UNICONET TECHNOLOGY LIMITED · IT HARDWARE REPAIR SERVICES
GPU Compute Cards · GPU Servers
Chip-Level Repair Specialists
Chip-Level Repair Specialists
When an H100 goes down, you lose not the hardware itself but hourly-billed compute, training progress and delivery milestones.
With chip-level repair capability, UnicoNet returns faulty compute cards and AI servers to the cluster through repair rather than scrapping the whole unit — and we do more than repair fast: at the very moment a fault occurs, we establish clearly whether the unit should be replaced and whether it can be repaired.
With chip-level repair capability, UnicoNet returns faulty compute cards and AI servers to the cluster through repair rather than scrapping the whole unit — and we do more than repair fast: at the very moment a fault occurs, we establish clearly whether the unit should be replaced and whether it can be repaired.
Supports NVIDIA H100 / H200 / A100 / A800 / H800 / L40S / A40 / RTX 4090·5090 / T4 · HGX 8-GPU baseboards · DGX and mainstream-brand AI servers (complete systems)
BGA chip level
HBM memory / GPU core reballing and resoldering
48h+
Shipped only after full-load stress testing
3–6 months
Repair warranty, repeat faults repaired free of charge
7×24
Fault response, with an optional on-site spare parts pool
Replacing a card and repairing a card differ in cost by an order of magnitude
High-end compute cards are subject to supply lead times and export controls: OEM repair channels are limited, and new-card lead times often run from weeks to months. In industry practice, board-level repair and full-card replacement typically differ in price by an order of magnitude.
The real challenge is never whether a part can be swapped, but judging whether it should be replaced and whether it can be repaired — a wrong call costs the customer money, and above all time.
UnicoNet's value lies in drawing that line clearly first: what can be repaired is repaired properly, and what cannot be repaired is reported honestly, so no downtime window is wasted.
CORE SERVICE 01
GPU Compute Cards · Chip-Level Repair
From SXM modules to PCIe accelerator cards, covering root-cause diagnosis and repair of six typical faults
1
Card drop-out / not detected
Typical symptoms: nvidia-smi cannot see the card, dmesg reports Xid 79, moving slots and rebooting make no difference
Root cause: SXM gold finger oxidation, power delivery or core BGA cold solder joints, PCIe Switch failure
Action: Thermal reproduction and localisation, then core / memory reballing, gold finger replacement and link re-test
2
Thermal throttling · collapsing compute
Typical symptoms: temperatures spike under full load, clock speeds halve, training time stretches abnormally
Root cause: dried-out and pumped-out thermal paste, dust build-up in the airflow path, blocked liquid-cooling cold plates, worn fan bearings
Action: Rebuild the thermal module, replace thermal paste / thermal pads, clean cold plates and pressure-test for leaks
3
ECC errors · training crashes
Typical symptoms: CUDA error: ECC error detected, abnormal loss, random process termination
Root cause: ageing HBM memory dies, out-of-spec memory supply voltage drift, exhausted Row-remapping resources
Action: Die-level memory diagnosis to pin down the bank, full HBM stack replacement, X-Ray re-inspection of solder joints
4
Short circuit on power-up · rack breaker trips
Typical symptoms: breaker trips as soon as the card is inserted, PSU protection trips, burning smell
Root cause: MOS transistor / filter capacitor breakdown, 12V short to ground, failed power-stage components
Action: Check impedance rail by rail, replace power-stage components, and power up only after the impedance re-test passes
5
Multi-card cluster communication failure
Typical symptoms: NVLink bandwidth drops, NCCL timeouts, multi-node training hangs
Root cause: cold solder joints and broken traces on NVLink / NVSwitch, degraded optical modules and high-speed cables
Action: Link-level diagnosis plus on-site swap-and-verify from our optical module / cable spare pool, so the GPU does not take the blame
6
Faults specific to liquid cooling
Typical symptoms: cold plate seepage, blocked flow channels, coolant corrosion of the PCB
Root cause: ageing hose seals, degraded coolant, stress cracking of quick connectors
Action: Drain the coolant, clean the cold plate, hold at 0.15MPa for 30 minutes to test for leaks, then repair coolant corrosion on the PCB
Our real capability: no heat gun improvisation
High-end compute cards demand extreme precision, and non-standard repair easily renders them scrap. The equipment and environment required for standard chip-level repair are set out below:
BGA rework station
Three-stage temperature-controlled desoldering and precise reballing, with X-ray re-inspection after full HBM stack replacement
3D X-Ray inspection
Reveals visually undetectable BGA voids, micro-cracks and solder bridging in a single pass
Infrared thermal imaging
Live thermal mapping to pinpoint shorts and abnormally hot components quickly
Oscilloscope + impedance testing
12V ripple, power-stage waveforms and rail-by-rail impedance to ground quantified, no guesswork
Constant temperature and humidity burn-in room
20℃ / 40–60% RH simulating data-centre conditions, with 24–72 hours of full-load burn-in
Full-load test platform
An industrial power environment supporting full-load power-up of complete systems, tested with real 3000W+ power supplies
Clean, ESD-safe work area
Full ESD-controlled handling, diagonal disassembly and assembly to prevent PCB warping, and oven dehumidification pre-treatment
PCB trace repair and jumper wiring
Micro-repair of broken, corroded and burnt traces, with reinforced insulation protection
Models we repair (partial list)
SXM / HGX modules
H100 · H200 · H800 · A100 · A800 · B200 / GB200 series (assessed subject to supply and parts availability)
PCIe accelerator / graphics cards
L40S · A40 · A100 PCIe · RTX 5090 / 4090 / 3090 · A6000 · T4 · V100
Domestic accelerator cards
Ascend 910 / 910B and other domestic compute cards (board-level repair assessed case by case by model)
Complete AI servers
HGX 8-GPU baseboards · DGX series · Inspur / H3C / Lenovo / Supermicro / Dell / Huawei and other branded systems
Note: enquiries about models not listed here are welcome; for models affected by export controls or parts supply, we will report repairability and alternatives honestly after diagnosis.
CORE SERVICE 02
AI GPU Servers · Complete System Repair
The failure profile of a GPU server is completely different from that of a traditional server: it is not about hard disks, but power, cooling and interconnect
Industry data: GPU server fault distribution
Compiled from publicly available industry data-centre operations statistics, for repair prioritisation reference only
Power subsystem faultsapprox. 35%
Ageing PSU modules, out-of-spec 12V ripple, unbalanced power phases, leading to system power loss or GPU throttling
Cooling and thermal control faultsapprox. 27%
Worn fan bearings, micro-leaks in liquid-cooling lines, dried-out thermal paste, leading to over-temperature throttling and interrupted training
PCIe / interconnect faultsapprox. 18%
Gold finger oxidation, link CRC errors, PCIe Re-Training, leading to single-card or multi-card communication failure
Memory / GPU core faultsapprox. 12%
Accumulated HBM / memory ECC errors, core micro-cracks, solder ageing, leading to inaccurate computation results
NVLink / NVSwitch and othersapprox. 8%
Loose bridges, rising bit error rates, NVSwitch ASIC faults, NVMe thermal throttling
Power delivery
3000W+ CRPS redundant power supply modules, power distribution boards (PDB), 12V ripple and three-phase balance testing, correction of unbalanced power phases
Cooling system
Fan wall module replacement, airflow path rebuild, liquid-cooling CDU and cold plate maintenance, line pressure-testing for leaks, standardised replacement of thermal interface materials
Motherboard and interconnect
Motherboards, PCIe backplanes, SXM baseboards, NVLink bridges, PCIe Switches, bent-pin correction and contact cleaning
Management and firmware
BMC / IPMI module repair, SEL log analysis, BIOS / VBIOS repair and flashing, power cap and topology correction
Storage and memory
Hard disk backplanes, RAID / HBA cards, NVMe thermal and throttling remediation, memory module testing and replacement
High-speed networking
InfiniBand / RoCE switches and NICs, optical module and high-speed cable degradation testing, cluster networking fault localisation
One point of contact for the whole data hall
Beyond compute equipment, UnicoNet also provides hardware sales, maintenance and system integration for enterprise network and server equipment
🔧
Switches / routers
Huawei · H3C · Juniper · Fortinet · Cisco and other leading brands; ports, power supplies, fans, firmware and configuration recovery
🖥️
Servers / workstations
Rack and tower servers, motherboards, power supplies, RAID, rails and cabling, including life-extension maintenance for out-of-warranty equipment
💾
Storage equipment
SAN / NAS, disk arrays, hard disk backplanes and controllers, together with data migration and backup planning
☁️
Cloud and integration services
Amazon Cloud and Microsoft Cloud reseller services, hybrid cloud architecture, migration and operations support
A standardised eight-step repair process
Every step is traceable: repair is not a black box
01
Intake registration and pre-inspection
Log the serial number, model and reported symptoms and open a dedicated file; visually inspect connectors, PCB and liquid-cooling lines for damage
02
Layered fault diagnosis
DCGM / nvidia-smi error analysis, Xid code interpretation, IPMI SEL logs, thermal imaging and X-Ray localisation
03
Repair plan and quotation confirmation
Issue the repair plan and quotation; units confirmed to have irreversible damage are returned as they are, with no repair charge
04
Dehumidification pre-treatment · controlled disassembly
Oven dehumidification and solder paste conditioning; diagonal removal and fitting of heatsinks and cold plates to prevent PCB warping; liquid-cooled cards drained in advance
05
Chip-level repair in a cleanroom
BGA reballing and rework, HBM replacement, power-stage component replacement, PCB trace repair, SXM connector and cold plate servicing
06
Standardised thermal reassembly
Fit new thermal paste / thermal pads and tighten diagonally; for liquid-cooled models, top up with the specified coolant and inspect the lines
07
Three-tier stress testing + burn-in
Detection test, then full-load stress test, then dedicated memory / dual-die verification; criteria: temperature < 80℃, ECC back to zero, no hardware errors; 24–72h burn-in
08
Report handover · warranty and shipment
Visual repair report with test data, with full transparency on parts and procedures; 3–6 month warranty, repeat faults repaired free of charge
Repair vs replacement: get the numbers straight first
Note: actual cost and lead time are subject to the diagnosis report; before quoting, we draw a clear line between repairable and must-be-replaced.
Three service levels, chosen around your downtime cost
STANDARD
Standard repair
Drop-off / on-site diagnosis, then chip-level repair, then stress testing, then warranty and shipment
Approx. 5 – 7 working days
Best for: non-critical nodes and spare cards whose compute can be rescheduled
EXPRESS · RECOMMENDED
Expedited repair
Priority scheduling, a dedicated engineer following up, priority allocation of spare parts, and test-and-dispatch immediately after repair
48 – 72 hours
Best for: running training jobs and clusters that stall the moment they are short of a card
ON-SITE
On-site presence and spare parts pool
Spare parts pre-positioned on your site for your model mix, so on failure the spare goes in first and the faulty unit is repaired afterwards
7×24 response · on site within 4 hours
Best for: large-scale clusters, IDC operators, cloud service providers and training platforms
The scenarios we serve
AI training / inference clusters
IDC and colocation facilities
Universities and research laboratories
Film and TV render farms
Cloud service providers / compute rental
Finance and quantitative trading
Enterprise self-built data centres
Life-extension maintenance for out-of-warranty equipment