The MLA100 PCIe Card brings ARIES AI inference to standard servers and PCs. It’s a single-slot card with a 25W TDP that runs vision, language and multimodal models without replacing your entire system infrastructure.
AI performance (Dense, INT8)
AI performance
80 TOPS (Dense, INT8)
NPU clock
1.25 GHz
NPU cores
8
Memory
LPDDR4X, 16 GB / 32 GB
Memory bandwidth
66.7 GB/s
Flash memory
32 MB
Host interface
PCI Express Gen4, 8-lane
UART
1 Port
LED indicators
2 (Power, Heartbeat)
Supported OS
Linux, Windows
Frameworks
PyTorch, TensorFlow, TFLite, ONNX, Keras
Power (TDP)
25 W
Input voltage
DC +12V
Cooling system
Heatsink, Fan cooling
Dimensions
167 (L) × 71 (W) × 16 (H) mm
Weight
225 g
MODEL
TASK
ARIES
FPS per watt vs. a 360W desktop GPU (YOLO11n-seg)
FPS per dollar vs. a 230W workstation GPU (YOLO11n-seg)
FPS vs. a 230W workstation GPU (YOLO11n), at a fraction of the power
Measured on the MLA100 PCIe Card, batch 8. Based on internal benchmark testing.
Processor operations only. At USD 0.08/kWh, the average US industrial electricity rate (May 2024).
About 9× less than a 230W-class GPU and 18× less than a 450W-class GPU at TDP.
The card’s own fan is enough, with no liquid cooling or added air handling at the facility.
An 8-slot AI server holds 8 MLA100 cards. The same number of dual- or triple-slot GPUs would need two or more servers.
MLA100 PCIE CARD
TYPICAL DUAL-SLOT GPU
MLA100 PCIe Card
1 slot
Dual-slot GPU
2 slots
Triple-slot GPU
3 slots
The qb Compiler applies mixed precision layer by layer and keeps over 99% of original FP32 accuracy, with no retraining.
mAP on COCO, measured on the MLA100 PCIe Card. Percentages show accuracy retained versus FP32. GPU INT8 results use a standard GPU INT8 inference toolchain.
Memory and key parts are industrial grade, rated for -40 to 85℃ operation.
Passed a 504-hour operating life test and a 700-cycle thermal test (-40 to 90℃), and met Korean National Police Agency environmental standards.
ISO 9001
KC
FCC
CE
RoHS
VCCI
BSMI
PSE
Tested for harsh environments (MLA100 PCIe Card): a 504-hour operating life test and a 700-cycle thermal test (-40 to 90℃), and it passed Korean National Police Agency environmental standards.
Certifications
ISO 9001
KC
FCC
CE
RoHS
VCCI
BSMI
PSE
Over 490 validated model architectures, from CNNs to LLMs and VLMs. Pre-compiled models are available in the Mobilint Model Zoo and on Hugging Face.
STEP 01
qb Compiler converts PyTorch, TensorFlow or ONNX models into an MLA100 executable, keeping over 99% of original accuracy.
STEP 02
qb Runtime provides Python and C++ APIs for loading models and running inference.
STEP 03
qb Driver supports Linux and Windows hosts.
Intersection video analyzers handle vehicle detection, license plate OCR and vehicle re-identification on-site for municipal traffic centers.
MLA100 cards replaced GPU servers for speech AI, handling STT at 240 calls per second and TTS at 45 calls per second.
Secondary battery lines and wafer inspection equipment use MLA100 to detect micro-defects and classify defect patterns.
