Vectorplay120 fps

Technology

What is between your controller and the frame

A full stack of specific engineering choices, most of which exist to remove milliseconds.

Inputraw HIDTransportQUIC datagramsRenderdedicated GPUEncodehardware AV1Presentclient decode

Decisions worth explaining

AV1 where the client supports it

At the same perceptual quality, AV1 costs about 30% less bitrate than H.265, which matters enormously on a constrained domestic line. Hardware encode keeps the latency cost under a millisecond.

QUIC datagrams, not TCP

Unreliable datagrams with our own loss concealment. A lost packet degrades one slice of one frame instead of stalling the stream while TCP retransmits.

Whole GPUs

Time-slicing a GPU between two players produces exactly the kind of frame-time variance that makes cloud gaming feel bad. We do not do it, and it is why our per-hour cost is what it is.

Client-side prediction where it helps

Mouse-look is predicted locally and reconciled, which removes the sensation of drag in first-person games without affecting what the server sees.

Measured, weekly

Added latency by region — p50
1FRA1AMS1LON1PAR0MAD0MIL1WAW1STO0IAD1ORD1DFW1LAX1NRT1SYD
14 ms
Best region
17 ms
Median
31 ms
p95 across regions
weekly
Measured