No. PCIe isn't running out of road. Its bandwidth has doubled with each generation: PCIe 6.0 was released on 11 January 2022, PCIe 7.0 on 11 June 2025, and PCIe 8.0 is planned for release to PCI-SIG members by 2028. What PCIe was never designed to be is a fabric that makes many GPUs work as one, and there NVLink stays several times ahead. A Blackwell Ultra GPU has both. Its PCIe Gen6 x16 host interface is rated at 256 GB/s bidirectional. Its NVLink 5 is rated at 1.8 TB/s bidirectional, about seven times more. The approach that is reaching its limit is building large multi-GPU systems by plugging cards into a PCIe bus. PCIe itself is fine.
How do PCIe generations compare with NVLink?
| Interface | Released | Rate per lane | Signalling | x16 per direction (usable) | x16 bidirectional (headline) |
|---|---|---|---|---|---|
| PCIe 5.0 | 2019 | 32 GT/s | NRZ, 128b/130b | about 63 GB/s | 128 GB/s |
| PCIe 6.0 | 11 Jan 2022 | 64 GT/s | PAM4 + FLIT | about 121 GB/s | 256 GB/s |
| PCIe 7.0 | 11 Jun 2025 | 128 GT/s | PAM4 + FLIT | about 242 GB/s | 512 GB/s |
| PCIe 8.0 | Planned for 2028 (announced 5 Aug 2025) | 256 GT/s | Not yet announced | Not confirmed | 1 TB/s |
| NVLink 5 (Blackwell, Blackwell Ultra) | Shipping | — | — | 900 GB/s per GPU | 1.8 TB/s per GPU |
| NVLink 6 (Vera Rubin) | Ramping, per NVIDIA (Oct 2026) | — | — | 1.5 TB/s per GPU | 3 TB/s per GPU |
Source: PCI-SIG releases via BusinessWire (PCIe 6.0, 11 Jan 2022; PCIe 8.0 announcement, 5 Aug 2025), Tom's Hardware (PCIe 7.0, 12 Jun 2025), NVIDIA NVLink page (updated 15 Sep 2026) and Vera Rubin NVL72 page (updated 21 Oct 2026). Usable per-direction PCIe figures are our calculation after 128b/130b or FLIT overhead. Headline PCIe figures are raw and bidirectional, the same basis NVIDIA uses for NVLink. NVIDIA's NVLink 6 figure was revised down from 3.6 TB/s announced at CES in January 2026 (TechRadar, 31 Oct 2026).
What did PCIe 6.0 change?
PCIe 6.0 doubled the rate per lane by changing how bits are sent rather than by simply clocking faster. It moved from two-level NRZ signalling to PAM4, which carries two bits per symbol, and to fixed-size FLIT (flow control unit) packets. PAM4 is more error-prone, so the specification added lightweight forward error correction and a cyclic redundancy check. Dr. Debendra Das Sharma of Intel, a PCI-SIG board member, put the FEC and CRC delay at about 1–2 ns per link. In his words, PCIe "can not afford the 100ns FEC latency as networking does with PAM4". Each 256-byte flit carries 236 bytes of transaction data, and the retry rate he cited costs about 0.05% of bandwidth. PCIe 6.0 is also backward compatible with every earlier generation, which is the reason PCIe is everywhere. PCIe 7.0 keeps PAM4 and FLIT and doubles the rate again.
Which products actually use PCIe 6.0 today?
Adoption arrived in AI systems first:
- NVIDIA Blackwell Ultra (B300) has a PCIe Gen6 x16 host interface. Tom's Hardware calls it the first data-centre GPU to support PCIe 6.x. That is press framing, not a verified first.
- NVIDIA ConnectX-8 SuperNIC integrates a 48-lane PCIe Gen6 switch alongside 800 Gb/s networking. NVIDIA calls it the industry's first SuperNIC with a Gen6-capable switch (a vendor claim). It ships in HGX B300, GB300 NVL72 and RTX PRO servers.
- Micron 9650, a Gen6 x4 NVMe SSD, entered mass production in February 2026 with reads of up to 28 GB/s (StorageReview).
- PCIe Gen6 switches were not yet in general release. Broadcom showed its 144-lane PEX90144 on a pre-release board at SC25. Microchip's Switchtec Gen 6 was sampling as of October 2025.
Host CPUs lag behind. NVIDIA documents Grace as PCIe Gen5, although one press report described it as Gen6-capable, so treat that claim with caution. Inside a DGX Station GB300, two M.2 slots run at Gen6 x4 off the ConnectX-8, while the slots off Grace are Gen5.
Why can't PCIe simply replace NVLink?
Bandwidth is the obvious reason. Per direction, PCIe Gen6 x16 at about 121 GB/s gives a GPU about one-seventh of NVLink 5's 900 GB/s. Even PCIe 7.0 at about 242 GB/s is about a quarter. PCIe 8.0's 1 TB/s bidirectional headline is still below NVLink 5's 1.8 TB/s today, and PCI-SIG coverage expects Gen8 devices around 2030 or later. By then NVLink will have moved on: NVLink 6 on Vera Rubin is already rated at 3 TB/s bidirectional per GPU. Three other differences matter as much as raw speed:
- Topology. PCIe is a host-rooted tree. GPUs behind the same switch share its uplinks, and a 144-lane Gen6 switch has room for only nine x16 devices. An NVLink 5 Switch chip has 72 ports and 7.2 TB/s of bidirectional all-to-all bandwidth. An NVL72 rack gives each of its 72 GPUs one link to each of 18 such chips.
- Coherence. On Grace systems, NVLink-C2C lets the CPU and GPU share one page table and cache each other's memory. On a typical x86 plus PCIe system, data has to be copied explicitly. See Why NVIDIA's Most Powerful AI Systems Don't Rely on PCIe Alone.
- Energy. NVIDIA rates NVLink-C2C at 1.3 picojoules per bit. It claims up to 6x better energy efficiency than a PCIe Gen6 PHY on its own chips (a vendor claim).
PCIe isn't running out of speed. It was simply never designed to be the fabric between GPUs.
Is there an open alternative to NVLink?
Several are in progress, and none is built on PCIe.
- UALink: 1.0 was released on 8 April 2025, at 200G per lane and up to 1,024 accelerators per pod. Its x4 "station" carries 800 Gb/s in each direction, which is 100 GB/s per direction. UALink 2.0 followed on 7 April 2026. The Register reported that 1.0 chips would reach labs in the second half of 2026, with products later in 2027.
- Ultra Ethernet: the UEC 1.1 specification (12 October 2026) added a "Unified Local" profile for short-range scale-up links.
- AMD Helios: AMD's 72-GPU rack quotes 260 TB/s of scale-up bandwidth over "UALink over Ethernet". AMD does not say whether that figure is per direction or bidirectional.
- NVLink Fusion: NVIDIA's own response, announced on 18 May 2025, licenses NVLink to third-party CPUs and accelerators. AWS is designing Trainium4 to integrate with NVLink 6.
The analysts are divided. Dell'Oro (28 July 2026) says NVLink leads scale-up today but expects Ethernet to lead by the end of the decade. We look at the business side in Is NVLink NVIDIA's Biggest Competitive Advantage?
What does this mean for workstation buyers?
In a workstation, the PCIe limit shows up when you add GPUs. RTX PRO Blackwell workstation cards connect to the host over PCIe, and SourceIT's listings for them do not mention NVLink. In a multi-card workstation, the cards therefore exchange data over PCIe Gen5 at about 63 GB/s per direction per x16 slot. That is fine when each card runs its own model or job. It is slower when one model is split layer by layer across cards. A GB300 system takes the opposite approach: one large GPU with 252 GB of HBM3e, joined to the Grace CPU's memory over coherent NVLink-C2C. Our post NVIDIA GB300 vs RTX Workstations compares the two. For the bigger picture, see NVLink vs InfiniBand vs PCIe.
For storage, PCIe 6.0 is where most buyers will meet it first. Gen6 SSDs such as the Micron 9650 roughly double the per-drive ceiling of Gen5. That matters for loading model weights and writing checkpoints, but only when the rest of the platform has Gen6 slots.
Buying in Singapore
SourceIT lists the NVIDIA RTX PRO 6000 Blackwell Workstation Edition at S$29,100.00 and the RTX PRO 5000 Blackwell at S$14,971.20. For PCIe Gen5 storage we list the Micron 7600 PRO 3.84 TB at S$5,183.00 and the Micron 9550 PRO 7.68 TB at S$10,456.00 (all listed 6 Nov 2026). We do not currently list a PCIe Gen6 SSD; ask us to source one. GB300 systems are price on request, including the HP ZGX Fury G1n, ASUS ExpertCenter Pro ET900N G3 and Dell Pro Max with GB300. See the AI Workstations and AI Servers collection for more. 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
How fast is PCIe 6.0 compared with NVLink? A PCIe 6.0 x16 link is rated at 256 GB/s bidirectional, or about 121 GB/s per direction after overhead. NVLink 5 on Blackwell is 1.8 TB/s bidirectional per GPU, 900 GB/s per direction, so about seven times more.
When will PCIe 7.0 and 8.0 products arrive? PCIe 7.0 was released to members on 11 June 2025. PCIe 8.0 is planned for release by 2028, with devices expected around 2030 or later. Product timing for both depends on CPU, switch and device vendors.
Is PCIe 6.0 already in use? Yes, in AI hardware first. Examples include NVIDIA's Blackwell Ultra host interface, the ConnectX-8 SuperNIC's built-in switch and the Micron 9650 SSD. General-purpose server CPUs and PCIe switches are following.
Will PCIe ever replace NVLink? Not as currently designed. Each generation narrows the raw-bandwidth gap, but NVLink also advances and adds switching across 72 GPUs, coherence and in-network reductions. The open challengers to NVLink are UALink and Ultra Ethernet, not PCIe.
Do RTX PRO workstation cards use NVLink? SourceIT's RTX PRO Blackwell listings do not mention NVLink. These cards connect over PCIe, so data moving between cards in a multi-GPU workstation crosses PCIe.
