Comparative Analysis of AI Server Types and Guidelines for Selecting GPU Servers

Different types of AI servers are compared using a two‑dimensional chart that plots CPU, GPU, FPGA, TPU, and ASIC architectures, showing that performance improves from left to right.

The vertical axis represents programmability/flexibility; ASIC offers the highest performance but the lowest flexibility, while CPU provides the greatest flexibility with the lowest performance.

Overall, GPU flexibility is lower than CPU but its performance is higher, followed by FPGA, TPU, and finally ASIC.

When selecting a server, factors such as power consumption, cost, performance, and real‑time requirements must be considered; ASIC is suitable for fixed, simple algorithms, whereas GPU is preferred for training or general workloads.

Basic principles for choosing GPU servers are introduced, starting with common GPU classifications by bus interface: NV‑Link, traditional bus, and PCI‑e.

NV‑Link GPUs, exemplified by NVIDIA V100 with SXM2/SXM3 interfaces, are used in DGX supercomputers and partner‑designed NV‑Link servers.

Traditional bus GPUs include PCI‑e models like V100, P40 (Pascal), P4, and the newer T4; the slim P4 and T4 are typically used for inference.

Traditional PCI‑e GPU servers are divided into OEM (e.g., Sugon, Inspur, Huawei) and non‑OEM offerings.

Selection criteria also cover precision, memory type and capacity, power, cooling, noise, temperature, and mobility requirements.

Choosing a GPU model should start from business needs: HPC tasks requiring double precision may need V100 or P100, while applications like oil exploration demand larger video memory; bus standards may also dictate the choice.

GPU servers are widely used in AI; in teaching scenarios, GPU virtualization is important, often requiring 30‑60 virtual GPUs, with V100 for training and P4/T4 for inference.

After fixing the GPU model, the type of server must be considered: edge servers may use T4 or P4, while central inference may need V100, taking throughput and usage scenarios into account.

Client IT operations capability influences the decision: large enterprises with strong ops may choose generic PCI‑e servers, whereas smaller teams or data scientists may have different priorities.

Additional factors include the value of accompanying software and services.

The maturity and engineering efficiency of the GPU cluster system matter; integrated solutions like DGX provide a fully optimized stack from OS and drivers to Docker, enhancing productivity.

Promotional Offer: "Full Store Technical Materials Pack (All)" e‑book is on a special New Year discount, originally valued at 240 CNY, now 168 CNY (valid 2019‑12‑14 to 2020‑01‑01). The pack includes 32 titles covering topics such as RDMA principles, data backup, container architecture, flash technology, virtualization, storage, IO performance tuning, Ceph, Kubernetes, DPDK/SPDK, InfiniBand, cloud PaaS, OpenStack, NVMe, and more.