Repository navigation
[Question] Clarification on FP8 Micro-block Scaling and FP4 Support Timeline #47
Description
Hi cuTile team,
I have two specific questions regarding the support for Blackwell-specific hardware features:
- Automatic Micro-block Scaling for FP8
When using fp8 with ct.matmul, how is the Micro-block Scaling (1x16) handled?
Automation: Does the tileiras compiler automatically handle the scaling logic and hardware invocation (5th-gen Tensor Core) under the hood?
Explicit Scaling: If it is not fully automatic, how should we provide the scale-factor tiles to the ct.matmul operator? Currently, the ct.matmul(A, B) signature seems to only accept data tiles. Is there a plan for a signature like ct.matmul(A, B, A_scale, B_scale)?
- NVFP4 (FP4) Support Roadmap
The current documentation and samples focus on fp8 and bf16. Since Blackwell's throughput peak is tied to NVFP4:
When can we expect the support for 4-bit narrow-precision tiles in cuTile Python?
Thanks for this great library!
Reacted by Edward, Pasha Khosravi and Yi Liu
Activity
haijieg commentedon Jan 8, 2026
Hi @ultranationalism
The current TileIR release (CTK 13.1) only provides Tensorcore enabled FP8 e4m3fn and e5m2 matmul without micro-scaling. Micro scaling (MXFP8) and NVFP4 support is being actively worked on and will be provided in the near future CTK releases. As for API in cutile, we haven't started designing it yet. Something like
ct.scaled_matmul(A, B, A_scale, B_scale)is a reasonable starting point.Edward-lyz commentedon Jan 12, 2026
That's great! This is a really important feature for developers.
odelame commentedon May 13, 2026
Hey, is there an update on this? An expected timeline? Is it still being worked on?
haijieg commentedon May 14, 2026
Hi @odelame, yes it's being worked on and will be ready soon.
ultism commentedon May 14, 2026
What a pity. I've already written my nvfp4 operator using the CUTE DSL and tuned it to production-ready level. I'll just have to wait for another chance to try out this framework.
https://github.com/ultranationalism/svdquant-kernels :)
haijieg commentedon May 27, 2026
mma_scaled, mxfp8 and nvfp4 is supported by v1.4.0.