For data center deployments, GroqRack provides an extensible accelerator network. Combining the power of an eight GroqNode™ set, GroqRack features up to 64 interconnected chips. The result is a deterministic network with an end-to-end latency of only 1.6µs for a single rack, ideal for massive workloads and designed to scale out to an entire data center.
Eight GroqNode™ servers with 64 interconnected cards plus 1 additional redundant node reduces unexpected downtime impact.
14 GB shared global SRAM delivers large globally sharable SRAM for high-bandwidth, low-latency access to model parameters.
Low latency and high performance delivers large globally sharable SRAM for high-bandwidth, low-latency access to model parameters.
704 RealScale™ chip-to-chip connectors enable near-linear multi-server and multi-rack scalability without the need for external switches.
Up to 3.2 TBps global bisectional bandwidth facilitates massive concurrency and data parallelism needed for bandwidth-sensitive applications.
End-to-end on-chip protection improves uptime and reliability with error-correction code protection throughout the entire GroqChip™ data path.
GroqRack 42U Server Chassis
Up to 9 x GroqNode 1 (GN1-B8C) servers with a fully connected internal RealScale network delivering accelerated compute performance up to 48 POPs (INT8), 12 PFLOPs (FP16)
288 x QSFP28 GroqNode connectors creating a switchless routing fabric with 3.2TBps total global bisection bandwidth across a single global hop
Primary Rack Controller Server with AMD EYPC 7413 Processor
NVMe Server with dual AMD EYPC 7413 Processors with 8 x 7.68TB NVME installed for data (61.44TB total storage)
2 x 100GE Ethernet Switches (data network + redundancy) with min 32 x QSFP28 ports
1GE Ethernet Switch (management network) with min 48 x RJ45 ports
4x 17.3kW PDUs (2N Redundancy)
OS & Software
Ubuntu Linux with GroqWare™ Suite (SDK and Utilities) pre-installed on dual SATA SSD drives (RAID1 config). Optional support for RHEL & Rocky Linux.