Build ID: 716300

Author: JT2026-6-12 - 09:35:25

Custom Label (SKU): 716300-1U-GB-R162-Z13-10BSFF-2PS

Hardware Highlights

This Server is a Gigabyte R162-Z13 a variant of the R162-Z10 model 

Similar to Dell PowerEdge R6515, Supermicro AS -1114S-WN10RT, Lenovo SR635

It uses a single High End AMD Epyc 7642 CPU at 48 Cores 

8 slots for DDR4 RAM Dimms  

10x U.2 NVME Drive Bays

Onboard Dual 1GB Ethernet ports

Dual 25GB SFP28 via OCP 2.0 card

* This server does not come with PCI-E Expansion risers 

* This Server does not come with outer rails

Configuration:

CPU: 1x AMD Epyc 7642 48 Cores

Memory: 64GB DDR4 RAM KIT 

Hard Drives:

- 1x NEW MZ-QLB9600 - 960GB Gen3 U.2 NVME Drive Like New

Controllers:

- 1x OCP 2.0 card for Dual 25GB SFP28 Ports intalled

OB NIC: Dual 1GB Ethernet ports Onboard via Intel i350-AM2

Secondary Chassis/ Motherboard specs:

Server Chassis/Case: 1U-GB-R162-Z13-10BSFF-2PS

Motherboard: MZ12-HD2

Backplane: 10x U..2 NVME

Chassis PCI slots:

Motherboard PCI slots: Slot

Front Bays: 10x SFF 2.5"

Caddies Included: 2x 2.5" Gigabyte Style Caddies

Rear Bays: None

PSU Slots: Dual PSU Slots

Power: 2x 1200W Power Supplies

Mounting: Inner Rails Only, no Outer Rails included

Warranty: Standard

Warranty Includes 30 Days Standard Warranty covers hardware for the duration stated Expert Pre-sales consulting. With our expertise we will help you find the best solutions for your needs.Expert Post sales Technical Support and consulting - We Set you up for Success

Standard Warranty and Terms of Sale

- Standard 30 day limited warranty. - Item tested to power on and all specs show in BIOS, We Warranty Hardware Only, not Software. - Processing times estimated 3 Days. Tracking number will be emailed after Shipped. - 30 Day Hardware parts Warranty. problems must be reported 30 days of when the item is delivered. - Buyer Pays for all shipping, even on returns. (shipping is non-refundable) - Restocking fee as described in the terms of the listing/ order. 20% Restocking fee if not stated. - If you have issues please message us with an item number and the issue you are having. - Our customer service representatives will assist you on getting your problem solved

Message Us

- Please have eBay item number ready when contacting, we will try to answer the questions within 24 hours.

SERVER USE CASE:

Cloud Computing

Server Hosting 

Edge HPC

Storage heavy applications

AI Inference via CPU power

Using an AMD EPYC 7642 (48 cores, 96 threads) for AI inference involves leveraging its massive memory bandwidth and high core count with optimized software like llama.cpp or vLLM to achieve efficient CPU-based throughput. For optimal performance, set the BIOS NPS (NUMA Nodes Per Socket) to 1, ensure high-speed DDR4 memory (3200MHz) is fully populated across all 8 channels, and use quantization (

-bit) to fit models within available system RAM. 

Key Setup and Optimization Steps

BIOS Configuration: Set NPS (NUMA Nodes Per Socket) to 1 for better performance in most inference tasks, as suggested in.

Memory Optimization: Ensure you have 8 channels of DDR4 3200MHz memory installed for maximum bandwidth.

Software Stack for CPU Inference:

llama.cpp: Best for running GGUF-quantized LLMs (like Llama 3, Mistral) on CPUs.

vLLM: Optimized for throughput and batch processing.

AMD AOCL (AMD Optimizing CPU Libraries): Use these libraries to speed up math operations.

Model Quantization: Use 4-bit, 3-bit, or 2-bit quantization (GGUF format) to fit large models within system RAM and improve token generation speed (tokens/second). 

Scenarios for 7642 Inference

Large Language Models (LLMs): The 7642 can run large models (e.g., 70B parameters) because of the large available RAM (up to 4TB+), although at lower latency than GPUs.

Batch Inference: Due to high core counts, the 7642 is efficient at processing multiple inference requests simultaneously.

GPU Backup: In servers with GPUs, the EPYC 7642 acts as a powerful host CPU and a fallback if GPUs fail,

Performance Considerations

Speed: While not as fast as GPU inference, an EPYC 7642 provides decent performance, particularly when using heavily optimized libraries.

Latency: Token latency is higher on CPU than GPU, but acceptable for many back-end or offline tasks, as reported by.