. The model file, the conversation, and the rest of the OS draw from the same capacity. That design shows up in deskside AI machines because it matches how MoE weights behave: capacity is driven by the full expert set, while speed still depends on how much of that set is active per token and how fast the memory can feed the GPU.

NVIDIA DGX Spark (SKU CSW-9020003-NA) is built around the NVIDIA GB10 Grace Blackwell Superchip. It has 128 GB of LPDDR5x coherent unified system memory, shared by the CPU and GPU, with about 273 GB/s of memory bandwidth.

NVIDIA positions the platform for local agent work and large models at the desktop, including AI models up to 200 billion parameters on a single Spark. For MoE work, the useful property is the large shared pool: the inactive experts and the active path sit in the same coherent memory instead of stopping at a smaller dedicated GPU budget.

The CORSAIR AI Workstation 300 (SKU CS-9080002-NA, with a higher-storage CS-9080003-NA configuration also listed) pairs an AMD Ryzen AI Max+ 395 with 128 GB of LPDDR5X-8000 and Radeon 8060S graphics at up to 96 GB of VRAM.

That “up to 96 GB” figure comes from AMD’s Variable Graphics Memory feature on Ryzen AI Max platforms: the same physical LPDDR5X can be reallocated so a large contiguous block shows up as dedicated graphics memory for workloads that expect VRAM. AMD’s own FAQ walks through a 128 GB Max+ 395 system with VGM set to 96 GB, and uses Meta’s Llama 4 Scout-class 109B total / 17B active model as an example of the large-parameter path that kind of pool is meant for.

A mixture-of-experts model keeps many specialist networks in memory and uses only a few for each word. Unified memory means the processor and the graphics share that memory.

NVIDIA DGX Spark is NVIDIA's system. The processor and the graphics share 128 GB. CORSAIR offers it. CORSAIR makes the AI Workstation 300. The processor and the graphics use the same memory. Variable Graphics Memory sets aside part of it for graphics, up to 96 GB of the 128 GB.

On a separate graphics card, the model has to fit in the memory on that card. For those sizes, see How much GPU memory you need for local AI.

CORSAIR PRO builds configure-to-order AI workstations, from compact deskside systems to multi-GPU towers. Match the machine to how your software places MoE weights: one shared pool on platforms such as DGX Spark or AI Workstation 300, or a separate GPU pool on a discrete-GPU workstation.

Explore options in the CORSAIR PRO configurator, or talk through requirements with a CORSAIR PRO solutions architect.