TPU SystemsPrivate AI Infrastructure
Book a Consultation

AI Systems

On‑premises AI server platforms, sized to your firm

TPU Systems builds private AI deployments on GIGABYTE enterprise hardware — NVIDIA‑certified platforms spanning desk‑side AI appliances to 10‑GPU rack servers. We size, install, integrate and manage the platform; you keep every byte of data.

Sizing Guide

Match the platform to your team

"32 users" isn't a hardware spec. Concurrency, document volume and response-time targets drive the platform choice — this is where our assessment starts.

Your teamRecommended platformTypical role
8–12 users · pilot or departmentAI TOP ATOMEvaluation, secure desk‑side inference, fine‑tuning experiments, light document AI
15–40 users · whole firmG293‑Z42 or G294‑Z22Full private AI service: 24/7 inference, permission‑aware RAG across all documents
40+ users · multi‑model, heavy useG493‑ZB1Concurrent multi‑model serving, on‑premises fine‑tuning, growing organizations
GIGABYTE AI TOP ATOM compact desktop AI supercomputer with patterned front grille

GIGABYTE AI TOP ATOM

Personal AI Supercomputer

Empower advanced AI with GIGABYTE AI TOP ATOM — a personal AI supercomputer delivering petaflop‑scale performance in a compact, energy‑efficient desktop form factor. At its core, the NVIDIA GB10 Grace Blackwell Superchip with fifth‑generation Tensor Cores drives prototyping, fine‑tuning and inference right at your desk, while 128 GB of coherent unified memory runs today's most demanding open models locally.

AI performance
Up to 1 petaflop of FP4 computing power
Accelerator
NVIDIA GB10 Grace Blackwell Superchip · 20‑core Arm CPU · NVLink‑C2C
Memory
128 GB coherent unified system memory
Model support
Up to 200B parameters (up to 405B with two systems via ConnectX‑7)
Storage
Up to 4 TB NVMe
Software
NVIDIA AI software stack · GIGABYTE AI TOP Utility
Form factor
Compact desktop
Where it fits Pilot programs and departments of 8–12 users; secure desk‑side inference for executives; fine‑tuning workbench. Pair two ATOM systems over ConnectX‑7 to scale to larger models.
GIGABYTE G293-Z42-AAP1 2U GPU server with cover removed, showing GPU slots, DIMMs and front drive bays

GIGABYTE G293‑Z42‑AAP1

2U DP · 8× PCIe Gen5 GPUs

An NVIDIA‑Certified 2U HPC/AI server built for dense GPU compute. Dual AMD EPYC 9005/9004 processors feed up to eight dual‑slot Gen5 GPUs through PCIe switches, with 12‑channel DDR5 memory and Titanium‑efficiency redundant power — the platform we specify most often for full‑firm private AI deployments of 15–40 users.

Certification
NVIDIA‑Certified Systems — data center servers
CPUs
Dual AMD EPYC 9005/9004 (Socket SP5), TDP up to 210 W
GPUs
Up to 8 × dual‑slot Gen5 GPUs (8 FHFL PCIe Gen5 x16 slots) + 2 LP slots
Memory
24 × DIMM, 12‑channel DDR5 RDIMM per CPU (up to 6400 MT/s with EPYC 9005)
Storage
4 × 2.5" Gen5 NVMe/SATA/SAS hot‑swap + 4 × 2.5" SATA/SAS
Network
2 × 10 Gb/s LAN (Broadcom BCM57416) + dedicated management LAN
Power
Dual 3000 W 80 PLUS Titanium redundant PSUs
Management
ASPEED AST2600 BMC
Where it fits The flagship platform behind our typical US$150,000, 32‑user deployment: sustained concurrent inference, permission‑aware RAG across large document sets, and room to grow GPU capacity without replacing the server.
GIGABYTE G294-Z22-AAP2 2U GPU server with exposed internals and front drive caddies

GIGABYTE G294‑Z22‑AAP2

2U UP · 8× PCIe Gen5 GPUs

A cost‑efficient single‑socket alternative with the same eight dual‑slot Gen5 GPU slots, driven by one AMD EPYC 9005/9004 processor and 12‑channel DDR5. Flexible accelerator options let you balance budget and capability — from NVIDIA H200 NVL pairs bridged with NVLink to workstation‑class Blackwell or AMD Radeon AI PRO cards.

CPUs
Single AMD EPYC 9005/9004, 12‑channel DDR5, 12 × DIMM
GPU options
4 × NVIDIA H200 NVL (2‑way NVLink bridges) · 4 × NVIDIA RTX PRO 6000 Blackwell Server Edition · 8 × AMD Radeon AI PRO R9700S
Expansion
8 × FHFL PCIe Gen5 x16 for GPUs + 2 × LP Gen5 x16
Storage
4 × 2.5" Gen5 NVMe/SATA/SAS‑4 + 4 × 2.5" SATA/SAS‑4 + 2 × M.2 Gen5
Network
2 × 10 Gb/s LAN (Broadcom BCM57416)
Power
Dual 3000 W 80 PLUS Titanium redundant PSUs
Where it fits Mid‑size firms that want full private AI capability at a lower entry point, or organizations standardizing on a specific accelerator family.
GIGABYTE G493-ZB1-AAP1 4U GPU server with rack handles, drive bays and internal cooling partition

GIGABYTE G493‑ZB1‑AAP1

4U DP · 10× PCIe Gen5 GPUs

When one model isn't enough. This NVIDIA‑Certified 4U platform carries ten dual‑slot Gen5 GPU slots and 48 DIMMs of 12‑channel DDR5 across dual AMD EPYC 9004 processors, with 3+1 redundant 3000 W Titanium power supplies. Built for organizations serving multiple models simultaneously, fine‑tuning on their own data, or pushing heavy concurrent inference across 40+ users.

Certification
NVIDIA‑Certified Systems — data center servers
CPUs
Dual AMD EPYC 9004 (Socket SP5)
GPUs
Up to 10 × dual‑slot Gen5 GPUs (10 FHFL Gen5 x16 slots) + 2 front LP slots
Memory
48 × DIMM, 12‑channel DDR5 RDIMM
Storage
8 × 2.5" Gen5 NVMe/SATA/SAS‑4 + 4 × 2.5" SATA/SAS‑4 + 2 × M.2
Network
2 × 1 Gb/s LAN (Intel I350‑AM2) + management LAN
Power
3+1 3000 W 80 PLUS Titanium redundant PSUs
Where it fits Larger professional firms, multi‑tenant deployments and organizations with on‑premises fine‑tuning programs or high‑volume document processing.

TPU Systems is an independent integrator. Server platforms are designed and manufactured by GIGABYTE Enterprise Solutions; product descriptions, specifications and photography courtesy of GIGABYTE. Platform availability and configuration may change — every quote is confirmed at time of order.

Not sure which platform?

We'll size it for you — for free.

Tell us your user count, document volume and workflows. We'll recommend a platform, estimate your business case and propose a paid pilot that proves it.

Book a Sizing Consultation Call 647‑827‑0502 x 880