Shameless

[josephbreda]

I can’t believe NVIDIA can with a straight face continue to tout “1 Petaflop of FP4 performance” in their promotion of RTX Spark.

NVIDIA RTX Spark — Slim Laptops & Small Desktops

Software support still lags behind the pack many months after release of GB10. Promised acceleration from NVFP4 never materialized. Perhaps they could be given a pass for overoptimistic expectations at the time the DGX Spark was released. Now it’s just apparent that there is no connection between marketing and engineering reality.

This enthusiast early-adopter community has been extremely patient with the pace of development. I don’t think mass-market consumers are going to be as forgiving.

[elsaco]

It’s actually Up to 1 Petaflop FP4 AI Performance!

[josephbreda]

Great. Give my comment a “like” and I’ll give you “up to a billion dollars.” :-)

[j0n]

+1 In for the “up to a billion dollars”

[j0n]

I did notice this as well and was slightly surprised the marketing hasn’t been responsive to the forum feedback. However as someone else commented, the rate of NVFP4 model releases is picking up lately:

and I noticed NVFP4 is a critical part of @tenari 's PrismaQuant mixed-precision quantization scheme.