BDTI has released BDTI DSP Kernel Benchmarks™ results for the CEVA-TeakLite-III core from CEVA. As we’ve written previously, CEVA-TeakLite-III is a 32-bit DSP core that primarily targets audio applications (both portable and high-definition) but also targets VoIP and cellular baseband. It is the third generation of CEVA’s TeakLite architecture, and the first to use a native 32-bit data size. The CEVA-TeakLite-III also supports SIMD (single-instruction, multiple data) dual-16-bit MACs.
The CEVA-TeakLite-III competes with a range of general-purpose DSP and CPU cores from vendors such as VeriSilicon, ARM, and MIPS, and also with application-specific audio solutions, such as Tensilica’s 330HiFi audio core and ARC’s Sound Subsystems cores. According to CEVA, the CEVA-TeakLite-III core runs at up to 550 MHz in a 65 nm process and 335 MHz in a 130 nm process; in 130 nm, it consumes 0.47 mm2 and 70 mW.
BDTI benchmarked the CEVA-TeakLite-III core using the BDTI DSP Kernel Benchmarks™, a suite of 12 common signal processing algorithms (including various FIR and IIR filters, an FFT, a Viterbi decoder, etc.) The results on these benchmarks are used to generate BDTI’s composite DSP speed score, the BDTImark2000™ (or BDTIsimMark2000™, for processors whose results have been verified on a simulator rather than on hardware). A higher BDTImark2000 score indicates a faster processor.
At a clock speed of 335 MHz (in a 130 nm process), the CEVA-TeakLite-III core obtains a BDTIsimMark2000 score of 2140. Though the TeakLite-III core supports single-cycle 32-bit data computations, the BDTI DSP Kernel Benchmarks were implemented using 16-bit data since this is the data width for which the core achieves its highest DSP performance. Therefore, the performance reported here is not indicative of the performance of the core using 32-bit data. (All benchmark results in this article assume use of the TSMC CL013G process and the ARM Artisan SAGE-X library. For more information, including additional scores, see www.bdti.com/Services/Benchmarks.)
Compared to CEVA’s CEVA-X1620, which has a BDTIsimMark2000 score of 2660 at 330 MHz, TeakLite-III is not quite as fast. This isn’t surprising since the X1620 targets higher-performance applications and has a somewhat more powerful instruction set. Compared to the superscalar VeriSilicon ZSP500 DSP core, the TeakLite-III core is much faster, primarily due to its faster clock rate. The ZSP500 runs at 205 MHz and has a score of 1620. TeakLite-III is also much faster than ARM’s ARM1136 CPU core, which achieves a score of 1160 at 330 MHz. (BDTI has not yet implemented the BDTI DSP Kernel Benchmarks on Tensilica’s 330HiFi or ARC’s Sound Subsystems, so we cannot make confident performance comparisons to these cores).
In addition to evaluating speed, BDTI uses this benchmark suite to evaluate cores’ area and energy efficiency. It is here that the CEVA-TeakLite-III core really shines; compared to the ZSP500, ARM1136, and CEVA-X1620, it has the best area efficiency (BDTIsimMark2000/mm2) and the best energy efficiency (BDTIsimMark2000/mW) of the group.
As signal processing has become ubiquitous in embedded applications, low-cost CPU cores (such as the ARM1136) are increasingly tasked with DSP workloads, and are becoming more capable in this realm. However, our benchmark results indicate that DSP cores still generally deliver superior area- and energy-efficiency on DSP-intensive workloads.
Additional details on the CEVA-TeakLite-III core and BDTI’s benchmark analyses are available in BDTI’s new white paper on the CEVA-TeakLite-III core. Additional benchmark scores are available at www.bdti.com/Services/Benchmarks.
Add new comment