Falcon 6040 is a 3U GPU-accelerated memory/storage server designed for AI workloads whose datasets extend beyond local GPU memory. Built on a high-density PCIe Gen6 fabric, the system integrates up to 40 E3.S NVMe SSDs and GPUs in a unified platform, delivering up to a theoretical 800 million IOPS at 512B for highly parallel, fine-grained data access.
Working with NVIDIA® SCADA™, Falcon 6040 enables GPU-initiated access to NVMe-backed data, allowing high-capacity SSDs to serve as a second-tier memory layer beyond GPU HBM. Combined with H3's software-defined PCIe fabric orchestration, the platform provides a scalable foundation for data-intensive workloads including Graph Neural Networks (GNNs), vector search, and other large-scale AI applications.
Use high-capacity NVMe as a second-tier memory/storage layer to scale AI data capacity more cost efficiently. SSD costs roughly 1/10 as much as DDR and 1/100 as much as HBM per GB, helping reduce memory infrastructure costs as datasets grow.
Scale GPU-accessible data capacity to 200+ TB with NVMe-backed second-tier memory, supporting data-intensive AI workloads such as GNNs and vector database, where graph, feature, embedding, and index data can exceed local GPU memory capacity.
Falcon 6040 works with NVIDIA® SCADA™ to enable GPU-initiated, fine-grained access across 200+ TB of NVMe-backed second-tier memory beyond local HBM, supporting highly concurrent data access for AI workloads.
Explore how NVIDIA® SCADA™ enables GPU-initiated access to NVMe storage through a PCIe interconnect. Select an element below to see how each part fits into the system.
ARCHITECTURE AT A GLANCE
Select a block to explore its role.
HOST-SIDE INITIALIZATION
SCADA provides a software layer for storage setup and GPU-initiated data access. The PCIe platform supplies the hardware interconnect.
Separate host-side initialization from GPU-initiated runtime I/O.
The CPU prepares access resources. At runtime, the GPU initiates requests and checks completion; the NVMe controller transfers data directly to or from GPU memory through PCIe.
Falcon 6040 is designed for AI workloads where datasets extend beyond local GPU memory and data access becomes highly concurrent, fine-grained, or irregular.
Support large graph and feature datasets that extend beyond local HBM, using a larger NVMe-backed data tier for highly parallel, fine-grained, and irregular access patterns.
Support large embedding and vector-index datasets beyond local GPU memory, with a high-capacity NVMe data tier enabling highly parallel and irregular data access for large-scale vector database workloads.
| Hardware Features | Falcon 6040 |
|---|---|
| Model Name | Falcon 6040 |
| BMC | ASPEED AST2600 Remote Management Processor |
| CPU |
|
| PCIe Switch | 2 x Broadcom PEX 90144 PCIe 6.0 Switches |
| Memory |
|
| E3.S NVMe SSD Bays |
|
| PCIe Slot |
|
| Power |
2+2 redundant 3200 W 80 PLUS® Titanium power supplies Hot-swappable |
| Fan |
|
| LAN |
|
| Dimension | 3U; 129.6 (H) x 438(W) x 850(D) mm |
| Compatible Devices | Falcon 6040 |
|---|---|
| Accelerator |
|
| Network Cards | NVIDIA ConnectX-7, ConnectX-8, ConnectX-9 |
| Software Features | Falcon 6040 |
|---|---|
| Management Interface | Redfish®, RESTful API, GUI and CLI |
| System Management |
|
If you want to apply for any product display, please write a form and we will contact you after receiving the message.