Calculator

PCIe and CXL bandwidth calculator

Circuit design and memory. Everything is calculated in your browser; nothing you enter is sent to a server.

KEY POINTS

PCIe bandwidth is transfer rate (GT/s) × lanes × encoding efficiency ÷ 8: about 63 GB/s one way for PCIe 5.0 x16 and about 118 GB/s for PCIe 6.0 x16. GT/s counts bit transfers, not data. CXL 2.0 uses the PCIe 5.0 physical layer and CXL 3.x uses PCIe 6.x.

Inputs

One-way bandwidth (after encoding)63.02GB/s
Total in both directions126.03GB/s
Raw transfer rate (one way)64.00GB/s
Encoding efficiency98.46%
TLP efficiency (header relative to MPS)91.4%
Payload bandwidth (estimate)57.61GB/s
Time to transfer 100 GB (estimate)1.74s

Payload bandwidth roughly subtracts TLP headers and DLLPs (ACKs and flow control); on real systems it is set by device processing and the DMA implementation. FLIT efficiency for PCIe 6.0 and later includes FEC and CRC.

Bandwidth by generation and lane count (one way, GB/s)

GenerationTransfer rateEncodingx1x4x8x16Spec released
PCIe 1.02.5 GT/s8b/10b0.2501.02.04.02003
PCIe 2.05 GT/s8b/10b0.5002.04.08.02007
PCIe 3.08 GT/s128b/130b0.9853.97.915.82010
PCIe 4.016 GT/s128b/130b1.9697.915.831.52017
PCIe 5.0 / CXL 2.032 GT/s128b/130b3.93815.831.563.02019
PCIe 6.0 / CXL 3.x64 GT/sPAM4, FLIT7.37529.559.0118.02022
PCIe 7.0128 GT/sPAM4, FLIT14.75059.0118.0236.02025

GT/s versus GB/s

PCIe speeds in GT/s (gigatransfers per second) count how many bits per second travel on one lane's signal pair, not the amount of data. Data is encoded to make the clock easy to recover (20% overhead for 8b/10b, about 1.5% for 128b/130b), so usable bandwidth is lower by that amount. From PCIe 6.0, signaling changed to four-level PAM4 with fixed-size packets (FLITs) that include error correction.

In AI servers, the bandwidth between GPU and CPU, between GPU and network card, and to CXL memory expansion shapes performance. PCIe roughly doubles in speed each generation, but signal loss grows with it, prompting lower-loss board materials, added retimers and discussion of a shift to optical links (co-packaged optics).

FAQ


What is the bandwidth of PCIe 5.0 x16?

32 GT/s × 16 lanes × 128/130 ÷ 8 is about 63 GB/s in each direction, about 126 GB/s in total.

What changed in PCIe 6.0?

It moved to four-level PAM4 signaling and fixed-size FLIT packets with forward error correction, doubling the rate to 64 GT/s.

Related equipment and companies


More calculators


Circuit design and memory

Last updated: October 23, 2026