Two RTX PRO 6000 GPUs vs One DGX Station: Which Makes More Sense for Local AI?

For most local AI teams, one DGX Station makes more sense than two RTX PRO 6000 cards when the goal is running one large model fast. It makes less sense when the goal is running several mid-sized models, mixing AI with graphics, or keeping spending under control. Two RTX PRO 6000 Blackwell cards give you 192 GB of GPU memory, but as two separate 96 GB pools joined by PCIe, so any model bigger than one card must be split in software. A DGX Station GB300 gives one GPU with 252 GB of HBM3e and 748 GB of coherent memory in total, so many of those models simply load. The two cards cost far less, and they run ordinary x86 software.

How do two cards and one GB300 compare on paper?

Source: NVIDIA RTX PRO 6000 Max-Q and DGX Station specifications; card-to-card figure from a Level1Techs p2pBandwidthLatencyTest on two RTX PRO 6000 cards (July 2025). The GB300 ballpark assumes public US pricing of roughly US$93,000 to US$125,000 for most OEMs, about S$119,000 to S$160,000 before GST at USD/SGD 1.279 (6 Nov 2026). Prices change, so ask for a current quotation.

We used the 300 W Max-Q edition for the two-card build on purpose. Two 600 W Workstation Editions in one tower need a chassis designed for it: Puget Systems says the Workstation Edition is "NOT suitable for most multi-GPU configurations", and HP's Z8 Fury G5 lists one Workstation Edition or up to four Max-Q cards. Dell's Pro Precision 9 T6 is one of the few towers that lists two 600 W GPUs.

Which models fit, and where?

Capacity decides whether a model loads at all. These are weight sizes only; you need another 10 to 20% for the KV cache and the runtime.

• gpt-oss-120b (65 GB in its native MXFP4 format) fits on one RTX PRO 6000. The second card can run a second copy or a different model.
• Llama 3.3 70B in BF16 (141 GB) needs both cards, which leaves roughly 50 GB across the pair for cache and overhead. At FP8 (71 GB) it fits on one card.
• Qwen3-235B-A22B at 4-bit (about 118 GB) needs both cards. At FP8 (236 GB) it needs four cards, or the GB300's HBM3e, which is a tight fit.
• DeepSeek-V3 or R1 at FP8 (689 GB) doesn't fit two cards at all. On a GB300 it loads, but about 437 GB of it sits in the slower LPDDR5X.

Our 96GB VRAM sizing guide covers more models and the KV cache arithmetic.

What does splitting a model across two cards actually involve?

When a model is bigger than 96 GB, the software has to split it. There are two common ways. Pipeline parallelism puts the first half of the layers on card one and the second half on card two, and passes activations across PCIe once per step. Tensor parallelism splits every layer across both cards and combines the partial results in every layer, which means far more traffic over PCIe.

Which is better on PCIe is genuinely contested. vLLM's documentation says that if GPUs on a node don't have NVLink, you should "leverage pipeline parallelism". Yet CloudRift and Helix both found tensor parallelism faster on RTX PRO 6000 cards in their own tests. Expert parallelism for mixture-of-experts models was described by one vLLM forum report as "catastrophic on PCIe", at 1.4 to 2.6 tokens per second on four cards. Benchmarks in this area age quickly; Helix measured a 7.8x difference between two inference engines on identical cards.

Card-to-card transfers also need tuning. Peer-to-peer (P2P) copies work on RTX PRO cards, unlike GeForce, but users report needing IOMMU and ACS changes in the BIOS or kernel, and hangs have been reported on some Threadripper PRO boards. With P2P off, one Level1Techs user measured NCCL all-reduce at about 31 GB/s between two Max-Q cards. Our post on multi-GPU AI without NVLink goes deeper.

Two RTX PRO 6000 cards give you 192 GB. They don't give you a 192 GB GPU.

How does memory access differ on the GB300?

On the GB300 there is nothing to split up to 252 GB. The model loads into one GPU's HBM3e and runs at 7.1 TB/s. Past 252 GB, NVLink-C2C lets the GPU reach the Grace CPU's 496 GB of LPDDR5X directly, about 10 times faster than the measured card-to-card PCIe figure (450 GB/s each way against about 46 GB/s). Overflow is still slow memory, though. One StorageReview test ran GLM-5.2 with about 215 GB of experts in Grace memory at about 35 tokens per second for a single user. See 748GB is not all GPU VRAM and HBM3e vs LPDDR5X for that trade-off.

Bandwidth sets the speed limit for generating tokens. For a 70B model at FP8 (71 GB read per token), the theoretical single-user ceiling is about 25 tokens per second on one RTX PRO 6000 and about 100 on GB300 HBM3e. Splitting across two cards with tensor parallelism can raise the RTX ceiling, because each card reads half the weights, but PCIe traffic eats into the gain. Our memory vs bandwidth explainer walks through the arithmetic.

What do the measurements say?

No published test puts exactly two RTX PRO 6000 cards against a DGX Station on the same model, so treat any claim to the contrary carefully. The closest like-for-like figure is StorageReview's gpt-oss-120b test at 128 concurrent requests: about 3,064 output tokens per second on one RTX PRO 6000 and 9,280 on an ASUS ExpertCenter Pro ET900N G3 (GB300), roughly three times more. Separate copies of a model scale well in practice: Helix measured 2.03x throughput when it ran two copies side by side instead of one, and Exxact measured 96% scaling for two-GPU LoRA fine-tuning. Even an ideal doubling (about 6,100 tokens per second) would stay below the GB300 figure on that model.

Which should you choose?

• Choose two RTX PRO 6000 cards if your models are mostly 120B-class or smaller, you want two models or two users running side by side, you need x86 software or graphics, or your budget is closer to S$60,000 than S$150,000. MIG lets each card split into up to four 24 GB instances.
• Choose a DGX Station if your main model is over 96 GB, you want the highest single-model throughput, you'd rather not tune parallelism settings, or you plan to move work to GB300 systems in a data centre later. Check your software against the Arm CPU first; see Grace Arm software compatibility.

For the single-card version of this decision, read RTX PRO 6000 vs DGX Station GB300. If two cards isn't enough, four RTX PRO 6000 cards is the next step to consider.

Buying in Singapore

SourceIT lists the RTX PRO 6000 Blackwell Max-Q Workstation Edition at S$29,100 (listed 6 Nov 2026) here and the 600 W Workstation Edition at S$29,100 here, with more in our NVIDIA graphics cards collection. We can quote a host workstation designed for two cards. DGX Station GB300 systems, including the ASUS ExpertCenter Pro ET900N G3, HP ZGX Fury G1n and MSI XpertStation WS300, are price on request. Tell us your largest model and how many users you have, and we'll price 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

Do two RTX PRO 6000 cards act as one 192 GB GPU? No. Each card has its own 96 GB. Software such as vLLM splits a larger model across both and exchanges data over PCIe.

Is a DGX Station faster than two RTX PRO 6000 cards? On large single models, usually yes. StorageReview measured about three times the gpt-oss-120b output of one RTX PRO 6000 on a GB300, and even perfect two-card scaling would fall short of that.

Which costs less? Two RTX PRO 6000 cards are S$58,200 at SourceIT (listed 6 Nov 2026) before the host workstation. A DGX Station is price on request, with around S$150,000 as a ballpark.

Can I fit two 600 W Workstation Editions in one tower? Only in systems designed for it, such as Dell's Pro Precision 9 T6. Most multi-GPU builds use the 300 W Max-Q edition.

What model is too big for two cards but fine on a GB300? DeepSeek-V3 or R1 at FP8, at about 689 GB, won't fit in 192 GB. It loads on a GB300's 748 GB, although most of it runs from the slower LPDDR5X.