AI servers

ПАК SOPHGO YU08-01 — ИИ-сервер для агентов и больших моделей
New
Price upon request

SOPHGO YU08-01 AI Server — All-in-One HW/SW Complex for LLMs and AI Agents (4× TPU, 1024GB)

5
Pre-order
RISC-V сервер SOPHGO SRA1-20
New
Price upon request

RISC-V server SOPHGO SRA1-20 - 2× SG2042 (128 RISC-V cores, 256GB DDR4, 2U)

5
Pre-order
RISC-V сервер SOPHGO SRA3-40
New
Price upon request

RISC-V server SOPHGO SRA3-40 — SG2044 with TPU (64 RISC-V cores, 128GB LPDDR5X, 4U)

5
Pre-order
RISC-V сервер хранения SOPHGO SRB1-20
New
Price upon request

SOPHGO SRB1-20 RISC-V storage server — 2× SG2042 (256GB DDR4, 8× 12TB HDD, 2U)

5
Pre-order
RISC-V сервер хранения SOPHGO SRB3-40
New
Price upon request

SOPHGO SRB3-40 RISC-V Storage Server — SG2044 with TPU (128GB LPDDR5X, 12-bay, 4U)

5
Pre-order
RISC-V сервер SOPHGO SRM1-20
New
Price upon request

RISC-V server SOPHGO SRM1-20 — 2× SG2042 with TPU acceleration (256GB DDR4, 2U)

5
Pre-order
RISC-V кластер SOPHGO SRC1-10 42U
New
Price upon request

SOPHGO SRC1-10 RISC-V Cluster — 42U, 16× SG2042 (1024 cores, 2TB DDR4, 192TB HDD)

5
Pre-order
Edge AI-сервер SOPHGO SE7 32-EA4-23
New
Price upon request

SOPHGO SE7 32-EA4-23 Edge AI Server — BM1684X (32 TOPS INT8, 16GB, 32 video channels)

5
Pre-order
Edge AI-сервер SOPHGO SE7 32-BP1-11
New
Price upon request

SOPHGO SE7 Edge AI Server 32-BP1-11 — BM1684X (32 TOPS INT8, 16GB, industrial, IP40)

5
Pre-order
Edge AI-сервер SOPHGO SE7 32-EP1-21
New
Price upon request

SOPHGO SE7 Edge AI Server 32-EP1-21 — BM1684X (32 TOPS INT8, 16GB, −40…+60 °C, IP40)

5
Pre-order
Edge AI-сервер SOPHGO SE8 288-EA6-72
New
Price upon request

SOPHGO SE8 Edge AI Server 288-EA6-72 — 9x BM1684X (288 video channels, x86 8C/16T, 144GB)

5
Pre-order
Edge AI-сервер SOPHGO SE9 16-BP1-11
New
Price upon request

SOPHGO SE9 16-BP1-11 Edge AI Server — BM1688 (16 video channels, 8GB, 15W, IP41)

5
Pre-order

AI servers: what hardware is needed for neural networks

The division brings together equipment for artificial intelligence—from camera modules to servers for eight accelerators:

  • GPU platforms and GPU servers — training and inference models on NVIDIA A100, H100, H200, B200, B300 and AMD Instinct GPUs, as well as machines on 1–4 video cards and used servers;
  • AI supercomputers — desktop stations on NVIDIA GB10 with 128 GB of total memory and on GB300;
  • Edge servers and AI modules from SOPHGO — video analytics near the cameras when power consumption is 12.5 W;
  • RISC-V servers — calculations and data storage on open-architecture processors.

Start with the model size: it determines how much memory the accelerators will need. Then consider the software stack — CUDA from NVIDIA, ROCm from AMD — and the power your server can handle. If the required configuration isn’t in the catalog, The manager will choose the right one for the task..

Frequently asked questions

How are GPU platforms different from GPU servers in the catalog?

GPU platforms are equipped with servers with 6–8 accelerators from Supermicro, ASRock Rack, GIGABYTE, ASUS, HPE, and Dell. There are both machines with one or four video cards and tower configurations as well as used servers in the GPU servers.

How much accelerator memory is needed for LLM?

In the FP16 format, the model takes up about 2 GB per billion parameters; in 4-bit quantization, it takes about 0.5–0.6 GB; plus a buffer for context. Models with 70 billion parameters in FP16 require about 140 GB, that is, at least two accelerators of 80–96 GB each.

Which server should I use for training, and which for inference?

For training large models, servers with 8 SXM accelerators with NVLink connectivity—H100, H200, B200, or B300—are used. For inference and pre-training, 2–8 PCIe cards are often sufficient, such as the RTX PRO 6000 with 96 GB of memory.