In a technical exchange about wide expert parallelism for trillion-parameter models, Will Brown noted that large Mixture of Experts systems only become cost-effective once batch sizes climb high enough to justify clusters of 32 or 64 GPUs, tying directly into ongoing conversations on reinforcement-learning infrastructure.
@Dorialexander pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way
https://www.primeintellect.ai/blog/rl-at-1t-scale
Yeah vllm/sglang do work quite well know, batching ensures a good theoretical throughout but then everyone want to Claude Code and you have to manage X parallel sessions, each with varying latency.
Brown’s observation aligns with current Wide-EP practice where communication overhead drops once enough tokens are processed together, yet exact batch-size thresholds for specific models remain unstated in the exchange.
The surrounding thread flags tension between high-throughput batch serving and the variable-latency demands of parallel agentic workloads, leaving open how 32-to-64-GPU setups will handle mixed traffic patterns.
No Digg Deeper questions have been answered for this story yet.
@willccbb I know (great post!). But wide-ep is requiring a large GPU stack?
@Dorialexander pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way
https://www.primeintellect.ai/blog/rl-at-1t-scale
@Dorialexander as with all big MoEs, economical serving kicks in at fairly high batch size
32 or 64 gpus is common
@willccbb I know (great post!). But wide-ep is requiring a large GPU stack?
@willccbb Yeah that was the point I intended to develop in next post. As equipment lags, many cos have GPUs lying around but more in the 16-32 h100 range.
@Dorialexander @willccbb I wonder whether there’s an org which helps you sourcing larger clusters for your needs at great prices
@Dorialexander guessing we get something close to glm-5.2-ish in the 400-500b range in the next few months? our results there were h200 fp8, should translate to something a bit smaller
tuning is fiddly but doable