### HPE Upgrades Supercomputer Lineup for 2025 #### Overview Hewlett Packard Enterprise (HPE) is set to upgrade its supercomputer lineup across the board in 2025. The company, which has a rich history in high-performance computing (HPC) through its acquisitions of Cray, Silicon Graphics, Convex, Compaq, and HP, is focusing on enhancing networking, compute, and storage capabilities to stay competitive in the exascale market. This move is crucial as HPE aims to maintain its dominant market share, which stood at just under a quarter in 2023, according to Hyperion Research. #### Networking Upgrades HPE's networking upgrades are centered around the new "Slingshot 400" family, which includes the "Rosetta-2" switch ASIC and the "Cassini-2" network adapter ASIC. The first generation "Rosetta" and "Cassini" chips, introduced in 2019 and fully ramped by 2022, supported 200 Gb/sec speeds. The new Rosetta-2 switch will likely be fabricated using TSMC's 7 nanometer process, reducing power consumption while doubling the native signaling speed to 56 Gb/sec per lane, resulting in 400 Gb/sec per port. The Cassini-2 adapter is expected to follow a similar pattern, doubling its speed to 400 Gb/sec. These enhancements are crucial for HPE to compete with other major players like Broadcom, Nvidia, and Cisco Systems in the HPC networking market. #### Compute Upgrades HPE is also introducing new compute blades for its Cray EX and ProLiant XD systems. The Cray EX4252 Gen 2 compute blade, available in early 2025, will feature eight AMD "Turin" Zen 5c Epyc 9965 processors, each with 192 cores running at a base clock speed of 2.25 GHz. This configuration allows for 98,304 cores in a single liquid-cooled rack, delivering 1.77 petaflops of base performance and potentially reaching 2 petaflops with all-core turbo. With 500 racks, an all-CPU exascale supercomputer can be built, showcasing the significant advancements in compute density and performance. By the end of 2025, HPE will release the EX154n compute blade, compatible with Nvidia's Grace-Blackwell GB200 NVL4 compute board. This board integrates a pair of Grace CPUs and a quad of Blackwell GPUs into a single unit, offering a six-way shared memory cluster. The EX154n blade will support 224 Blackwell GPUs in a single EX4000 cabinet, achieving over 10 petaflops of performance with 42 TB of HBM3E stacked memory and 52.9 TB of LPDDR5 memory. Despite the higher compute density, the price/performance ratio is only 35 percent better for the GPU setup compared to the CPU setup, primarily due to the high cost of HBM memory. #### Storage Enhancements Storage is another critical component of supercomputers, and HPE is not overlooking this area. The new Cray E2000 all-flash array, set to replace the E1000, will feature 32 NVM-Express flash drives and two storage servers with PCI-Express 5.0 slots. This upgrade promises more than double the read and write performance, enhancing the overall efficiency of HPE's supercomputing solutions. #### ProLiant XD Line The ProLiant XD line, designed for AI service providers and large enterprise customers, is also receiving significant updates. The ProLiant XD680 will pair two Intel "Emerald Rapids" Xeon SP processors with eight Intel Gaudi 3 accelerators, available before the end of 2025. This system is expected to meet the demand for AI workloads, although Intel has acknowledged lower-than-expected demand for the Gaudi 3 devices. The ProLiant XD685, available in early 2025, will feature a pair of AMD Turin Epyc CPUs and offer the choice of eight Nvidia Hopper or Blackwell GPUs or eight AMD "Antares" MI300X or MI325X GPUs. This flexibility in GPU options will cater to a wide range of AI and HPC workloads. The XD685 will be available in both air-cooled and liquid-cooled versions, providing options for different cooling requirements. #### Market Context The HPC market is highly competitive and price-sensitive. While HPE leads with a significant market share, other major players like Eviden (Atos), Fujitsu, Lenovo, and Dell are also vying for a piece of the pie. The presence of indigenous labs in China and the reluctance of European centers to buy outside the EU add layers of complexity to the market dynamics. HPE's strategy to continuously upgrade its systems with new networking, compute, and storage technologies is essential to maintaining its leadership position. #### Conclusion HPE's comprehensive upgrades for 2025, encompassing networking, compute, and storage, highlight the company's commitment to staying at the forefront of the HPC and AI supercomputing markets. These enhancements, particularly in the Cray EX and ProLiant XD lines, are designed to meet the evolving needs of HPC centers and large enterprises, ensuring that HPE remains a key player in the exascale computing landscape. As the market continues to evolve, HPE's ability to innovate and deliver high-performance, cost-effective solutions will be crucial for its ongoing success.