On March 6, ” of new era forum leads in “AI In 5G—— , development of artificial intelligence estate is allied (AIIA) releases “AIIA DNN Benchmark V0.5” formally first-run evaluate a result. Should evaluate in AIIA authority test platform is finished, be based on upright side to conclude the processor of deepness nerve network of the task is fiducial test. Below setting of 4 big representative application, can reflect the performance level of the processor that has deepness study to handle ability or accelerator objectively. Evaluate a result to show, hai Saiqi Lin the beautiful on expression of 980 He Ruixin small RK3399.
The standard that “AIIA DNN Benchmark V0.5” evaluates is relatively severe exacting, include 4 grand ceremony setting and index of two kinds of big evaluation to wait, index includes rate (Fps) and algorithmic function, wait like Top1, Top5, MAP, MIoU, PSNR. In the meantime, this also is deepness study processor domain distinguishs first whole model the Benchmark that contrasts with floating-point. Development board shows the RK3399 of microelectronics of Fuzhou luck core a data that grab an eye.
The RK3399 of microelectronics of Fuzhou luck core that uses 28nm technology is developed board, strong performance of favorable balance of trade is shown in evaluate. Evaluate data to show, do not need to nod the case that turns new training to fall surely in floating-point model, int8 is calculated with precision loss most greatly the cost of 1% , achieve opposite Yu Fu to nod computation the function of double.
In addition, in Interpretation evaluation, attempt of AIIA first time is quantified in fiducial test lieutenant general and floating-point model is evaluated apart. And RK3399 of microelectronics of Fuzhou luck core is developed board obtain multinomial and exceedingly good data likewise, front row level is in inside course of study.
Be worth what carry is, RK3399 of microelectronics of Fuzhou luck core is developed board the backside that data grabs an eye, cannot leave the support of forward position technology. For instance RK3399 blended in Tengine, latter is a lightweight that develops by OPEN AI LAB modular engine of inference of high-powered nerve network. Tengine is aimed at Arm technically embedded equipment is optimized, and need not depend on tripartite library, can cross platform to use supportive Android, Liunx to wait.
Tengine support is of all kinds and common roll accumulate nerve network, include SqueezeNet, mobileNet, alexNet, resNet, supportive layer confluence, 8 are quantified etc optimize strategy. Through calling the HCL library that in the light of different CPU small structural frame optimizes, can dig the function of Arm CPU adequately disinter to come. And the Cortex-A72 single-track Cheng of RK3399 runs shift to carry commonly used MobileNet, need 111ms only.
In IoT equipment, intelligence on the innovation of the alternant equipment, artificial intelligence equipment such as personal computer, robot and research and development, microelectronics of Fuzhou luck core has shown a banner advantage on the technology, besides RK3399, RK3399Pro of artificial intelligence chip has admiral class extremely likewise hotspot, its use the hardware structure design of CPU+GPU+NPU first. The processor of NPU(nerve network with this one compositive chip) shirt-sleeve microelectronics of Fuzhou luck core learns the crucial skill that waits for a respect in processing of machine vision, speech, deepness, piece on NPU operation function is as high as 3.0TOPs, have high-powered, low power comsumption, development to wait for an advantage easily.
Times of revolution of artificial intelligence chip, development of artificial intelligence estate is allied (authority of AIIA) “AIIA DNN Benchmark V0.5” checks platform appear and first-run data is released, or the situation that lacks terminative AI chip unified data to test a level, quicken the technical decay of artificial intelligence chip and evolution.