PAI introduces the new General Unit (GU) series instance types, starting with the GU30 series. This series is suitable for training and inference of various models, such as AIGC text-to-image generation, large language models, multimodal models, NLP, CV, and ASR. Compared to traditional PAI instance types, GU series instances with similar performance are nearly 45% cheaper.
Pricing and scenarios for GU series instance types
The following table lists the GU series instance types supported by PAI, along with their pricing and scenarios.
The GPU card for GU30 series instance types has 24 GB of VRAM, 31 TFLOPS of FP32 computing power, and supports a PCIe 4.0 connection between the GPU and CPU.
The performance of GU30 series instance types is comparable to the NVIDIA A10. They are also more than 20% cheaper than the A10 instance types on PAI.
Name | Instance type | Specifications | Scenarios |
GU30 single-card dense | ml.gu7i.c8m30.1-gu30 | 8 vCPU + 30 GB RAM + 1 × 24 GB VRAM | For models such as image classification and detection, and AIGC image generation. |
GU30 single-card balanced | ml.gu7i.c16m60.1-gu30 | 16 vCPU + 60 GB RAM + 1 × 24 GB VRAM | For models such as ASR and OCR. |
GU30 single-card sparse | ml.gu7i.c32m188.1-gu30 | 32 vCPU + 188 GB RAM + 1 × 24 GB VRAM | For models such as custom search and recommendation search. |
GU30 dual-card | ml.gu7i.c64m376.2-gu30 | 64 vCPU + 376 GB RAM + 2 × 24 GB VRAM | For AIGC LLM models with 13.5B parameters. |
GU30 quad-card | ml.gu7i.c128m752.4-gu30 | 128 vCPU + 752 GB RAM + 4 × 24 GB VRAM | For AIGC LLM models with 30B parameters. |
How PAI GU series instance types achieve high cost-effectiveness
PAI has ultra-large-scale clusters for AI training and inference. Using advanced elastic capabilities such as auto scaling and elastic resource pools, PAI reduces the cost of individual physical resources and achieves the same performance with fewer resources.
PAI's AI optimization and acceleration improve the overall utilization of heterogeneous resource clusters, resulting in higher performance for AI training and inference using the same physical resources.
How to purchase PAI GU series instance types
PAI GU series instance types support two billing methods: subscription and pay-as-you-go.
Subscription (upfront payment)
You can go to the EAS Dedicated Subscription purchase page to view pricing information and purchase GU30 series instance types on a subscription basis. For more information, see Use EAS resource groups.
Pay-as-you-go
You can go to the EAS Dedicated Pay-As-You-Go purchase page to view pricing information and purchase GU30 series instance types on a pay-as-you-go basis. Billing starts after the instance is created, even if you do not use it to deploy a service. For more information, see Use EAS resource groups.
You can also directly use GU30 series instance types from public resources when you deploy a service, without making a separate purchase. After the service is deployed, you are billed on a pay-as-you-go basis. For information about the billing of GU series instance types in a public resource group, see Elastic Algorithm Service (EAS) billing.