Lattice Computing
Lattice Computing — The Next-Generation AI Compute Architecture
· GPU/SSD Tight Coupling · Microsecond Latency · 10x Bandwidth◆Definition in One Sentence
Powered by Lattice Fabric / Extended PCIe + , this pioneering architecture achieves tight GPU/SSD coupling: GPUs and NVMe SSDs are directly paired through the extended PCIe interconnect, with end-to-end data exchange completed entirely within the expanded PCIe domain.
▍Latency Comparison: Milliseconds → Microseconds
Traditional Path· Traditional PCIe Networks — Millisecond-Level Latency
SSD→
System Memory→
NIC→
Network→
NIC→
System Memory→
GPU
Millisecond-level ms · Slow
Lattice Interconnect· Extended PCIe — Microsecond-Level Latency
NVMe SSD→
Extended PCIe Domain→
GPU
Microsecond-level μs · Order-of-magnitude gain
▍Why It Matters: Industry Pain Points
▣All-Flash Storage
Already a hundred-billion-dollar market, yet high reliability still depends on expensive dual-port SSDs.
▣Container Cloud (K8s)
Kubernetes has become mainstream, yet a high-performance, easy-to-deploy storage subsystem remains missing.
▣GPU / AI Compute
AI compute is throttled by storage performance; the network becomes the bottleneck and GPU power cannot be fully unleashed.
▍Lattice Interconnect PCIe Card
Note: The gold fingers plug into the server motherboard's PCIe slots; GPU interconnect ports bring external GPUs into the same PCIe domain via cables/backplanes, enabling tight GPU↔SSD coupling and microsecond-level data exchange.
▦Key Components & Specifications
| Component | Specification / Function |
|---|---|
| Form Factor | Full-height full-length (FHFL) PCIe card, 312 × 111.15 mm, dual-slot thickness (with heatsink); optional E3.S backplane card form |
| PCIe Switch | Onboard high-lane-count PCIe Switch (64–96 lanes) that expands the PCIe domain, enabling extended PCIe interconnect |
| NVMe SSD Slots | 8 E3.S / U.2 (or M.2) NVMe SSD slots; each drive delivers ~100 GB/s over Gen6 x16, forming an SSD pool |
| GPU Interconnect Ports | PCIe x16 high-speed interface (cable/backplane) that connects external GPUs into the same PCIe domain for direct GPU↔SSD data exchange |
| Host Interface | PCIe x16 gold fingers connecting to the server CPU/motherboard |
| Power & Cooling | Auxiliary power (8-pin) + onboard VRM power module; large passive heatsink, optional active fan module |
| Software | KFNative™ driver & firmware: SSD pool discovery, KV-Cache mapping, Ring Reduce support |
Standard PCIe Card (Plug into Server)
OCP 3.0 Accelerator Form
E3.S Rack-Scale Backplane Integration
GPU Server Companion Card
· PUF 物理不可克隆 · 防破解存储 · 先进封装 · 全链路测试