𝗔 𝗖𝗮𝘀𝗲 𝗦𝘁𝘂𝗱𝘆 𝗼𝗳 𝟲𝟰-𝗯𝗶𝘁 𝗙𝗹𝗼𝗮𝘁𝗶𝗻𝗴 𝗣𝗼𝗶𝗻𝘁 𝗘𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗶𝗻 𝗤𝘂𝗮𝗻𝘁𝘂𝗺 𝗘𝘀𝗽𝗿𝗲𝘀𝘀𝗼

The evolution of computer architecture, driven by AI applications, has significantly enhanced the reduced- and mixed-precision computing capabilities of GPUs, offering various low-precision representations such as FP16, BF16, and FP8. There is growing interest in utilizing these capabilities through mixed-precision algorithms and emulation techniques to boost scientific computing performance without compromising accuracy.

This presentation explores the potential of leveraging lower precision Tensor Cores for scientific applications by emulating FP64 and FP32 matrix multiplications without sacrificing accuracy, thereby achieving improved performance per watt.

We will discuss a novel library developed at NVIDIA, which transparently accelerates DGEMM and ZGEMM calls using INT8 Tensor Cores. Results from several use cases in Quantum Espresso will be presented, highlighting the effectiveness of this approach.

Would you like to attend the seminar? Send an e-mail to our Director Ivan Girotto: igirotto@ictp.it