NVIDIA DGX Station Has 748GB of Memory. But It's Not All GPU VRAM.

Only about a third of a DGX Station's 748GB is GPU memory. NVIDIA's DGX Station GB300 pairs 252GB of HBM3e attached to the Blackwell Ultra GPU (7.1 TB/s) with 496GB of LPDDR5X attached to the 72-core Grace CPU (396 GB/s). NVLink-C2C joins the two pools at up to 900 GB/s, so the GPU can reach all 748GB. The GPU's own memory is about 18 times faster than the CPU's, though. A model that fits inside the 252GB runs at full speed. A model that spills into the LPDDR5X still runs, but whatever part lives there is limited by a far narrower pipe. The five DGX Station GB300 systems on SourceIT's store (HP, MSI, ASUS, GIGABYTE and Supermicro) share this memory layout, and all are price on request.

How is the 748GB actually split?

Figures from NVIDIA's DGX Station product page and datasheet (Aug 2026) and the five OEM listings on SourceIT (4 Nov 2026). The four-module SOCAMM layout is from SourceIT's HP and Supermicro listings and StorageReview's HP coverage (9 Sep 2026). We have not tested a DGX Station ourselves.

Why is the GPU share only 252GB?

Because that is the part bonded to the GPU package. HBM3e is stacked high-bandwidth memory placed right next to the Blackwell Ultra die. It is expensive and limited in capacity, but it is what keeps the GPU's tensor cores fed. The LPDDR5X belongs to the Grace CPU, just as DDR5 belongs to the CPU in a normal workstation. The difference is that NVLink-C2C lets the GPU read it directly instead of copying it across PCIe.

You may also see 784GB in older articles. NVIDIA's March 2025 GTC announcement quoted "784GB of coherent memory", which was 288GB of HBM3e plus 496GB. By GTC 2026 the GPU figure had been revised to 252GB at 7.1 TB/s, according to ServeTheHome (20 Mar 2026). ServeTheHome suggests seven of the GPU's eight HBM3e stacks are enabled, but NVIDIA has not confirmed that explanation. NVIDIA's current product page and datasheet use 748GB, and every OEM listing on our store does too. There is also a small gap on the CPU side: four 128GB SOCAMM modules add up to 512GB, but 496GB is exposed. ServeTheHome describes NVIDIA as reserving the last 16GB.

Does bandwidth matter more than capacity?

For running large language models, often yes. When a model generates text one token at a time, the GPU reads every active weight for every token, so the generation speed of a single stream is roughly capped by memory bandwidth divided by the size of the weights. That gives a quick, theoretical ceiling:

• A 70-billion-parameter model at FP8 is about 70GB of weights. Read from HBM3e at 7.1 TB/s, the ceiling is about 101 tokens per second for one stream.
• The same 70GB read from LPDDR5X at 396 GB/s would cap at under 6 tokens per second.

These are upper bounds that ignore the KV cache, batching and software overheads, so real numbers are lower, but the ratio is the point. HBM3e is about 18 times faster than the LPDDR5X (7,100 ÷ 396). Even NVLink-C2C at 900 GB/s does not change the picture: StorageReview measured roughly 390 GB/s from Grace memory into the GPU on the MSI XpertStation WS300 (24 Aug 2026), because the LPDDR5X itself is the bottleneck.

Capacity decides what you can load. Bandwidth decides how fast it runs. On a DGX Station, 252GB of the 748GB is fast.

Which models fit in the fast 252GB?

Weight size is parameters multiplied by bytes per parameter: 2 bytes at FP16, 1 at FP8 and 0.5 at FP4.

• A 70B model needs about 140GB at FP16, 70GB at FP8 and 35GB at FP4. It fits in HBM3e at any of these precisions, with room for a large KV cache and many concurrent users.
• Llama 3.1 405B is 810GB at FP16 and 405GB at FP8 according to Meta's model card. Neither fits in HBM3e. At FP4 it is about 203GB, which fits in the 252GB with roughly 50GB to spare for cache.
• DeepSeek-R1 has 671 billion parameters, but only 37 billion are active per token because it is a mixture-of-experts model. Its published FP8 weights are about 689GB. That only fits by using both pools, which leaves under 60GB of the 748GB for the operating system, cache and everything else. Because only the active experts are read per token, the placement of frequently used experts matters a great deal.

So the honest summary is that 748GB lets a trillion-parameter-class model load on one machine, which is NVIDIA's "up to 1 trillion parameters" claim. Speed depends on how much of the working set stays in HBM3e. For the full workload ladder, see What Can You Run on a DGX Station GB300? and DGX Station GB300 vs DGX Spark GB10.

What happens when a model spills into CPU memory?

It still works, which is the real advantage over a PCIe workstation. Because the memory is coherent, the GPU can read data that lives in the Grace CPU's LPDDR5X without the application copying it over first. Inference engines can keep cold weights, older KV cache entries or long-context data in the CPU pool and pull them in as needed. The cost is speed: anything the GPU reads from LPDDR5X on every token runs at LPDDR5X pace. We explain what "coherent" means in practice in 748GB Coherent Memory Sounds Incredible. But What Does Coherent Actually Mean?

There is a second trade-off when you share the machine. Multi-Instance GPU (MIG) partitions are carved out of the HBM3e only. NVIDIA's MIG documentation lists profiles from a one-seventh slice of about 31GB up to a full 251GB instance, although profile names vary by driver release. Seven teams each get a slice of the fast memory, not of the 748GB headline. We cover sharing in Can Seven AI Teams Share One DGX Station Without Fighting for GPU Resources?

What should buyers take from this?

• Size your main model to 252GB. If your production model and its cache fit in HBM3e, you get the speed the 20 PFLOPS FP4 rating implies.
• Treat the 496GB as a large, slower tier. It is excellent for data preparation, overflow and very large mixture-of-experts models, and far larger than any PCIe GPU's memory.
• Quantisation matters. Moving a 405B model from FP8 to FP4 is the difference between spilling into CPU memory and fitting entirely in HBM3e.
• Every OEM has the same memory. The five GB300 systems differ in chassis, cooling, storage, software and warranty, not in the 252GB plus 496GB layout. See Two DGX Stations, Same 748GB Memory. Why Might Their Performance Still Differ?

Buying in Singapore

SourceIT supplies the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro Super AI Station ARS-511GD-NB-LCC, all with 252GB of HBM3e and 496GB of LPDDR5X. These systems are price on request: public US pricing puts most configurations at roughly US$93,000 to US$125,000, or about S$119,000 to S$160,000 before GST at 3 Nov 2026 exchange rates, depending on OEM, storage and options. If your models fit in 128GB, the NVIDIA DGX Spark GB10 128GB/4TB is S$9,300 (listed 4 Nov 2026). Browse the NVIDIA DGX Spark and DGX Station collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

How much GPU memory does the DGX Station GB300 have? 252GB of HBM3e at 7.1 TB/s. The other 496GB of the 748GB total is LPDDR5X attached to the Grace CPU at 396 GB/s.

Can the GPU use all 748GB? Yes. NVLink-C2C makes the two pools coherent, so the GPU can read CPU memory directly. Data read from the LPDDR5X is limited to that memory's much lower bandwidth.

Why do some articles say 784GB? NVIDIA's 2025 announcement quoted 288GB of HBM3e plus 496GB. The shipping specification is 252GB plus 496GB, which is 748GB, and that is what NVIDIA and all five OEMs now publish.

Does Llama 3.1 405B fit in the GPU memory? Only at FP4, where its weights are about 203GB. At FP8 (405GB) or FP16 (810GB) part of it must sit in CPU memory.

Do the HP, MSI, ASUS, GIGABYTE and Supermicro versions have different memory? No. All five use the same NVIDIA GB300 board with 252GB of HBM3e and 496GB of LPDDR5X. They differ in chassis, cooling, storage, software and warranty.