Name: Perspic GPU Servers
Brand: Perspic AI

Inference Optimized

H9508

5U 8-GPU PCIe 5.0 full-mesh P2P.

8x GPU

PCIe 5.0

Form

76 GB/s

P2P

6-7 kW

Power

Max Density

H9816

8U 16-GPU dual-node, 2.26 TB VRAM.

16x GPU

Dual

Form

2.26 TB

VRAM

12 kW

Power

Training

H8236

8U 8-GPU NVLink 5th Gen training.

8x GPU

SXM

Form

900 GB/s

NVLink

10 kW

Power

The Challenge

Enterprise AI demands more GPU compute than ever. Traditional solutions cost too much.

Organizations deploying large language models, computer vision pipelines, and AI video generation face a critical bottleneck: OEM GPU servers from major vendors carry massive premiums. NVSwitch licensing, proprietary interconnects, and vendor lock-in inflate total cost of ownership by 40% or more. Meanwhile, model sizes continue to grow -- the latest open-source models exceed 671 billion parameters, requiring multi-terabyte VRAM configurations that few platforms can deliver affordably.

What if you could get DGX-class performance at 40% lower cost, in a 5U form factor?

The Solution

PCIe 5.0 full-mesh P2P eliminates the NVSwitch tax.

Perspic Hardware (Jowox) GPU servers use a proprietary PCIe 5.0 full-mesh peer-to-peer topology that delivers 76 GB/s per GPU link -- without NVSwitch. Every GPU communicates directly with every other GPU, enabling optimal tensor and pipeline parallelism for large model inference and training. The result is enterprise-grade performance at a fraction of the cost.

⍔

Universal GPU Support

Any PCIe 5.0 GPU. No vendor lock-in.

⇄

Full-Mesh Topology

76 GB/s per link. Optimal parallelism.

⇧

2x Gen4 Bandwidth

64 GT/s per lane for AI workloads.

The Proof

2,073 tokens per second. Measured, not estimated.

KEY

2,073

tok/s at 70B @ 300

COMPACT

40% smaller

SAVINGS

40%

Cost reduction

STABLE

<85°C

Zero throttle

DeepSeek-R1-Distill-Llama-70B | 8x RTX 5090 (32GB) | sglang + evalscope

The Portfolio

Three platforms. Every AI workload covered.

From inference-optimized 5U servers to 16-GPU density monsters and NVLink-equipped training rigs, the Jowox lineup covers the full spectrum of enterprise AI requirements. Each platform supports flexible GPU configurations through standard PCIe 5.0 or optional SXM/OAM baseboards.

2,073 tok/s

H9508

5U PCIe 5.0 Inference

Key Advantages

GPUDirect RDMA support
Full-mesh PCIe 5.0 P2P (76 GB/s)
37% cost reduction vs DGX

Select GPU

Slots

1.1 TB

VRAM

600W each

TDP

~6-8 kW

Power

LLM InferenceComfyUIAI Video

671B+

H9816

8U 16-GPU Max Density

Key Advantages

2.26 TB total VRAM
No NVSwitch required
Unified liquid cooling

Select GPU

Slots

2.26 TB

VRAM

600W each

TDP

~15-20 kW

Power

671B ModelsDigital TwinsHPC

18 PFLOPS

H8236

8U SXM/OAM Training

Key Advantages

NVLink 5th Gen (900 GB/s)
700W+ TDP per GPU
Direct liquid cooling

Select Module

8x OAM/SXM

Modules

1.1 TB

VRAM

NVLink 900 GB/s

Interconnect

~8 kW

Power

~32 PFLOPS

FP8 compute (H200 SXM)

Pre-trainingFine-tuningResearch

The Comparison

Shorter cost bar = better value. We lead.

Vendor / Model	Form	Memory	Interconnect	Cost
Perspic H9508	5U/8x	1.13TB	PCIe 5.0 P2P	Best Value
Supermicro SYS-821GE	8U/8x	640GB	NVLink 4.0	OEM Premium
Inspur NVSwitch-based	6U/8x	640GB	NVSwitch	OEM Premium
HPE ProLiant DL380a Gen11	2U/4-8x	320-640	PCIe Gen5	OEM Premium
Lenovo ThinkSystem SR685a V3	8U/8x	640GB	PCIe+NVLink	OEM Premium

* Comparison based on selected specifications from publicly available datasheets and vendor documentation as of Q1 2026. Actual performance, pricing, and availability may vary by configuration, region, and vendor terms.

* "Best Value" reflects Perspic Hardware's estimated TCO advantage based on MSRP-equivalent pricing for comparable GPU configurations.

All product names, logos, and brands are property of their respective owners. NVIDIA, NVLink, NVSwitch, and DGX are trademarks of NVIDIA Corporation. Supermicro is a trademark of Super Micro Computer, Inc. HPE and ProLiant are trademarks of Hewlett Packard Enterprise. Lenovo and ThinkSystem are trademarks of Lenovo Group. Use of these names is for identification and comparison purposes only and does not imply endorsement or affiliation.

The Configurator

Not sure which server? Let the data decide.

Our GPU Server Estimator analyzes your workload requirements -- model size, concurrent users, precision needs -- and recommends the optimal platform and GPU configuration. No guesswork, no over-provisioning.

LLM Inference Chatbots, APIs

AI Image ComfyUI, SD

AI Video Sora, Runway

Training Fine-tuning

3D / Sim Digital Twins

HPC Scientific

Model Size70B

7B13B70B180B671B+

Concurrent Users100

10501005001000+

Recommended Configuration

H9508

8x H100 PCIe · H100 PCIe (80GB HBM3) · Total VRAM: 640 GB

Why this configuration? Optimal for Llama 3 70B inference at 2,073 tokens/s. Full-mesh P2P topology with 76 GB/s bandwidth.

*Theoretical analysis. Actual performance varies by model and configuration.

The Edge

Four pillars of the Perspic Hardware advantage.

5U Form Factor

40% smaller than industry-standard 8U. More compute density per rack, lower colocation costs, better thermal management.

5U · -40%

40% Cost Savings

No NVSwitch licensing fees, no proprietary connectors, no vendor lock-in premiums. Direct savings on total cost of ownership.

-40% TCO

Perspic Software Stack

Ships with Perspic OS + Ohm Studio pre-installed. One-command deployment for containers, inference, and training workloads.

1-Click

Full Compliance

RoHS certified, CE marked, 80+ Titanium PSU efficiency. Enterprise procurement-ready from day one.

80+ Ti

Deep Dive

The H8236: Where PCIe meets NVLink.

The H8236 is uniquely versatile: it natively supports PCIe GPUs, but with an optional OAM/SXM baseboard upgrade, it unlocks NVLink 5th Gen at 900 GB/s -- the same interconnect used in NVIDIA DGX systems. This makes it the ideal platform for organizations that need training performance today but want a flexible upgrade path.

PCIe NativeSXM Upgrade Path

8x GPU

Capacity

Form

1.1 TB

VRAM

PCIe Support

H200 NVL (141GB) ★600W

H100 NVL (94GB)400W

L40S (48GB)350W

OAM / SXM (with upgrade)

B200 SXM (192GB)1000W

H200 SXM (141GB)700W

H100 SXM (80GB)700W

OAM/SXM baseboard upgrade enables NVLink 900 GB/s connectivity.

Click to enlarge

H8236 Modular Architecture Diagram
GPU Module · Motherboard Module · 8U Chassis · Backplane Bracket

After-sale Support

Hardware Warranty & Service

Every GPU server comes with comprehensive warranty coverage. We stand behind our hardware.

4h

Remote Support Response

Phone, email, and remote access during business hours (Mon-Fri, 9:00-18:00).

10d

Return-to-Depot Repair

Confirmed hardware failures repaired within 10 business days. Return shipping covered by Perspic.

2d

On-Site Service (Optional)

On-site repair with 2 business day response. Available as add-on during warranty period.

Warranty Coverage

Motherboard • CPU • Memory (RAM) • HDD/SSD • Power Supply • RAID Controller • Network Cards • Cooling Fans • Chassis & Backplane

Service Region

Currently covering Mainland China, Hong Kong SAR, and Macau SAR.
Additional regions coming soon. Governed by the laws of Hong Kong SAR.

Warranty effective from date of delivery. Replacement parts carry 90-day warranty or remaining warranty period, whichever is longer. Full warranty terms provided with purchase documentation. Contact info@perspic.net for details.

AI-Ready. Rack-Ready.

Hardware Warranty & Service