Benefits

Decoupling

In a conventional GPU-accelerated instance, the CPU, memory, and GPU are deployed on the same physical server. EAIS instances decouple the CPU from the GPU. The CPU, memory, and GPU can exist on different physical servers. You can select an Elastic Compute Service (ECS) instance based on your CPU and memory needs and then attach an EAIS instance. This creates a new GPU-accelerated instance type that meets your requirements.

Low cost

EAIS instances can reduce model inference costs by up to 50%. You can specify the exact amount of inference acceleration you need. This avoids over-provisioning GPU resources. You can select the instance type that best fits your application.

For example, consider a scenario where you need an instance with more than 128 GiB of memory and a single GPU:

• Using a conventional GPU-accelerated instance:

  Among conventional GPU-accelerated instance types, the following options meet your requirements:

  Instance Type	vCPU	Memory (GiB)	GPU	GPU Memory (GB)
  ecs.gn6i-c24g1.12xlarge	48	186.0	T4*2	32
  ecs.gn6v-c8g1.8xlarge	32	128.0	V100*4	64
  ecs.gn6e-c12g1.12xlarge	48	368.0	V100*4	128

  As the table shows, the GPU-accelerated instance you purchase may come with multiple GPUs. This results in wasted, over-provisioned GPU resources.

• Using an EAIS instance:

  You only need to purchase the following compute resources:

  Product	Instance Type	Metric data
  Elastic Compute Service	ecs.r6.6xlarge	24 vCPUs, 192 GiB
  Elastic Acceleration Instance (EAIS)	eais.ei-a6.4xlarge	16 TFLOPS (FP32), 32 GB GPU memory

In summary, if you purchase a GPU-accelerated instance, you must choose from fixed instance types and pay for all its resources. With EAIS, you can select an ECS instance without a GPU that meets your memory needs and then attach an EAIS instance with the required computing power and GPU memory. This flexible approach solves the same problem and offers a clear cost advantage.

Elasticity

EAIS instances allow you to obtain the exact resources you need by flexibly matching GPU resources. You can easily scale the amount of inference acceleration up or down to meet your business needs. This prevents over-investment in provisioned resources:

• When you need to add ECS instances to meet growing demand, you can attach an EAIS instance to each one.

• When demand decreases, you can release the EAIS instance attached to any ECS instance at any time. When needed, you can create and attach a new EAIS instance to the ECS instance. This lets you flexibly match and use GPU resources and avoid charges for idle resources.

Versatility

EAIS instances are highly versatile. They support various heterogeneous hardware types, such as GPUs, NPUs, and FPGAs, and offer a wide range of compatible options.

High performance

EAIS instances provide acceleration for model inference. Compared to conventional GPU-accelerated instances, an EAIS instance with the same computing power for inference delivers higher performance.