Show HN: Our GLM-5.3 Flash Switchless recipe is now out for 4x DGX Sparks
github.com
1 thread
This is the culmination of several weeks' worth of R&D and testing. It builds on OSS components and we credit others.
Could a 200-400G switch be faster? Potentially. Sparks are compute/memory bound and we've had at least one person with said 1300GBP switch comment and say how our numbers matched his on X.
No "count to 200" as a benchmark, we've had this hooked up to opencode for Slicer/OpenFaaS/inlets development over the course of the week along with various OSS contributions.
And my current favourite for a 2x Spark setup is DS4F all day long NVFP4 (which is a mixed quant, of higher quality than straight 4-bit)