Specifications
AI Accelerator (AI-N)
NVIDIA L4 Tensor Core GPU, 24 GB GDDR6, 300 GB/s memory bandwidth
AI Compute (AI-N)
FP32 30.3 TFLOPS — TF32 120, FP16/BF16 242, FP8 485, INT8 485 TOPS (with sparsity)
Media Engines (AI-N)
2x NVENC, 4x NVDEC, 4x JPEG decoders for multi-stream video analytics
AI Accelerator (AI-F140)
Intel Flex 140 card, Intel Xe architecture, 12 GB high-bandwidth memory
CPU
Intel Xeon D-1700 — 4, 8 and 10-core SKUs (10C in P/N 80500-0217-G02-SL03A)
Memory
3 channels DDR4, up to 256 GB
Storage
eMMC 4 GB to 256 GB, 2x NVMe (PCIe x4), 2x 2.5 inch SATA
Network — 25G
4x 25G/10G SFP28 (SKU dependent)
Network — 10G
4x 10G/1G SFP+ (SKU dependent)
Network — Copper
3x 2.5GbE (Intel i226) plus 1x 1GbE dedicated BMC management
PCIe Expansion
1x PCIe Gen4 x16 full-height 3/4-length, 1x PCIe Gen3 x8 half-height half-length
Cellular (optional)
4G and 5G wireless modules with antenna access through extender
Time Sync (optional)
SMA inputs 10 MHz / PPS / GNSS, SMA outputs 10 MHz / PPS
Management
AST2620 BMC, NC-SI to CPU for optional in-band management
I/O
2x USB 3.0/2.0, Cisco-pinout RS232 RJ45 console
Ordering P/N
80500-0217-G02-SL03A — Marbella 10C 64G, NVMe, NVIDIA L4, 2x AC, US
Power
Redundant AC inputs; short-depth 1U rackmount form factor
Overview
The Silicom Marbella AI-N and Marbella AI-F140 are edge AI analytics appliances that pair Silicom's Xeon D-1700 Marbella networking platform with a datacenter-class accelerator in a single short-depth chassis. The platform supplies high-performance Intel server-class compute, generous ECC-addressed memory and up to 4x 25G SFP28 plus 4x 10G SFP+ connectivity, while the accelerator handles inference workloads locally at the edge rather than round-tripping data to a central datacenter.
Marbella AI-N ships the NVIDIA L4 Tensor Core GPU: 24 GB of GDDR6 at 300 GB/s delivering 485 INT8 TOPS and 485 FP8 TFLOPS with sparsity, plus two NVENC, four NVDEC and four JPEG engines for dense multi-camera video analytics. Marbella AI-F140 instead uses the Intel Flex 140 card with its Xe architecture and 12 GB of high-bandwidth memory, giving integrators an open-software alternative where Intel oneAPI tooling is preferred.
Because both variants share the same Marbella baseboard, the appliances support the same storage layout (eMMC plus dual NVMe plus dual 2.5 inch SATA), optional 4G/5G cellular, optional GNSS/PPS/10 MHz time synchronisation, and a dedicated BMC management path. They are designed for telco edge, manufacturing and IoT deployments where rack depth, latency and power budget are all constrained.
Key Benefits
Single-chassis edge inference: datacenter-class GPU acceleration in a short-depth 1U appliance. High-speed edge uplinks: up to 4x 25G SFP28 plus 4x 10G SFP+ for camera and sensor aggregation. Deterministic timing: optional GNSS, PPS and 10 MHz synchronisation for time-sensitive workloads. Field serviceability: AST2620 BMC with NC-SI in-band management and redundant AC power.
Applications
Real-time video analytics and multi-camera inference, telecom edge MEC nodes, smart-manufacturing quality inspection, retail and campus analytics, IoT data aggregation with local AI processing, and low-latency edge services where WAN backhaul is limited.
Request a Quote — SILICOM MARBELLA AI-N & AI-F140 — EDGE AI ANALYTICS APPLIANCES
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →