Pyxis collects sparse accelerators and their perforamnce data. We hope this dataset is helpful for inereted researchers in the fileds of computer architecture, FPGA, systems, perforamnce and algorithm. Specifically,
- (1) the collected acclerators may be a good refrence design for hardware related researchers,
- (2) the collected performance data serves as a large scale (73.8K) labels of graphs with various sizes and my be helpful for graph algorithm researchers.
- HLS tool: Vitis 2020.2
- U250 SpMM Accelerator: Sextans-U250
- Number of collected perforamnce instances: 18,459
- Runing on 2,637 sparse matrices from SuiteSparse
- Files: Pyxis/perfdata/20211003_U250_binary.txt, Pyxis/perfdata/20211003_U250_real.txt.
- One line is a perforamcne instance with 8 properties: [Matrix ID] [Row Number: M] [Column Number: K] [NNZ] [N] [Latency(ms)] [Throughput(GFLOP/s)]
- HLS tool: Vitis 2020.2
- U280 SpMM Accelerator: Sextans-U280
- Number of collected perforamnce instances: 18,459
- Runing on 2,637 sparse matrices from SuiteSparse
- Files: Pyxis/perfdata/20211003_U280_binary.txt, Pyxis/perfdata/20211003_U280_real.txt.
- One line is a perforamcne instance with 8 properties: [Matrix ID] [Row Number: M] [Column Number: K] [NNZ] [N] [Latency(ms)] [Throughput(GFLOP/s)]
We also collocted perfoamnce data on two GPUs.
- CUDA version: 10.2
- Number of collected perforamnce instances: 18,459
- Runing on 2,637 sparse matrices from SuiteSparse
- Files: Pyxis/perfdata/20211003_K80_binary.txt, Pyxis/perfdata/20211003_K80_real.txt.
- One line is a perforamcne instance with 8 properties: [Matrix ID] [Row Number: M] [Column Number: K] [NNZ] [N] [Latency(ms)] [Throughput(GFLOP/s)]
- CUDA version: 10.2
- Number of collected perforamnce instances: 18,459
- Runing on 2,637 sparse matrices from SuiteSparse
- Files: Pyxis/perfdata/20211003_V100_binary.txt, Pyxis/perfdata/20211003_V100_real.txt.
- One line is a perforamcne instance with 8 properties: [Matrix ID] [Row Number: M] [Column Number: K] [NNZ] [N] [Latency(ms)] [Throughput(GFLOP/s)]
More details coming soon.
Or you can go to this link to find details.
If you find this dataset useful, please cite:
@misc{pyxis2021,
Author = {Linghao Song and Yuze Chi and Jason Cong},
Title = {Pyxis: An Open-Source Performance Dataset of Sparse Accelerators},
Year = {2021},
Eprint = {arXiv:2110.04280},
}