The RTX PRO 6000 Blackwell uses PCIe 5.0 because, for a single card, almost all the heavy traffic stays inside its own 96 GB of memory, and the slot only has to load data in and results out. NVIDIA GB300 systems use NVLink because they are built for the opposite case: memory and GPUs that must act together. In a DGX Station, NVLink-C2C joins the GB300's GPU to the Grace CPU's memory at 900 GB/s. In a GB300 NVL72 rack, NVLink switches join 72 GPUs at 1.8 TB/s each. PCIe 5.0 is fine for what an RTX PRO 6000 normally does, and becomes the limit when a job spreads across several GPUs or spills out of GPU memory.
How fast is each link?
| Link | Where it is used | Per direction | Bidirectional total |
|---|---|---|---|
| GDDR7 on the RTX PRO 6000 (local memory) | Inside the card | 1,792 GB/s (Workstation and Max-Q); 1,597 GB/s (Server Edition) | |
| PCIe 5.0 x16 | RTX PRO 6000 to CPU and to other cards | About 63 GB/s | 128 GB/s raw |
| Measured RTX PRO 6000 card to card (P2P) | Two cards in one workstation | 45.85 GB/s | 89.93 GB/s |
| PCIe 6.0 x16 | Blackwell Ultra B300 GPU host interface; ConnectX-8 switch | About 121 GB/s | 256 GB/s raw |
| NVLink-C2C | GB300 GPU to Grace CPU memory | 450 GB/s | 900 GB/s |
| NVLink 5 (per GPU) | GPU to GPU via NVLink switches in GB300 NVL72 | 900 GB/s | 1.8 TB/s |
| NVL72 NVLink domain | All 72 GPUs in one rack | — | 130 TB/s aggregate (72 x 1.8 TB/s) |
Source: NVIDIA RTX PRO 6000 Blackwell datasheets; NVIDIA Blackwell Ultra and NVLink-C2C technical material; GB300 NVL72 specifications; card-to-card figures from a Level1Techs p2pBandwidthLatencyTest (July 2025). PCIe per-direction figures are after encoding overhead. NVLink figures are as NVIDIA states them, bidirectional per GPU unless marked.
The table shows how big the gaps are. The RTX PRO 6000's own memory is about 28 times faster than its PCIe 5.0 slot. NVLink-C2C is about 7 times faster per direction than that slot. NVLink 5 is about 14 times faster, which is where NVIDIA's "14x PCIe Gen5" comparison comes from (1.8 TB/s against 128 GB/s, both bidirectional).
Why is PCIe 5.0 enough for one RTX PRO 6000?
A single-card job crosses the slot mostly at the start and end. Loading a 65 GB model such as gpt-oss-120b takes about a second at PCIe 5.0's theoretical rate; in practice, your SSD is usually the slower part. After that, every token, ray or simulation step reads from the card's 1,792 GB/s GDDR7. Rendering, CAD, single-model inference and fine-tuning that fits in 96 GB all behave this way, which is why the card's 125 TFLOPS FP32 and 4,000 AI TOPS (FP4, with sparsity) are usable over an ordinary slot.
PCIe also has practical advantages. It works in any modern workstation or server, it lets you add cards one at a time, and it is the same standard used by NVMe drives and network cards. The RTX PRO 6000 Server Edition goes into ordinary x86 servers for exactly that reason; the HPE ProLiant DL385 Gen11 (price on request) lists it as a supported GPU.
When does PCIe become the bottleneck?
- When a model is split across cards. Tensor parallelism exchanges partial results in every layer. Over PCIe that traffic moves at tens of GB/s, not hundreds. CloudRift found eight NVLink-connected H100 or H200 GPUs delivered 3 to 4 times the throughput of eight "PCIe-bound" RTX PRO 6000 cards on the same model with eight-way tensor parallelism.
- When data spills out of GPU memory. Out-of-core rendering in Redshift or Octane, or offloading model layers to system RAM, pulls data across the slot repeatedly. Maxon warns that many such transfers hurt performance.
- When the CPU and GPU share work closely. Without a coherent link, data must be copied across PCIe before the other side can use it.
For two to four cards, careful software choices keep PCIe workable; our posts on multi-GPU AI without NVLink and two RTX PRO 6000 cards vs a DGX Station show how.
PCIe is a doorway into the GPU. NVLink turns several GPUs, or a GPU and a CPU, into one room.
What does NVLink-C2C do in a GB300?
In a DGX Station, the GB300 has one Blackwell Ultra GPU and one Grace CPU, joined by NVLink-C2C at 450 GB/s each way. The link is coherent, so the GPU can read and write the CPU's 496 GB of LPDDR5X directly, alongside its own 252 GB of HBM3e, for 748 GB in total. A model larger than the HBM can therefore still run without being split across GPUs. It runs more slowly, because LPDDR5X itself delivers 396 GB/s against HBM3e's 7.1 TB/s. The Station's GPU has no NVLink to any other GPU. See DGX Station coherent memory explained for the details.
Why does GB300 NVL72 need a full NVLink fabric?
A GB300 NVL72 rack puts 72 Blackwell Ultra GPUs and 36 Grace CPUs into one NVLink domain. Each GPU has 18 NVLink links, one to each of 18 NVSwitch chips in nine switch trays, so any GPU can reach any other at the full 1.8 TB/s. That lets models far larger than one GPU's memory be split with tensor and expert parallelism, which exchange data constantly, without leaving the rack's NVLink domain. The 130 TB/s figure is the sum of all 72 GPUs' bandwidth, not what one transfer sees. Our explainer on why GB300 NVL72 needs NVLink covers the rack design.
Is PCIe catching up?
PCIe 6.0 doubles the rate to about 121 GB/s per direction for x16, and NVIDIA already uses it: the B300 GPU has a Gen6 x16 host interface, and the ConnectX-8 SuperNICs in NVIDIA's RTX PRO Servers include a PCIe Gen6 switch. The RTX PRO 6000 itself is a PCIe 5.0 card. Even at Gen6, NVLink 5 stays about 7 times faster per GPU (1.8 TB/s against 256 GB/s, both bidirectional). Our PCIe 6.0 vs NVLink post looks at this in depth.
What should RTX PRO 6000 buyers take from this?
If your work fits on one card, PCIe 5.0 costs you almost nothing. If it needs two to four cards, plan the software split and expect some efficiency loss; four RTX PRO 6000 cards is the practical ceiling for most workstations. If one model needs hundreds of gigabytes of fast memory, that is the point where NVLink platforms, from a DGX Station to an NVL72 rack, start to pay for themselves. See from RTX PRO 6000 to GB300 NVL72 for that path.
Buying in Singapore
SourceIT lists the RTX PRO 6000 Blackwell Workstation Edition here and Max-Q here (both S$29,100, listed 6 Nov 2026), and the passive Server Edition at S$32,999 here. Prices change, so ask for a current quotation. DGX Station GB300 systems such as the HP ZGX Fury G1n are price on request, as are GPU servers in our AI workstations and AI servers collection. We don't list NVL72 racks, but we can help you source one. 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 RTX PRO 6000 Blackwell support NVLink? No. NVIDIA's specifications list no NVLink for it, and Lenovo's spec sheets state it is not supported. It connects over PCIe 5.0 x16.
How much faster is NVLink than PCIe 5.0? NVLink 5 gives each GPU 1.8 TB/s bidirectional, about 14 times a PCIe 5.0 x16 slot's 128 GB/s. NVLink-C2C gives 900 GB/s, about 7 times the slot per direction.
Does a DGX Station have GPU-to-GPU NVLink? No. It has one GPU, joined to its Grace CPU by NVLink-C2C. GPU-to-GPU NVLink is in rack systems such as GB300 NVL72.
Will PCIe 5.0 slow down a single RTX PRO 6000? Rarely. Once data is on the card, work runs from its 1,792 GB/s memory. The slot matters when data spills out of the card or a job spans several cards.
Is PCIe 6.0 enough to replace NVLink? Not yet. PCIe 6.0 x16 offers 256 GB/s bidirectional, while NVLink 5 offers 1.8 TB/s per GPU, about 7 times more.
