From WikiChip
Difference between revisions of "hisilicon/kirin/970"
< hisilicon‎ | kirin

(added utilizing devices)
(Neural Network Processing Unit (NPU))
Line 98: Line 98:
 
== Neural Network Processing Unit (NPU) ==
 
== Neural Network Processing Unit (NPU) ==
 
[[File:kirin 970 npu.png|right|350px]]
 
[[File:kirin 970 npu.png|right|350px]]
The Kirin 970 incorporates a new Neural Network Processing Unit (NPU) designed specifically to be used as an AI accelerator. According to CEO Mr. Richard Yu who introduced the Kirin 970 at 2017 IFA, the NPU uses up the die area of roughly half of the CPU while consuming 50% less power and performing around 25 times faster than a traditional CPU for tasks such as photo recognition. The exact architectural detials of the NPU have been withheld. The NPU is said to deliver 1.92 [[TFLOP]]s (HP 16-bit).
+
The Kirin 970 incorporates a new Neural Network Processing Unit (NPU) designed specifically to be used as an AI accelerator. According to CEO Mr. Richard Yu who introduced the Kirin 970 at 2017 IFA, the NPU uses up the die area of roughly half of the CPU while consuming 50% less power and performing around 25 times faster than a traditional CPU for tasks such as photo recognition. The NPU is said to deliver 1.92 [[TFLOP]]s (HP 16-bit). While the exact architectural detials of the NPU have been withheld, the NPU appear to be a licensed IP design from [[Cambricon Technologies]].
  
 
== Utilizing devices ==
 
== Utilizing devices ==

Revision as of 01:04, 16 September 2017

Template:mpu Kirin 970 is a 64-bit octa-core high-performance mobile ARM LTE SoC introduced by HiSilicon in mid-2017 at the 2017 IFA. This chip, which is fabricated on a 10 nm process, features four Cortex-A73 big cores operating at up to 2.4 GHz along with four Cortex-A53 little cores operating at up to 1.8 GHz. The 970 incorporates ARM's Mali G72 (12 core) IGP operating at 850 MHz and supports up to 6 GiB of dual-channel LPDDR4-1866 memory.

Overview

Introduced at the 2017 IFA, the overall core organization is identical to the Kirin 960 which was introduced the previous year, but features 20% power efficiency and 40% smaller die area due to the process shrink. The 970 ballooned to over 37.5% more transistors from 4 billion in the 960 to 5.5 billion. The 970 adds many enhancements, including a more powerful Mali G72 GPU and incorporates a new Neural Network Processing Unit (NPU) designed for AI acceleration. The 970 has two improved ISPs and a more powerful LTE modem supporting up to User Equipment (UE) category 18 capable of reaching a maximum downlink of 1.2 Gbps (4x4 MIMO, 256 QAM, 3CC CA).

Cache

Main articles: Cortex-A53 § Cache and Cortex-A73 § Cache

[Edit/Modify Cache Info]

hierarchy icon.svg
Cache Organization
Cache is a hardware component containing a relatively small and extremely fast memory designed to speed up the performance of a CPU by preparing ahead of time the data it needs to read from a relatively slower medium such as main memory.

The organization and amount of cache can have a large impact on the performance, power consumption, die size, and consequently cost of the IC.

Cache is specified by its size, number of sets, associativity, block size, sub-block size, and fetch and write-back policies.

Note: All units are in kibibytes and mebibytes.
New text document.svg This section requires expansion; you can help adding the missing info.

Memory controller

[Edit/Modify Memory Info]

ram icons.svg
Integrated Memory Controller
Max TypeLPDDR4-1866
Supports ECCNo
Max Mem6 GiB
Controllers1
Channels2
Width64 bit
Max Bandwidth27.82 GiB/s
28,487.68 MiB/s
29.871 GB/s
29,871.498 MB/s
0.0272 TiB/s
0.0299 TB/s
Bandwidth
Single 13.91 GiB/s
Double 27.82 GiB/s

Graphics

[Edit/Modify IGP Info]

screen icon.svg
Integrated Graphics Information
GPUMali-G72
DesignerARM Holdings
Execution Units12Max Displays2
Frequency850 MHz
0.85 GHz
850,000 KHz
OutputDSI

Standards
DirectX12
OpenCL2.0
OpenGL ES3.2
OpenVG1.1
Vulkan1.0
  • Hardware Acceleration
    • Decode: 2160p @ 60fps
    • Encode: 2160p @ 30fps

Wireless

  • LTE Modem
    • Up to User Equipment (UE) category 18
      • Downlink of up to 1.2 Gbps (4x4 MIMO, 256 QAM, 3CC CA)
      • Uplink of up to 150 Mbps (2x20MHz CA, 64-QAM)
  • Wi-Fi 802.11 ac Dual Band
  • Bluetooth 4.2
  • NFC
  • GPS / A-GPS / GLONASS / BDS / Galileo

Expansions

Neural Network Processing Unit (NPU)

kirin 970 npu.png

The Kirin 970 incorporates a new Neural Network Processing Unit (NPU) designed specifically to be used as an AI accelerator. According to CEO Mr. Richard Yu who introduced the Kirin 970 at 2017 IFA, the NPU uses up the die area of roughly half of the CPU while consuming 50% less power and performing around 25 times faster than a traditional CPU for tasks such as photo recognition. The NPU is said to deliver 1.92 TFLOPs (HP 16-bit). While the exact architectural detials of the NPU have been withheld, the NPU appear to be a licensed IP design from Cambricon Technologies.

Utilizing devices

  • Huawei Mate 10

This list is incomplete; you can help by expanding it.

Facts about "Kirin 970 - HiSilicon"
has ecc memory supportfalse +
integrated gpuMali-G72 +
integrated gpu base frequency850 MHz (0.85 GHz, 850,000 KHz) +
integrated gpu designerARM Holdings +
integrated gpu execution units12 +
max memory bandwidth27.82 GiB/s (28,487.68 MiB/s, 29.871 GB/s, 29,871.498 MB/s, 0.0272 TiB/s, 0.0299 TB/s) +
max memory channels2 +
supported memory typeLPDDR4-1866 +
used byHuawei Mate 10 +