Home › Products › G494-ZB4-AAP2
GIGABYTE · 4U GPU server · AMD EPYC
4U GPU server with dual AMD EPYC processors and up to eight dual-slot PCIe Gen5 GPUs. Validated with NVIDIA H200 NVL, NVIDIA RTX PRO 6000 Blackwell and AMD Instinct MI350P to train and serve AI models at scale.
Up to 8 dual-slot GPUs
2 x AMD EPYC 9005/9004
48 DDR5 DIMMs
12 NVMe Gen5 bays
4 x 3000 W Titanium PSUs

Applications
The G494-ZB4-AAP2 brings together eight GPUs, two AMD EPYC processors and 48 memory modules in a single server. It is the foundation for training models, serving generative AI to many users or running GPU-accelerated HPC.
Eight PCIe Gen5 GPUs, including NVIDIA H200 NVL in two sets of four with NVLink bridge.
Generative AI and language models serving many users at once, with up to eight GPUs per server.
Two AMD EPYC processors, 48 DDR5 modules and 12 NVMe Gen5 bays for scientific computing and large-scale data analysis.
Rendering and digital twins with up to eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.
Engineering highlights

Acceleration
Eight FHFL PCIe Gen5 x16 slots connected through two Broadcom PEX89144 switches. Supports GPUs up to 600 W with the optional fan kit.
Validated GPUs
8 x NVIDIA H200 NVL, 8 x NVIDIA RTX PRO 6000 Blackwell Server Edition or 8 x AMD Instinct MI350P, at 25 °C ambient.
Processors
9005 and 9004 series, SP5 socket, cTDP up to 400 W (500 W at 30 °C ambient).
Memory
12 channels per CPU. Up to 5200 MT/s with EPYC 9005 and 4800 MT/s with EPYC 9004 (1 module per channel).
Expansion
In addition to the GPUs: 3 FHHL and 3 low-profile PCIe Gen5 x16 slots for high-speed networking and DPUs.
Power
Four redundant power supplies with 80 PLUS Titanium certification (C19 cable).
Datasheet

Front

Rear
Form factor
4U rack · 448 x 176 x 880 mm (W x H x D) · 1000.5 mm deep with the optional fan kit
Motherboard
MZB3-G41
Processor
2 x AMD EPYC 9005 / 9004 series · SP5 socket (LGA 6096) · cTDP up to 400 W (500 W at 30 °C ambient)
Memory
48 DDR5 RDIMM slots · 12 channels per CPU · up to 5200 MT/s (EPYC 9005) and 4800 MT/s (EPYC 9004) with 1 module per channel
GPU
Up to 8 dual-slot PCIe Gen5 GPUs · up to 600 W each with the optional fan kit
Validated GPUs
8 x NVIDIA H200 NVL (2 sets of 4 with NVLink bridge) · 8 x NVIDIA RTX PRO 6000 Blackwell Server Edition · 8 x AMD Instinct MI350P
PCIe switches
2 x Broadcom PEX89144
Front storage
12 x 2.5" Gen5 NVMe/SATA/SAS-4 hot-swap
Internal storage
2 x M.2 (2280/22110) PCIe Gen3 x4
PCIe slots
8 x FHFL PCIe Gen5 x16 for GPU · 3 x FHHL PCIe Gen5 x16 (2 rear and 1 front) · 3 x LP PCIe Gen5 x16
Networking
2 x 1 Gb/s (Intel I350-AM2, front) · 1 dedicated management port
Management
ASPEED AST2600 BMC · VGA output
Security
TPM 2.0 header (optional module)
Power supply
4 x 3000 W 80 PLUS Titanium redundant · C19 cable
Cooling
12 x 60 x 60 x 56 mm fans and 2 x 40 x 40 x 28 mm fans
Operating temperature
10 °C to 35 °C
Weight
66.8 kg (net)
Product code
6NG494ZB4DR000ACP2
Specifications per the manufacturer. Validated GPUs require 25 °C ambient, and 600 W GPUs require the optional fan kit. The server is validated with a single GPU model at a time. GPUs sold separately. With a single processor, some memory and slots are unavailable.
For channels and resellers

Strategic partnership with Giga Computing and access to Gigabyte server platforms.
Technical support to choose GPU, processor, memory, networking, power and cooling based on the workload.
Stock availability and invoicing through the distributor to speed up the reseller's project closing.
Project quote
Describe your scenario and a technical specialist will get back to you with the recommended configuration.
Configuration based on workload: training, inference, HPC or visual computing
Service for resellers, integrators and businesses
Pre- and after-sales with a technical specialist