Availability of Pre-generated Hopper / Blackwell Traces

[yashvarr]

Hi, thanks for the great work on Accel-Sim 2.0 and the new support for modern architectures such as Hopper.

I had a question regarding trace generation and accessibility of traces for newer GPUs such as H100/H200 and B200.

As I understand it, the Accel-Sim tracer relies on NVBit and dynamic instrumentation, meaning that generating a trace for a particular architecture requires access to actual hardware from that architecture. For example, generating Hopper-specific traces would require access to an H100/H200 system.

This creates a practical barrier for researchers and students who want to experiment with or extend the simulator but do not have access to expensive modern GPUs. While one can build and run the simulator locally, reproducing workloads targeting H100 or newer architectures seems difficult without access to the physical hardware needed for trace generation.

Would it be possible to provide a collection of pre-generated traces for common workloads on newer architectures, similar to the existing benchmark trace distributions?

[JRPan]

Yes! We are working on that. We expect to push something with a week or two for Hopper.

[yashvarr]

Hi, any updates on this?

[JRPan]

Hopper traces are now available with Accel-Sim v2.0.1. The release includes H200 traces for 20 ML workload configurations (BERT, ResNet-50, DLRMv2, 3D U-Net, RetinaNet, Whisper, SDXL, Llama 3.1 8B inference and training, Qwen 2.5 7B, DeepSeek R1 and Mixtral 8x7B), each with matching Nsight Compute stats for correlation.

The Pre-collected Traces section of the README covers what is available, how to download it, and how to simulate and correlate it. In short:

./get-accel-sim-traces.py -l                                  # list what is available
./get-accel-sim-traces.py -a H200/inference_bert.sh/32_256    # download one configuration

Blackwell traces are not included yet.

[yashvarr]

Thanks!