For large AI models, a fast coherent link is the missing ingredient, but on a GB300 workstation it isn't the GPU-to-GPU NVLink people usually picture. The NVIDIA GB300 in a DGX Station has one Blackwell Ultra GPU joined to a Grace CPU by NVLink-C2C at 900 GB/s. That gives the GPU 252 GB of HBM3e plus direct access to 496 GB of CPU memory. A traditional workstation with two or four RTX PRO 6000 Blackwell cards has more GPUs and, with four cards, more total GPU memory, but each card holds 96 GB and talks to the others and to the CPU over PCIe. For one big model, GB300 usually wins. For many smaller jobs, graphics and x86 software, the RTX workstation still has real advantages.
How do the two designs compare on paper?
| Attribute | GB300 workstation (DGX Station class) | 2x RTX PRO 6000 Blackwell WE | 4x RTX PRO 6000 Blackwell WE |
|---|---|---|---|
| GPU memory | 252 GB HBM3e (one GPU), plus 496 GB LPDDR5X addressable | 192 GB GDDR7 (2 x 96 GB) | 384 GB GDDR7 (4 x 96 GB) |
| Largest single pool | 252 GB fast, 748 GB coherent | 96 GB per card | 96 GB per card |
| CPU–GPU link | NVLink-C2C, 450 GB/s each way | PCIe Gen5 x16 per card, about 63 GB/s each way | Same, per card |
| GPU–GPU link | Not applicable (one GPU) | PCIe | PCIe |
| Power | 1,600 W total system budget | 1,200 W for the two cards alone | 2,400 W for the four cards alone |
| SourceIT price (listed 6 Nov 2026) | Price on request | S$58,200 for the cards only | S$116,400 for the cards only |
| Measured: gpt-oss-120b, 128 concurrent | 9,280 output tok/s (ASUS ET900N G3) | 3,064 output tok/s on one RTX PRO 6000 (multi-card not tested) | |
Source: NVIDIA DGX Station and Grace Hopper/Blackwell Ultra technical material; RTX PRO 6000 Blackwell Workstation Edition listing (96 GB GDDR7, 600 W). Throughput from StorageReview's ASUS ExpertCenter Pro ET900N G3 review (vLLM, 512 in / 512 out). Card prices exclude the host workstation, CPU, RAM, storage and power supply.
On the one like-for-like measurement available, a single GB300 produced about 3 times the output of a single RTX PRO 6000 on the same model and concurrency. On GPT-OSS-20B, StorageReview's MSI XpertStation WS300 result was about 2.5 times the RTX PRO 6000 figure (22,161 vs about 9,000 tokens per second at 128 concurrent requests). Most of that gap comes from HBM3e bandwidth and Blackwell Ultra's FP4 tensor cores, not from the interconnect.
So where does NVLink come into it?
NVLink matters when one model is bigger than one GPU's memory. Then the model must be split, and the GPUs exchange data constantly. Tensor parallelism, the usual way to split a large model, combines partial results with an all-reduce in every layer. In NVIDIA's data-centre systems, that traffic runs over NVLink 5 at 1.8 TB/s bidirectional per GPU, about 14 times a PCIe Gen5 x16 slot. RTX PRO Blackwell workstation cards don't have that path: they share data over PCIe, often through the CPU's root complex, at roughly 63 GB/s each way per slot.
That has a measurable cost. Even on NVLink-connected DGX H100 systems, Microsoft's DeepSpeed team found tensor-parallel communication took 17% to 47% of iteration time. Over PCIe, at a fraction of NVLink's bandwidth, you should expect that share to grow. That's why multi-card PCIe workstations usually split large models by layer (pipeline parallelism), which passes activations between cards once per stage rather than every layer, or run a separate model on each card. Our explainers on PCIe 6.0 vs NVLink and why NVIDIA doesn't rely on PCIe alone go further into the link technology.
GB300 sidesteps the problem differently. With 252 GB in one GPU, many models that would need splitting across three 96 GB cards simply fit, so there's no GPU-to-GPU traffic to worry about. Beyond 252 GB, the overflow goes to LPDDR5X over NVLink-C2C, which is slower than HBM but faster than any PCIe slot.
The question isn't whether a workstation has NVLink. It's whether your largest model has to cross a slow link to run.
Where does coherent CPU–GPU memory help large models?
- Fewer splits: a 120B-class model at 8-bit (about 120 GB of weights) fits in 252 GB of HBM3e with room for KV cache. On 96 GB cards it must be quantised further or split.
- Graceful overflow: models larger than HBM can run with part of their weights in LPDDR5X. StorageReview ran GLM-5.2 with about 215 GB of experts in Grace memory at about 35 tokens per second for a single user.
- Simpler software: one GPU, one memory space and one shared page table means fewer parallelism settings to tune.
- Faster staging: data prepared on the CPU reaches the GPU over 450 GB/s each way instead of about 63 GB/s.
Our companion post, Inside NVIDIA GB300: Why 748GB of Coherent Memory Changes the AI Workstation Game, looks at what this means for daily workflows.
Where does a traditional RTX workstation still win?
- Graphics and visualisation: RTX PRO cards drive displays and run OpenGL, Vulkan, rendering and simulation. The GB300 has no display output at all.
- x86 software: CAD, CAE and analytics tools that only ship for x86 Windows or Linux run as usual. Grace is an Arm CPU.
- More GPUs for parallel jobs: four cards can run four independent models or users at once, with fault isolation between them.
- More total GPU memory: four cards give 384 GB, more than GB300's 252 GB of HBM3e, when the work splits cleanly.
- Incremental upgrades: start with one or two cards and add more later. A GB300 is one superchip.
The limits are just as real. Every model shard is capped at 96 GB, cross-card traffic runs over PCIe, and a four-card build at 600 W each needs a chassis, power supply and cooling designed for it. That's one reason the 300 W Max-Q edition exists. Our GB10 vs RTX PRO 6000 Blackwell comparison covers the single-card side of this trade-off, and Could DGX Station replace a room full of engineering workstations? covers the design-software side.
Can you have both in one box?
Partly. A DGX Station can take an RTX PRO card for display, graphics and rendering alongside the GB300. The trade-off is power: the system has a fixed 1,600 W budget, and fitting a card over 300 W, such as an RTX PRO 6000 Workstation Edition, lowers the GB300's power cap to 950 W. A card of 76 to 300 W lowers it to 1,250 W. See why a DGX Station needs another graphics card before choosing one. If you're weighing a GB300 against an existing multi-GPU server instead, read Can NVIDIA GB300 replace your company's multi-GPU AI server?
Buying in Singapore
For an RTX workstation build, SourceIT lists the RTX PRO 6000 Blackwell Workstation Edition (S$29,100, listed 6 Nov 2026) here, the 300 W RTX PRO 6000 Blackwell Max-Q (S$29,100) here and the RTX PRO 5000 Blackwell (S$14,971.20) here, with more in our NVIDIA graphics cards collection. GB300 workstations, including the HP ZGX Fury G1n, MSI XpertStation WS300 and Dell Pro Max with GB300, are price on request in our NVIDIA DGX collection. Tell us your largest model and your graphics needs and we'll help you compare both routes. 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
Does the GB300 in a DGX Station use NVLink? Yes, but NVLink-C2C between its Grace CPU and Blackwell Ultra GPU, at 900 GB/s total. It has only one GPU, so there is no GPU-to-GPU NVLink inside the box.
How do RTX PRO 6000 cards talk to each other in a workstation? Over PCIe. A Gen5 x16 slot carries about 63 GB/s in each direction, far less than NVLink in data-centre GPU systems.
Is four RTX PRO 6000 cards better than one GB300? It depends on the work. Four cards give 384 GB in total and suit parallel jobs, graphics and x86 software. One GB300 gives 252 GB in a single GPU, which suits large single models better.
Can a large model be split across RTX PRO cards? Yes, usually by layers (pipeline parallelism) to limit traffic over PCIe. Splitting every layer across cards (tensor parallelism) is possible but loses more time to communication.
Can I add an RTX PRO 6000 to a DGX Station? Yes, it's a supported option for display and graphics, but a card over 300 W lowers the GB300 power cap from 1,600 W to 950 W.
