Published Kal-El (TEGRA3) performance: is NVIDIA SoC truly faster than a Core2 ?
All peoples involved in IT and computer technologies in the last decade know very well NVIDIA: this graphic & compute chips design company reached many important target and set new state-of-the-art performance in about any marked where it operated.
Some day ago, NVIDIA uncovered its next-generation SoC project, codenamed Kal-El. This SoC is going to set new performance standard in its area, featuring four ARM Cortex A9 core (each with NEON support, thanks to the integrated MPE) and a renewed, 12 core wide graphic controller. Anandtech did a great job in explaining Kal-El architecture, so I advice anyone interested in SoC performances to read his article here: http://www.anandtech.com/show/4181/nvidias-project-kalel-quadcore-a9s-coming-to-smartphonestablets-this-year
However, NVIDIA did not limit itself to announce its new hardware beast, but showed some benchmark numbers to prove the superior speed of Kal-El. The used benchmark was CoreMark 1.0 (you can read about it here: http://www.coremark.org/home.php), a synthetic ALU and FPU benchmark. Kal-El performs very well here, doubling Tegra2 performance: this is an extremely great accomplishment, as the tested Kal-El silicon was only 12 days old, and this is a testament to NVIDIA's ability to design top-notch chip.
Moreover, Kal-El was not only compared to current Tegra2 hardware, but also against an aging Core2 T7200 (a dual core mobile processor @ 2.0 GHz with 4 MB of L2 cache) and, by NVIDIA measurement, Kal-El is faster than this x86 processor:
However, many readers discovered a strange thing: the compiler version used for the Core2 processor is very old, and the optimization flags used are much more conservative than the ones used on the two NVIDIA SoCs:
As you can see, NVIDIA benchmarked its processors with a recent GCC version (4.4.x branch) and very aggressive optimization settings (they used not only O3 but tuned some hardware-specific settings also). On the other hand, the Intel processor was benchmarked with a very old GCC version (3.4.x branch) and only “normal” (O2) optimization settings.
These consideration raises an important question: is Kal-El truly faster than a Core2 processor, or this result is only an artifact of the different compilers and optimization settings? Well, in this article we will try to answer to this question...
Comments
RSS feed for comments to this post