Kunpeng AI Inference-accelerated ECSs

Kunpeng AI inference-accelerated ECSs are designed to provide acceleration services for AI services. These ECSs are provided with the Ascend AI Processors and Ascend AI Software Stack.

Kunpeng AI inference-accelerated ECSs use Ascend 310 processors for AI inference acceleration.

**Table 1** Kunpeng AI Inference-accelerated ECSs
Type	Compute	Disk Type	Network
kAi1s	vCPU to memory ratio: 1:1 or 1:2 Number of vCPUs: 4 to 48 Kunpeng 920 processor Base frequency: 2.6 GHz	High I/O General Purpose SSD Ultra-high I/O Extreme SSD General Purpose SSD V2	Ultra-high packets per second (PPS) throughput An ECS with higher specifications has better network performance. Maximum PPS: 2,000,000 Maximum intranet bandwidth: 12 Gbit/s
kAi2	vCPU to memory ratio: 1:4 Number of vCPUs: 16 to 96 Kunpeng 920 processor Base frequency: 2.6 GHz	High I/O General Purpose SSD Ultra-high I/O Extreme SSD General Purpose SSD V2	Ultra-high packets per second (PPS) throughput An ECS with higher specifications has better network performance. Maximum PPS: 7,000,000 Maximum intranet bandwidth: 38 Gbit/s

Kunpeng Enhanced AI Inference-accelerated kAi1s (Type I)

Overview

Kunpeng AI inference-accelerated kAi1s ECSs use Ascend 310 processors for AI acceleration. Ascend 310 processors feature low power consumption, high computing capabilities, and significantly improved energy efficiency ratio (EER), facilitating the wide application of AI inference. kAi1s ECSs deliver the computing acceleration capabilities of the Ascend 310 processors on the cloud platform. This helps you quickly and simply use the Ascend 310 processors.

kAi1s ECSs are based on Atlas 300I accelerator cards. For details, see Ascend Community.

kAi1s ECSs are used for general technologies, such as computer vision, speech recognition, and natural language processing to support smart retail, smart campus, robot cloud brain, and safe city scenarios.

Specifications

**Table 2** kAi1s ECS specifications
Flavor	vCPUs	Memory (GiB)	Max./Assured Bandwidth (Gbit/s)	Max. PPS (10,000)	Max. NIC Queues	Max. NICs	Ascend 310 Processors	Virtualization
kai1s.xlarge.1	4	4	3/0.8	20	2	2	1	KVM
kai1s.2xlarge.1	8	8	4/1.5	40	2	3	2	KVM
kai1s.4xlarge.1	16	16	6/3	80	4	4	4	KVM
kai1s.3xlarge.2	12	24	8/4	100	4	4	4	KVM
kai1s.4xlarge.2	16	32	10/6	140	4	5	6	KVM
kai1s.6xlarge.2	24	48	12/8	200	8	6	8	KVM
kai1s.9xlarge.2	36	72	12/8	200	8	6	12	KVM
kai1s.12xlarge.2	48	96	12/8	200	16	6	12	KVM

Features

kAi1s ECSs have the following features:

1:1 or 1:2 ratio of vCPUs to memory
CPU: Kunpeng 920 (2.6 GHz)
Ascend 310 processors, four of which in an Atlas 300I accelerator card
8 TeraFLOPS of half-precision computing (FP16) on one processor
16 TeraOPS of integer-precision computing (INT8) on one processor
8 GiB of GPU memory with a memory bandwidth of 50 GiB/s on one processor
Built-in hardware video codec engine, supporting H.264/H.265

Notes

kAi1s ECSs support the following OSs:
- Ubuntu Server 18.04 64bit
- CentOS 7.6 64-bit
kAi1s ECSs do not support modification of specifications.
kAi1s ECSs support automatic recovery when the hosts accommodating such ECSs become faulty.

Kunpeng Enhanced AI Inference-accelerated kAi2 (Type II)

Overview

Kunpeng enhanced AI inference-accelerated kAi2 ECSs use Ascend 310P processors for AI acceleration. The Atlas 300I Pro accelerator card with the 24 GiB of GPU memory and Ascend 310P processor helps boost your AI inference services.

Specifications

**Table 3** kAi2 ECS specifications
Flavor	vCPUs	Memory (GiB)	Max./Assured Bandwidth (Gbit/s)	Max. PPS (10,000)	Max. NIC Queues	Max. NICs	Ascend 310P	GPU Memory (GiB)	Virtualization
kai2.4xlarge.4	16	64	15/5.5	200	8	8	1	1 × 24	KVM
kai2.8xlarge.4	32	128	24/8	250	8	8	2	2 × 24	KVM
kai2.16xlarge.4	64	256	30/16	350	16	8	4	4 × 24	KVM
kai2.20xlarge.4	80	320	38/30	700	32	16	5	5 × 24	KVM
kai2.24xlarge.4	96	384	38/30	700	32	16	5	5 × 24	KVM

Features

kAi2 ECSs have the following features:

1:4 ratio of vCPUs to memory
CPU: Kunpeng 920 (2.6 GHz)
Ascend 310P processors, one of which in an Atlas 300I Pro accelerator card Each processor with eight DaVinci AI cores and eight self-developed CPU cores
70 TeraFLOPS of half-precision computing (FP16) on one processor
140 TeraOPS of integer-precision computing (INT8) on one processor
24 GiB of GPU memory with a memory bandwidth of 240.8 GiB/s on one processor
Built-in hardware video codec engine, supporting H.264/H.265
JPEG image codec

Notes

kAi2 ECSs support the following OSs:
- HCE 2.0 64bit
- CentOS 7.6 64bit
- Ubuntu 18.04 server 64bit
kAi2 ECSs support automatic recovery when the hosts accommodating such ECSs become faulty.
kAi2 ECSs support specifications modification but do not support the instance type change.

Using a kAi ECS

Perform the following steps:

Create an ECS. For details, see "Step 1: Configure Basic Settings" in the Elastic Cloud Server User Guide.
- In the Specifications field, select kAi-accelerated specifications.
- In the Image field, select Public image or Private image.
  - Public image: The CANN 3.1.0 development kit has been included and environment variables have been configured in public images by default. You need to verify the environment availability.
  - Private image: You need to install the driver, firmware, and development kit, and configure environment variables by yourself. For details, see the CANN Software Installation Guide of the corresponding version in Ascend Documentation.
Remotely log in to the ECS.
If your ECS runs Linux, use an SSH password to log in to the ECS.
Verify the environment availability.
Use a sample for compilation and running. For details, see "Sample Overview" in the Model Development Learning Map of the corresponding CANN edition in Ascend Documentation.

The sample shows how to classify images (decode, resize, and infer images) based on the Caffe ResNet-50 network.