Since H3 began developing PCIe Gen 3 switch solutions, we continuously pioneer revolutionizing composable infrastructure for data centers. Following the announcement of PCIe 5.0 in 2019, H3 dedicated itself to accumulating energy and is now ready to launch the PCIe Gen 5 composable GPU expansion and NVMe MR-IOV in May 2023. We are boldly blowing the trumpet for the coming era of highly performing disaggregated data management solutions. Here are going to use three keywords to feature the form factors and breakthroughs of the next-generation Falcon series: Speed, Scale, and Flexibility.
Speed —The replacement of Gen 4 with Gen 5 doubles the throughput, beautifully heightening the performance of x16 with 63 GB/s compared to the 31.5 GP/s of Gen 4. With the double performance and low latency, the data transfer between GPUs and NVMe will be much faster for AI applications, especially for peer-to-peer data transfer (P2P).
Scale —The breakthrough of cross-domain technology enables the scalability of devices from one single switch to multiple switches. With this new cutting-edge switch-to-switch cascade, 4 hosts, 16 GPUs, and 32 NVMe SSDs can be connected within one unified PCIe switch fabric by cascading 4 Broadcom PCIe switches. By enlarging the fabric scale, PCIe/ CXL will be the king of the rack-level composable fabric.
Flexibility —The design change of the system from all-in-one to ToR, JBOG, and JBOf liberates the ratio of host and devices. The user can decide how many racks with NVMes and GPUs to install and collocate needed fabric boxes and hosts rather than being restricted to chassis with a fixed ratio between servers and devices. This flexibility brings the hardware composability for different workloads to better utilize the more and more expensive devices.
Learn more about how this future-proof composable infrastructure as a service can aid your enterprise to operate more efficiently, at a larger scale.