PCIe The data link between the processor, the graphics card and the SSD.
Short answer. PCIe (PCI Express) is the high-speed serial connection that links a computer’s processor to graphics cards, SSDs and network cards. A link is made of lanes, written ×1, ×4 or ×16, and each generation doubles the speed per lane: PCI-SIG gives 32 GT/s for PCIe 5.0, with up to 128 GB/s across a ×16 link counting both directions, and 64 GT/s for PCIe 6.0. In AI work PCIe matters mainly when loading models from an SSD, when a model spills out of GPU memory, or when several GPUs share work. NVIDIA’s RTX PRO 4000 Blackwell uses PCIe Gen 5; AMD lists PCIe 4.0 for the Ryzen AI Max+ 395.
01
Lanes and generations
PCI-SIG, the body that defines PCIe, states that PCIe 5.0 runs at 32 GT/s and 128 GB/s for a ×16 link, and that PCIe 6.0 doubles both to 64 GT/s and 256 GB/s, the latter counting both directions. NVIDIA says Gen 5 doubles the bandwidth of Gen 4, and NVM Express says a Gen 3 ×4 link delivers 4 GB/s. The Gen 3 and Gen 4 rows below follow from that doubling. Real throughput is a little lower than these link rates.
Generation
Rate per lane
×16 link, both directions
×4 link, one direction
PCIe 3.0
8 GT/s
about 32 GB/s
about 4 GB/s (NVM Express)
PCIe 4.0
16 GT/s
about 64 GB/s
about 8 GB/s
PCIe 5.0
32 GT/s
128 GB/s (PCI-SIG)
about 16 GB/s
PCIe 6.0
64 GT/s
256 GB/s (PCI-SIG)
about 32 GB/s
02
Where PCIe shows up in an AI box
Loading a model: weights travel from the SSD to memory once at start-up. A 20 GB file over a link that carries 4 GB/s takes at least 5 seconds; real drives are slower than the link.
Spilling past GPU memory: if a model does not fit in VRAM, offloaded layers cross PCIe. At about 64 GB/s each way on a Gen 5 ×16 link, that is roughly a tenth of the 672 GB/s NVIDIA lists for the card’s own memory.
Several GPUs: cards that share one model exchange data over the links between them.
Chips with shared memory avoid the hop: AMD says the CPU and GPU in the Ryzen AI Max+ 395 share data without expensive memory copies.
03
Reading a specification sheet
Note the generation and the lane count for each device: AMD lists PCIe 4.0 with 16 usable lanes for the Ryzen AI Max+ 395, so ask the system maker how they are shared between any expansion slot and any SSDs.
Ask for the electrical lane count of a slot, not only its physical size; a ×16 slot may be wired for fewer lanes.
PCI-SIG says each generation is backwards compatible, so a newer card works in an older slot, at the older speed.
Match the SSD generation to the slot: a fast drive in a slow slot is held back by the slot.
04
Where ANVE stands
ANVE Office Pro uses an NVIDIA RTX PRO 4000 Blackwell, a PCIe Gen 5 card per NVIDIA’s specification, and ANVE Office uses its SFF Edition, while ANVE One uses an AMD Strix Halo class chip, which AMD lists with PCIe 4.0. ANVE has published no speed figures, and no sealed unit exists yet.
05
Questions people ask
Does PCIe speed affect how fast an AI answers?
Mostly not once the model is in GPU memory. It affects loading time and any layers offloaded to system memory.
What does GT/s mean?
Billions of transfers per second on one lane. PCI-SIG quotes GT/s for the raw bit rate; usable data per second is a little lower.
Can I put a PCIe 5.0 card in a PCIe 4.0 slot?
Yes. PCI-SIG states that PCIe generations stay backwards compatible, so the link runs at the slower generation’s speed.
Do I need a PCIe 5.0 SSD for AI?
Probably not. SSD speed sets how quickly a model loads, not how fast it answers. A 20 GB model at 4 GB/s loads in about 5 seconds even on a Gen 3 ×4 link.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
