Immutable Descriptor Sets vs. Pre-recorded Updates for Dynamic Resources

Answer to the question

When resource volatility is high, immutable descriptor set switching offers more predictable performance across NVIDIA, AMD, and Intel drivers, but at the cost of increased memory usage. If the number of unique descriptor combinations per frame is small, pre‑recording all updates remains preferable because it eliminates the memory overhead and avoids the bind‑cache churn that can occur on some drivers.

Confirmed facts

• Updating a bound descriptor set after vkCmdBindDescriptorSets is undefined behavior per the Vulkan spec.
• Pre‑recording all vkUpdateDescriptorSets calls before command‑buffer recording guarantees spec compliance but fixes descriptor state at recording time.
• Immutable descriptor sets require pre‑allocating a set per frame (or per unique descriptor combination) and switching between them during binding.
• Driver implementations differ: NVIDIA and AMD often handle frequent set switches efficiently, while Intel drivers can suffer from bind‑cache thrashing when many sets are bound per frame.
• Memory overhead scales linearly with the number of descriptor sets created; for complex scenes this can exceed pool limits.

Likely explanation / trade‑off

• High resource volatility → frequent descriptor changes → immutable switching reduces driver‑level synchronization surprises.
• Low volatility → few descriptor changes → pre‑recording keeps memory footprint low and keeps the pipeline state simple.
• Driver bind‑cache behavior can make immutable switching more or less efficient depending on the GPU vendor and the size of the descriptor set.
• Memory overhead can become a bottleneck if you need dozens of descriptor sets per frame for many dynamic objects.

Steps for this case

1. Measure the number of unique descriptor combinations you need per frame (e.g., per dynamic object).
2. Calculate the expected memory cost: num_sets × descriptor_set_size and compare against your pool limits.
3. If the cost is acceptable, pre‑create a descriptor set for each frame and bind the appropriate one during draw calls.
4. Otherwise, keep a single set and perform all updates before recording the command buffer; ensure no updates occur after binding.
5. Profile bind‑cache performance on each target GPU (NVIDIA, AMD, Intel) using validation layers or vendor counters to confirm that set switching does not introduce stalls.
6. If stalls are observed on a particular vendor, consider a hybrid approach: pre‑record updates for the majority of objects and use immutable sets only for the most volatile ones.

Missing diagnostic detail to refine recommendation

Could you provide the average number of descriptor set bindings per frame on your target hardware? Knowing this will help determine whether the bind‑cache overhead of immutable switching is likely to outweigh the memory cost.