It changes what fits and how much memory juggling your team has to do. On a traditional AI workstation, the GPU has its own memory (up to 96 GB on an RTX PRO 6000 Blackwell), the CPU has separate RAM, and anything that doesn't fit on the card must be copied across a PCIe slot. On an NVIDIA GB300 system such as the HP ZGX Fury G1n or MSI XpertStation WS300, the Grace CPU and Blackwell Ultra GPU share one address space across 252 GB of HBM3e and 496 GB of LPDDR5X, 748 GB in all, joined by a 900 GB/s NVLink-C2C link. The GPU can use CPU memory directly. It is not the same as having 748 GB of fast GPU memory, but it removes a whole class of workarounds from daily work.
What is physically different from an earlier workstation?
| Attribute | Typical x86 workstation with RTX PRO 6000 Blackwell | GB300 workstation (DGX Station class) |
|---|---|---|
| GPU memory | 96 GB GDDR7 per card | 252 GB HBM3e at 7.1 TB/s |
| CPU memory | Separate system RAM | 496 GB LPDDR5X at 396 GB/s, GPU-addressable |
| CPU–GPU link | PCIe Gen5 x16, about 63 GB/s each way | NVLink-C2C, 450 GB/s each way (900 GB/s total) |
| Page tables | Separate for CPU and GPU | One shared per-process page table (ATS) |
| Moving data to the GPU | Explicit copies or managed-memory migration over PCIe | GPU reads CPU memory directly, cache line by cache line; pages can migrate |
| CPU | x86 (Intel or AMD) | 72-core Arm Neoverse V2 (Grace) |
| Display output | From the RTX card | None from the GB300; needs an add-in RTX PRO card |
Source: NVIDIA Grace Hopper architecture blog (coherence, ATS, first-touch placement), NVIDIA Blackwell Ultra blog and DGX Station documentation. PCIe per-direction figure calculated from the PCI-SIG rate. RTX PRO 6000 memory per NVIDIA's product specification.
NVIDIA describes NVLink-C2C as a "hardware-coherent interconnect". The CPU can cache GPU memory at cache-line granularity, both processors can read and write each other's memory, and hardware atomics work across them. Data is placed in HBM or LPDDR5X when it is first touched and can migrate later. The operating system sees the two memory pools as separate NUMA nodes. If you want the full memory-hierarchy explanation, our earlier posts cover it: what coherent actually means, why 748 GB is not all GPU VRAM and HBM3e vs LPDDR5X. This post looks at what changes in day-to-day work.
What changes when a model is bigger than GPU memory?
On a single 96 GB card, a model that doesn't fit forces a choice: quantise it harder, buy more cards and split it, or offload layers to system RAM and pay the PCIe penalty on every pass. On GB300, the first 252 GB runs at HBM speed and the remainder can sit in LPDDR5X, where the GPU reads it over C2C. It is slower, but it runs without a second machine or a re-architected pipeline.
StorageReview's tests on an ASUS ExpertCenter Pro ET900N G3 show what that looks like in practice. GLM-5.2 in NVFP4, with about 215 GB of expert weights offloaded to Grace memory, generated about 35 tokens per second for one user and about 139 tokens per second across 32 concurrent requests. Nemotron 3 Ultra 550B reached about 159 tokens per second at peak. Those are usable speeds for development and internal tools, though well below what the same system does with a model that fits in HBM (9,280 output tokens per second for gpt-oss-120b at 128 concurrent requests). For trillion-parameter sizing, see Can 748 GB handle a trillion-parameter model?
What changes for data science and preprocessing?
A lot of AI work is not model inference. It is loading a large dataframe, cleaning it, building features or embeddings, and handing results to the GPU. On a discrete-GPU workstation, a dataset larger than VRAM has to be chunked and each chunk copied across PCIe. With a shared page table, CPU code and GPU code can work on the same allocation, and the GPU can reach into the 496 GB LPDDR5X pool without a copy step written into the program. The CPU-to-GPU link is also about 7 times faster than PCIe Gen5 x16, so when data does move, it moves sooner.
The practical result is less memory-management code and fewer "out of memory" restarts. It doesn't remove the need to think about placement: data first touched by the CPU lands in LPDDR5X, and GPU kernels reading it there run at about 390 GB/s (StorageReview's measured Grace-to-HBM figure), not 7.1 TB/s.
What changes for agents and multi-model work?
An internal AI assistant rarely runs one model. A typical stack has a main LLM, an embedding model, a reranker, perhaps a vision model, plus a vector index and each user's KV cache. On a 96 GB card, these compete for space, and teams often end up spreading them across machines. With 748 GB addressable, all of them can stay resident, with the busiest kept in HBM and the rest in LPDDR5X. Long-context sessions benefit too, because KV cache that outgrows HBM has somewhere to go other than being dropped. NVIDIA's own marketing talks about "hundreds of agents running simultaneously" on DGX Station; how many your workload supports depends on model size and concurrency.
Coherent memory doesn't make slow memory fast. It makes big memory usable without rewriting your software around it.
What are the limits you'll still run into?
- Two speeds of memory: HBM3e at 7.1 TB/s is about 18 times faster than LPDDR5X at 396 GB/s. Anything latency-sensitive should fit in the 252 GB of HBM.
- Arm software: Grace is a 72-core Arm CPU. Most AI frameworks run natively, but some x86-only engineering and analytics tools don't; see why your software might still struggle.
- No GPU-to-GPU NVLink: coherence covers one CPU and one GPU. Scaling past one system goes over ConnectX-8 networking, not NVLink. See from DGX Station to GB300 NVL72.
- Display and power: the GB300 has no display output, so you need an RTX PRO card for monitors, and a card over 300 W lowers the GB300 power cap from 1,600 W to 950 W. See why a DGX Station needs another graphics card.
For how this compares with a multi-card RTX workstation in more depth, read NVIDIA GB300 vs traditional RTX workstations.
Does it matter whether you pick HP or MSI?
Not for the coherent memory itself. Both use the same GB300 superchip, and reviewers found DGX Station units from different OEMs within about 2% of each other on CPU, C2C and inference tests. The differences are around it. HP's ZGX Fury G1n listing states a 3-year parts, labour and onsite warranty, ships with an RTX PRO 2000 or RTX PRO 4000 SFF for display, includes 5U rack rails, and comes with HP Z Runtime and the HP Z Toolkit. MSI's XpertStation WS300 is a tower with ports on top and a custom closed-loop cooler; warranty terms are confirmed at quotation. Our HP ZGX Fury vs MSI DGX Station summary and software experience comparison go through these in detail.
Buying in Singapore
The HP ZGX Fury G1n and MSI XpertStation WS300 are both price on request, as are the ASUS, GIGABYTE, Supermicro and Dell GB300 systems in our NVIDIA DGX collection. Our model guides cover each in more detail: HP ZGX Fury G1n and MSI XpertStation WS300. Tell us the largest model and dataset you need to hold in memory and we can help you check the fit. 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
Is 748 GB of coherent memory the same as 748 GB of VRAM? No. 252 GB is HBM3e at 7.1 TB/s and 496 GB is LPDDR5X at 396 GB/s. The GPU can address all of it, but only the HBM portion runs at full GPU memory speed.
Do I need to rewrite my code to use coherent memory? Usually not. CUDA programs and frameworks run as before, and system-allocated memory becomes GPU-accessible. Performance tuning still benefits from controlling where large buffers are first allocated.
How much faster is NVLink-C2C than PCIe? NVIDIA rates NVLink-C2C at 900 GB/s total, about 7 times a PCIe Gen5 x16 slot's 128 GB/s bidirectional figure.
Can a GB300 workstation run models that don't fit in 252 GB? Yes, by placing part of the model in LPDDR5X. Reviewers measured usable but much lower speeds when weights are offloaded, so the largest models suit development and lower-concurrency use.
Do HP and MSI GB300 workstations handle memory differently? No. The memory system is part of NVIDIA's GB300 superchip. OEMs differ in chassis, cooling, storage, bundled software and warranty.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
