The Federal Bureau of Investigation (FBI) is considering acquiring AI supercomputer hardware.

In a request for information (RFI), the FBI said that it sought to understand "AI computing products capable of supporting large language model (LLM) training, inference, advanced analytics, computer vision, and other AI-enabled workloads within secure government computing environments."

RFIs are a precursor to procurement and do not guarantee that the agency will purchase hardware.

At this stage, the FBI could deploy one or more systems and "anticipates acquiring equipment in varying quantities and configurations through individual delivery orders," RFI documents disclose.

The FBI breaks down its AI compute needs into four categories. The first would feature Intel Xeon 6767P processors or equivalent/better and a minimum of 2TB of RAM (with pricing options to increase memory capacity up to 4TB). It would also include one HGX B300 eight-GPU air-cooled assembly or equivalent.

Storage is expected to include two 500GB or larger SSD boot drives and two 7.68TB or larger SSD data drives. Networking requires two 25Gb Ethernet interfaces and two Nvidia BlueField-3 single-port 400 GbE DPUs or equivalent.

Category two requires "integrated rack-scale AI systems designed to support large-scale AI training and inference workloads."

Solutions are expected to meet or exceed one integrated GB300 NVL72 rack-scale system or equivalent, an integrated high-speed accelerator interconnect fabric, integrated Ethernet networking equivalent to Spectrum-X architecture, integrated cooling solution utilizing an in-rack or sidecar coolant distribution unit (CDU); and compatibility with Government-provided facility water cooling infrastructure.

It should have two 500GB or larger SSD boot drives and two 3.84TB or larger SSD data drives.

Category three seeks "AI pod architectures optimized for enterprise artificial intelligence workloads."

This system would consist of one integrated AI pod architecture equivalent to Google's TPU v8, TPU v7L, or better, and provide overall compute capability that is equivalent to, or exceeds, the performance of five GB300 NVL72 rack-scale systems. It should support the execution of large language models and enterprise AI workloads, as well as the deployment of multiple AI frameworks and models.

Google Cloud’s Vertex AI services are specifically mentioned, although that service was recently rebranded as the Gemini Enterprise Agent Platform. The inclusion of TPUs in the procurement document is unusual, with the Google chip not normally offered as an on-prem product, beyond its Edge variant.

Previously exclusively available on Google Cloud, the TPU is set to come to a Google-co-owned cloud, be available in Fluidstack data centers, and potentially on neocloud services. Whether the FBI could currently get TPUs for an on-prem deployment remains unclear.

Finally, category four is focused on AI inference accelerators featuring Nvidia L40-S or equivalent, with at least 48GB of GPU memory optimized for convolutional neural network (CNN), embedding, and generative AI inference workloads.

Accelerators need to support mixed precision inference, including FP32, FP16, BF16, and INT8, and be compatible with industry-standard AI frameworks, including CUDA, TensorRT, PyTorch, and TensorFlow.

For all four categories, the majority of orders are expected to be delivered to the FBI Criminal Justice Information Services (CJIS) Division facility in Clarksburg, West Virginia.