got lightrail backfilling over 400 repos/sec for bits, though it's settling closer to 300/s so looking like over 24h for full-network backfill (it's not actually disk- or cpu-bound, this is all limited by scheduling resyncs from enough different PDS hosts concurrently)
Mar 18, 2026 at 6:05 PM
Run Workers for up to 5 minutes of CPU-time
Workers now support up to 5 minutes of CPU time per request. Allowing more CPU-intensive workloads.
A Full-Network Relay for $34 a Month | bryan newbold
This is an update to a Summer 2024 blog post. At the time, atproto relays required a cache of the full network on local disk to validate data structures. With the Sync v1.1 updates, relays don't need all that disk I/O. What impact does that have on hosting setup and operating costs? Turns out the d...
Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster. Powered by Cerebras, it delivers up to 750 output tokens per second.
Backfilling the Network | Bluesky
Backfilling is the process of syncing all the data in the network from scratch.
the tiny corp on Twitter / X
Qwen 3.5 27B getting 18.5 tok/s on Mac Mini with external 7900XTX. It should be able to be 3x faster than this with work, SSM stuff is still in PR. Hopefully Mac eGPU support brings in devs. pic.twitter.com/2aMkUpXY1S— the tiny corp (@__tinygrad__) April 1, 2026
Steeve Morin on Twitter / X
For instance, we are now loading from SSD at about 14-16GB/s *in a composable manner*.That means about 1.1s to load an 8B BF16 model. https://t.co/i0c7xu8HhP— Steeve Morin (@steeve) November 1, 2025
Benchmarks | EXO
Transparent benchmarks for LLMs tested on real hardware. Coming soon.
GrapheneOS on Twitter / X
We need more 10Gbps or higher dedicated servers for hosting our OS and app updates. We have North America covered well enough via sponsored servers from ReliableSite in both Miami and Los Angeles and a sponsored server from Xenyth in Toronto but no longer have any left in Europe.— GrapheneOS (@GrapheneOS) February 12, 2026
full-network space efficiency update: bad news: event though 84% of all repos are stale (no activity in 30 days), stale repos only make up 20% of the data in the network. moving them to cheap storage doesn't buy much. good news: the data in network is small?? like sub-1TB, minimal magic needed
Bluesky's AppView performance has not been so great recently, right? Long loading times with "Please check your internet connection" or "rate limit exceeded" Thinking it may be folks off of the Bluesky PDS's. Thought it was android oddities, but I get it on desktop and have heard from others
It's 100mb now in the reference PDS installer: github.com/bluesky-social/pds/blob/main/… Though each PDS may be configured with different limits, and I'm not sure if the actual limit is advertised anywhere.
Am I missing something or is the Bluesky new-PDS daily relay limit for an _entire PDS_ lower than the limits they give to a _one single repo_ on their hosted PDSses? docs.bsky.app/docs/advanced-guides/rate-lim…
random atproto sync notes - for whole-network backfill, you have to spread your `getRepo` requests across PDS hosts to get reasonable throughput, so it's a scheduling problem. it's also a spec-SHOULD to make these requests directly to your upstream (not PDSes directly) and follow its redirects.
