Research Overview
In today's world, innovations and new technologies are rapidly transforming our daily lives. We use technology at every moment of our lives, and as a result, we are surrounded by computers in one or the other form. At the heart of every computer lies the central processing unit (CPU), responsible for processing data. We are interested in the architecture of these processing units, often containing multi-/many-cores, to get the work done faster. We are developing techniques to improve the performance and energy efficiency of these multi-/many-core processors including graphics processing units (GPU). In other words, we are happy when we get more work done from these computing devices with less energy. Through our research, we hope to contribute towards the global goal of reducing carbon emissions and creating a sustainable future.
To demonstrate the potential impact of our work, let's consider the case of discrete GPUs. In 2017, approximately 357 million discrete GPUs were sold worldwide, consuming a total of 89 million kW of energy per hour, assuming a typical power consumption of 250 W per GPU. This amounts to a cost of approximately 27 million euros per hour. Now let's assume that we can increase the energy efficiency of a GPU by a mere 5%. When we scale over the 357 million discrete GPUs, it can save us 4.5 million kW of energy per hour, which translates to a cost savings of 1.35 million euros per hour. If we put it into a different perspective, this much energy savings would be sufficient to provide electricity to 11.3 million households in Germany for one hour. In our opinion, this is a lot!
Dr. Lal's Ph.D. research has already shown promising results, improving GPU energy efficiency by about 20%. While this research has been limited to the academic community and not implemented in real-world GPUs, it demonstrates the potential impact of our work. We are actively exploring ways to translate these research findings into practical, real-world solutions to help reduce energy consumption and carbon emissions.
If you are interested in doing research with our group, you may send an e-mail. If you are already at Hamburg University of Technology, you can also schedule an in-person meeting.
Broad Areas of Research
- Energy-Efficient GPU architecture
- Applying machine learning for microarchitectural improvements
- Power and performance modeling
- Runtime support for HPC clusters
- Memory system, MAG-aware memory compression
- Heterogeneous computing (porting/optimizing applications)
- Approximate computing
- GPU security