- Nvidia’s acquisition brings Enfabrica engineers directly into its AI ecosystem
- EMFASYS chassis pools up to 18TB of memory for GPU clusters
- Elastic memory fabric frees HBM for time-sensitive AI tasks efficiently
Nvidia’s decision to spend more than $900 million on Enfabrica was something of a surprise, especially as it came alongside a separate $5 billion investment in Intel.
According to ServeTheHome, “Enfabrica has the coolest technology,” likely because of its unique approach to solving one of AI’s largest scaling problems: tying tens of thousands of computing chips together so they can operate as a single system without wasting resources.
This deal suggests Nvidia believes solving interconnect bottlenecks is just as critical as securing chip production capacity.
A unique approach to data fabrics
Enfabrica’s Accelerated Compute Fabric Switch (ACF-S) architecture was built with PCIe lanes on one side and high-speed networking on the other.
Its ACF-S “Millennium” device is a 3.2Tbps network chip with 128 PCIe lanes that can connect GPUs, NICs, and other devices while maintaining flexibility.
The company’s design allows data to move between ports or across the chip with minimal latency, bridging Ethernet and PCIe/CXL technologies.
For AI clusters, this means higher use and fewer idle GPUs waiting for data, which translates into better return on investment for costly hardware.
Another piece of Enfabrica’s offering is its EMFASYS chassis, which uses CXL controllers to pool up to 18TB of memory for GPU clusters.
This elastic memory fabric allows GPUs to offload data from their limited HBM memory into shared storage across the network.
By freeing up HBM for time-critical tasks, operators can reduce token processing costs.
Enfabrica said reductions could reach up to 50% and allow inference workloads to scale without overbuilding local memory capacity.
For large language models and other AI workloads, such capabilities could become essential.
The ACF-S chip also offers high-radix multipath redundancy. Instead of a few massive 800Gbps links, operators can use 32 smaller 100Gbps connections.
If a switch fails, only about 3% of bandwidth is lost, rather than a large portion of the network going offline.
This approach could improve cluster reliability at scale, but it also increases complexity in network design.
The deal brings Enfabrica’s engineering team, including CEO Rochan Sankar, directly into Nvidia, rather than leaving such innovation to rivals like AMD or Broadcom.
While Nvidia’s Intel stake ensures manufacturing capacity, this acquisition directly addresses scaling limits in AI data centers.
