"Coherent" means the GPU and CPU in a DGX Station see one shared memory space and agree on its contents. Neither has to copy data to the other before using it. In the NVIDIA DGX Station GB300, NVLink-C2C links the Grace CPU's 496GB of LPDDR5X and the Blackwell Ultra GPU's 252GB of HBM3e at up to 900 GB/s, with hardware cache coherence. That is why NVIDIA can quote 748GB as one figure. Coherent does not mean uniform, though. The two pools still have very different speeds, and the software still decides where data lives. This post explains what coherence buys you and what it does not.
What does "coherent memory" mean in plain terms?
In a conventional workstation, a GPU sits on a PCIe slot with its own memory. The CPU has separate system memory. To use data, the application allocates GPU memory, copies the data across PCIe, runs the work and copies results back. The two memories are separate worlds.
On Grace-based superchips, NVIDIA connects CPU and GPU with NVLink-C2C, a chip-to-chip link that "provides memory coherency to simplify programming", in the words of NVIDIA's Grace Performance Tuning Guide. The same guide describes the two memories as two NUMA nodes joined by NVLink-C2C, "making access to both CPU and GPU memory seamless", and says the platforms support "all features of CUDA Unified Memory". In practice that means three things:
- One address space. A pointer to data in CPU memory can be used by GPU code, and the other way round.
- Hardware keeps caches consistent. If the CPU writes a value, the GPU reads the new value, without the program flushing caches or copying buffers.
- Fine-grained access. The GPU can read just the cache lines it needs from CPU memory, instead of moving whole buffers first.
How does that differ from a PCIe GPU workstation?
| DGX Station GB300 (coherent) | Typical PCIe GPU workstation | |
|---|---|---|
| GPU memory | 252GB HBM3e, 7.1 TB/s | The card's own memory, e.g. 96GB on an RTX PRO 6000 Blackwell |
| CPU memory the GPU can use | 496GB LPDDR5X, read directly | System RAM, normally copied across PCIe first |
| CPU–GPU link | NVLink-C2C, up to 900 GB/s | PCIe 5.0 x16, about 64 GB/s per direction |
| Address space | Shared and cache-coherent | Separate; unified memory is emulated by page migration |
| Headline memory | 748GB "coherent memory" | Card memory and system memory quoted separately |
DGX Station figures from NVIDIA's product page and datasheet (Aug 2026) and the Grace Performance Tuning Guide (accessed 4 Nov 2026). PCIe 5.0 x16 bandwidth is the nominal specification for the slot. We have not benchmarked these systems ourselves.
Does coherent mean the GPU sees 748GB at full speed?
No, and this is the most common misunderstanding. Coherence is about correctness and convenience, not bandwidth. Data the GPU reads from the LPDDR5X is limited by that memory and the link, not by HBM3e speed.
- HBM3e runs at 7.1 TB/s. The Grace LPDDR5X runs at 396 GB/s.
- NVLink-C2C is rated at up to 900 GB/s, but StorageReview measured roughly 390 GB/s from Grace memory into the GPU on the MSI XpertStation WS300 (24 Aug 2026). The LPDDR5X itself is the ceiling.
- An academic study of the earlier Grace Hopper GH200 (Schieffer et al., arXiv, July 2024) measured 375 GB/s host-to-device and 297 GB/s device-to-host over NVLink-C2C. It found that data left in CPU memory and "accessed over NVLink-C2C at a low bandwidth" could throttle a workload.
The same study shows the upside too. With system-allocated memory (ordinary malloc), the GPU could start working on CPU-resident data immediately, and frequently used pages migrated into GPU memory over time. After the working set had moved, reads over NVLink-C2C dropped to almost zero. That was on GH200, an earlier generation, but GB300 uses the same Grace CPU and NVLink-C2C design.
Coherent memory removes the copy step and the out-of-memory wall. It does not turn slow memory into fast memory.
What does coherence change for AI teams in practice?
- Bigger models load without crashing. A model larger than 252GB can still run, with part of it in CPU memory, instead of failing with an out-of-memory error. That is the basis of NVIDIA's "up to 1 trillion parameter" claim.
- Simpler code for data pipelines. Preprocessing on the Grace cores and training or inference on the GPU can share buffers without explicit transfers.
- Offloading becomes cheaper. Inference engines that offload KV cache or inactive weights to CPU memory get a much faster path than PCIe offers.
- Placement still matters. For best speed, keep the hot data (active weights and the live KV cache) in the 252GB of HBM3e. Our companion post, NVIDIA DGX Station Has 748GB of Memory. But It's Not All GPU VRAM., works through which models fit.
Is the memory still coherent when you link two DGX Stations?
NVIDIA's datasheet says two units linked over ConnectX-8 give "over 1.4 TB of coherent memory and up to 40 petaFLOPS". That link is networking, not NVLink. NVIDIA's own two-node setup guide uses two 400G cables between the ConnectX-8 ports with RoCE (RDMA over Ethernet), and there is no NVLink between separate DGX Stations. In practice, software such as an inference engine with tensor or pipeline parallelism splits the model across two machines and exchanges data over the network. That is a different and slower mechanism than the on-chip coherence inside one Station, so plan for it as a cluster rather than one larger box. We cover the networking side in DGX Station's Secret Weapon Isn't Just the GPU. It's the Networking.
Do developers need to change anything?
Mostly no, which is the point. CUDA code written for unified memory runs, and NGC containers for PyTorch and other frameworks are published for Arm64. Two practical notes from NVIDIA's DGX Station porting guidance are worth knowing. The operating system uses a 64KB memory page size, and NVIDIA says applications that require 4KB pages "are not supported at this stage". Performance tuning also benefits from treating CPU and GPU memory as separate NUMA nodes, as NVIDIA's tuning guide describes. For broader Arm software issues, see Your DGX Station Has 72 CPU Cores. Why Might Your Software Still Struggle?
Buying in Singapore
Every DGX Station GB300 we supply uses the same coherent 252GB plus 496GB design: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. If a PCIe workstation fits your workload better, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition with 96GB is S$29,100 (listed 4 Nov 2026). For that comparison on GB10, read NVIDIA GB10 vs RTX PRO 6000 Blackwell. 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
What does coherent memory mean on the DGX Station? The Grace CPU and Blackwell Ultra GPU share one address space over NVLink-C2C, and hardware keeps their caches consistent, so either processor can use data in the other's memory without the program copying it first.
Is all 748GB as fast as GPU memory? No. 252GB is HBM3e at 7.1 TB/s. The 496GB of LPDDR5X runs at 396 GB/s, and StorageReview measured about 390 GB/s from that memory into the GPU.
How fast is NVLink-C2C compared with PCIe? NVIDIA rates NVLink-C2C at up to 900 GB/s. A PCIe 5.0 x16 slot is nominally about 64 GB/s per direction.
Are two linked DGX Stations one coherent 1.4TB machine? NVIDIA describes two linked units as offering over 1.4TB of coherent memory, but the link is ConnectX-8 networking with RDMA, not NVLink. Software has to split work across the two systems.
Do I need to rewrite my code for coherent memory? Usually not. CUDA unified memory code runs as normal. Check for 4KB page-size assumptions, which NVIDIA says are not supported at this stage on DGX Station.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
