Choose your country or region.

EnglishFrançaispolskiSlovenija한국의DeutschSvenskaSlovenskáMagyarországItaliaहिंदीрусскийTiếng ViệtSuomiespañolKongeriketPortuguêsภาษาไทยБългарски езикromânescČeštinaGaeilgeעִבְרִיתالعربيةPilipinoDanskMelayuIndonesiaHrvatskaفارسیNederland繁体中文Türk diliΕλλάδαRepublika e ShqipërisëአማርኛAzərbaycanEesti VabariikEuskera‎БеларусьíslenskaBosnaAfrikaansIsiXhosaisiZuluCambodiaსაქართველოҚазақшаAyitiHausaКыргыз тилиGalegoCatalàCorsaKurdîLatviešuພາສາລາວlietuviųLëtzebuergeschmalaɡasʲМакедонскиMaoriМонголулсবাংলা ভাষারမြန်မာनेपालीپښتوChicheŵaCрпскиSesothoසිංහලKiswahiliТоҷикӣاردوУкраїнаO'zbekગુજરાતીಕನ್ನಡkannaḍaதமிழ் மொழி

Vice President of Marketing for ARM: The key to heterogeneous computing is to enable third-party access to chip-level performance

Support for heterogeneous computing Arm introduces two new mainstream ML processors

Heterogeneous computing has gradually become a mainstream direction in the industry. Intel and NVIDIA have launched a unified architecture platform to achieve heterogeneous computing. Recently, to support heterogeneous computing, Arm introduced two new mainstream ML processors, Ethos-N57 and Ethos-N37 NPUs, which are two processors following the Arm ML processor Ethos-N77.

In this regard, Arm marketing vice president Ian Smythe said in an interview: "The key to supporting heterogeneous computing is not to build a hardware platform, but more importantly, to enable third-party developers to access chip-level performance. Because if they can't get chip-level performance, IP capabilities, then these capabilities, these performances are already wasted. So we emphasize the cooperation of the whole ecosystem."

“In fact, it is very clear that there is a specific area of ​​computing, and we are considering how to meet the energy consumption of next-generation computing and the performance across CPU, GPU and NPU. This means that we not only need In the product design phase, it is also necessary to effectively establish data or combinations of different scene calculations during the operation phase of the developer's product deployment. In order to do this, we need a unified tool chain to implement CPU, GPU, NPU. Support," Ian Smythe said.

In addition, Ian Smythe pointed out that in fact, Arm has many such partners who have implemented heterogeneous system-on-chip on their TV or mobile phone products, such as using video, graphics, accelerators and CPUs. It is heterogeneous. Only ARM has implemented a better, system-wide data flow at the system level from a developer's perspective.

"For Arm, we focus on total computing, whether it's total computing or heterogeneous computing or proprietary computing, we focus on the balance between power and performance," Ian Smythe said.

It is reported that the design concept of Ethos-N57 and Ethos-N37 includes some basic principles, such as: optimization for the support of Int8 and Int16 data types; advanced data management technology to reduce data movement and related power consumption; The landing of the innovative Winograd technology has improved performance by more than 200% over other NPUs.

In addition, the Ethos-N57 features include: optimized for ML performance and power efficiency and performance range for 2 megabits per second. The Ethos-N37 also features: Designed to provide the smallest ML inference processor (less than 1 square millimeter); optimized for a performance range of 1 megabits per second.