Today SGI’s Gabriel Broner announced that the company’s SGI UV 3000 system had set two new benchmarking records, demonstrating the machine’s advanced throughput prowess enabled by its shared memory architecture. The test machine — a 128 socket UV 3000 system equipped with 1,280 E5-4627 v3 2.60 GHz Xeon cores — was assessed using the industry-standardized SPECrate metrics (specifically SPECint_rate_base2006 and SPECfp_rate_base2006), which measure the throughput or rate of a machine carrying out a number of simultaneous tasks.
According to results published by the Standard Performance Evaluation Corporation (SPEC), the system achieved a score of 45,100 for SPECint_rate_base2006 (source) and 42,600 for SPECfp_rate_base2006 (source). Broner notes that the new results put SGI ahead of its nearest competitors by more than 25 percent for SPECfp_rate_base2006 and more than 10 percent for SPECint_rate_base2006 — using 20 percent fewer cores. As you can see in the chart below, SGI overtook a ScaleMP machine comprising 1,536 Xeon E5-2680 v3 2.50 GHz cores.
“SGI has also published performance results of over twice its closest competitors for both the SPECompG_base2012 shared-memory parallel processing benchmark and the STREAM TRIAD memory bandwidth benchmark,” adds Broner.
The SGI UV 3000 test platform incorporates 128 10-core Intel Xeon E5-4627 v3 chips (2.60 GHz with Intel Turbo Boost Technology up to 3.20 GHz) and 16 TB of memory. The operating system is SUSE Linux Enterprise Server 12.
The SGI UV 3000 is a symmetric multiprocessing (SMP) system that targets compute-intensive, fast algorithm workflows. Such workflows are the hallmark of CAE, genome assembly and scientific simulation, domains which require the ability to solve complex problems at advanced scales. SPEC’s benchmark suites are designed using workloads developed from real-world applications and this particular suite (part of SPEC CPU 2006) is considered especially valuable for assessing large, multi-processor systems.
Sections of the relevant results are reproduced below (sourced from SGI):