Copied


NVIDIA Grace CPU Enhances Mathematical Optimization Efficiency and Performance

Rongchai Wang   Jul 13, 2024 17:50 0 Min Read


In a recent development, NVIDIA's Grace CPU has demonstrated substantial advancements in mathematical optimization performance and energy efficiency, according to the NVIDIA Technical Blog. These improvements are poised to benefit industries requiring high computational power and energy-saving solutions.

Enhanced Optimization Capabilities

Mathematical optimization is a crucial tool enabling businesses to make smarter decisions, improve operational efficiency, and reduce costs. However, the complexity of models and the size of datasets necessitate sophisticated AI algorithms and high-performance computing. NVIDIA's new Grace CPU aims to meet these demands with superior computational capabilities.

Founded in 2008, Gurobi Optimization, a leading mathematical optimization solver, received a Supermicro NVIDIA MGX-based system powered by the NVIDIA GH200 Grace Hopper Superchip. This system promises high performance with low power consumption, addressing the need for efficient and fast optimization solutions.

Benchmarking Performance

The benchmark tests utilized a single NVIDIA Grace Hopper Superchip server and a cluster of four AMD EPYC 7313P servers. The test setup included Gurobi Optimizer 11.0 on Ubuntu 22.04, with the Grace Hopper Superchip featuring an Arm-based NVIDIA Grace CPU combined with the NVIDIA Hopper GPU.

Performance evaluations were conducted using the Mixed Integer Programming Library (MIPLIB) 2017, which includes 240 real-world optimization instances. The NVIDIA Grace CPU's results were compared against the commonly used AMD EPYC servers.

Key Findings

The initial benchmarks indicated that the NVIDIA Grace Hopper Superchip outperformed AMD EPYC servers on most hard models, achieving an average runtime of 80 seconds compared to 130 seconds for AMD—a 38% improvement. Additionally, the NVIDIA Grace CPU demonstrated a 23% faster throughput while consuming 46% less energy than the AMD EPYC 7313P.

Further analysis showed energy consumption benefits, with the Grace Hopper using about 1.4 kWh at 8 threads versus 1.75 kWh for AMD, a 20% improvement. At 12 threads, the Grace Hopper used 1.6 kWh compared to 2.6 kWh for AMD, marking a 38% improvement.

Geometric mean runtime
Figure 1: Geometric mean of runtime on NVIDIA Grace CPU compared to AMD EPYC 7313P
Throughput and energy consumption
Figure 2: Throughput and energy on NVIDIA Grace CPU compared to AMD EPYC 7313P
Energy consumption in kWh
Figure 3: Energy consumption for MIPLIB Benchmark set in kWh on NVIDIA Grace CPU compared to AMD EPYC 7313P

Future Outlook

Preliminary benchmarks suggest that the Gurobi Optimizer, when run on the NVIDIA Grace Hopper Superchip, supports faster computational performance with lower energy consumption. This development holds promise for various industries seeking to enhance their energy efficiency while tackling complex business challenges with improved performance.

For an in-depth look at the tests and results, interested readers can view the on-demand session from NVIDIA GTC. More insights into how mathematical optimization can address complex challenges can be found at the Gurobi Resource Center.


Read More