Nvidia Rubin PCIe GPUs & HPE ProLiant Gen13

Nvidia Rubin PCIe GPUs, expected to follow the Blackwell architecture, will bring a generational leap in AI compute and memory bandwidth to HPE ProLiant Gen13 servers, but only in specific configurations. These GPUs will primarily benefit ProLiant Gen13 models designed for AI and HPC workloads, while standard rack servers will see limited impact. This article explains what Rubin PCIe means for Gen13, covering architecture, compatibility, power, and practical considerations.

Background: The Nvidia GPU Roadmap and Rubin

Nvidia’s GPU architecture roadmap follows a consistent cadence: Ampere, Hopper, Blackwell, and then Rubin. Rubin, named after astronomer Vera Rubin, is the next major architecture after Blackwell. While Blackwell (the B100, B200, and GB200) is currently rolling out, Rubin is slated for release in 2026. The Rubin family will include both data-center GPUs in the HGX form factor and, critically for this discussion, PCIe add-in cards. The PCIe versions are the ones that will fit into standard ProLiant Gen13 servers, as opposed to the proprietary HGX baseboards that require specialized chassis.

HPE ProLiant Gen13: The Target Platform

HPE ProLiant Gen13 servers are the next generation of HPE’s rack and tower servers, following Gen12. They are built around Intel Xeon (Granite Rapids) and AMD EPYC (Turin) processors, with support for DDR5 memory, PCIe Gen5 (and some Gen6) I/O, and high-speed networking like 400GbE. The Gen13 line includes models like the DL380, DL385, and the AI-optimized DL380a, which is specifically designed to house up to eight double-width GPUs. The DL380a is the primary server for Rubin PCIe GPUs, as it provides the necessary power, cooling, and physical space.

Rubin PCIe GPU Specifications and Features

While final specs are not public, Rubin is expected to deliver significant improvements over Blackwell. Based on Nvidia’s public statements and industry trends, the Rubin PCIe GPU will likely feature:

• HBM4 memory: The next generation of High Bandwidth Memory, offering even higher bandwidth and capacity than HBM3e used in Blackwell. Expect 288GB to 384GB per GPU.
• Advanced process node: Likely TSMC’s N3 (3nm) or better, improving performance per watt.
• Next-gen NVLink: For high-speed GPU-to-GPU communication, crucial for multi-GPU AI training.
• Increased FP4 and FP8 performance: For AI inference and training, with possibly new precision formats.
• PCIe Gen6 support: While the physical slot is Gen5, the GPU may support Gen6 for higher bandwidth to the CPU.

Compatibility: Which ProLiant Gen13 Servers Support Rubin?

Not all ProLiant Gen13 servers can support Rubin PCIe GPUs. The key requirements are:

• Physical space: Rubin PCIe GPUs will be double-width, full-height, full-length cards. They need the large chassis of the DL380a or similar GPU-optimized servers.
• Power delivery: Each GPU is expected to draw 600W or more. The server must have 2400W or 3000W power supplies and the appropriate power cabling (12VHPWR or similar).
• Cooling: The DL380a uses high-performance fans and a direct-to-chip liquid cooling option. Air cooling may be insufficient for the highest-power Rubin cards.
• BIOS and firmware: HPE will likely release BIOS updates for Gen13 to properly initialize and manage Rubin GPUs. Ensure you have the latest system ROM.
• Operating system and drivers: You’ll need a recent Linux distribution (e.g., RHEL 9.4+, Ubuntu 24.04+) with the latest Nvidia driver (R560+).

How Rubin PCIe GPUs Benefit ProLiant Gen13

For users who can properly configure a Gen13 server, Rubin PCIe GPUs offer several concrete advantages:

• Faster AI training: The increased memory bandwidth and compute throughput (especially for FP4) can cut training time for large language models by 30-50% compared to Blackwell.
• Larger model inference: With up to 384GB HBM4, you can run models with trillions of parameters on a single GPU, reducing the need for multi-GPU sharding.
• Better energy efficiency: The advanced process node means higher performance per watt, lowering operating costs for AI workloads.
• Future-proofing: Investing in Rubin-ready Gen13 servers positions you for the next wave of AI software and algorithms.

Power, Cooling, and Physical Requirements

Integrating Rubin PCIe GPUs into a ProLiant Gen13 server requires careful planning:

• Power: A server with eight Rubin GPUs could draw over 6kW. That requires 240V AC power, and you must ensure your data center power distribution can handle it. HPE offers 3000W power supplies for the DL380a.
• Cooling: The DL380a’s air cooling can handle up to 350W GPUs, but for 600W+ Rubin cards, liquid cooling is recommended. HPE offers a liquid-cooled version of the DL380a. Ensure your rack has the necessary plumbing.
• Physical dimensions: Rubin PCIe cards are long (over 12 inches) and heavy. They need anti-sag brackets and proper support. The DL380a chassis is designed for this.

Cost and Procurement Considerations

Rubin PCIe GPUs will be expensive, likely in the $30,000-$40,000 range per GPU. For a full server with eight GPUs, the GPU cost alone could exceed $250,000. Additionally, you’ll need to budget for:

• HPE ProLiant DL380a server base (around $15,000-$20,000).
• High-performance networking (e.g., InfiniBand or 400GbE) for multi-node scaling.
• Software licenses for AI frameworks (though many are open-source).
• Increased power and cooling costs in your data center.

Consider whether your workloads truly require Rubin, or if Blackwell (B200) is sufficient, as it will be more mature and potentially more available.

Common Problems and How to Avoid Them

Users integrating high-end PCIe GPUs into servers often encounter these issues:

• Insufficient power: The server may not have enough power supply capacity or the correct cabling. Always verify the total system power draw and use HPE’s power configurator.
• Overheating: Air-cooled servers may throttle the GPU under sustained load. Monitor GPU temperatures with nvidia-smi and ensure adequate airflow.
• Driver issues: New GPUs require the latest Nvidia drivers. Use the HPE-tested driver versions to avoid stability problems.
• BIOS compatibility: Older BIOS versions may not recognize the GPU. Update the system ROM to the latest HPE version.
• Physical fit: Some PCIe slots may not have enough clearance. Use the GPU support brackets provided by HPE.

Frequently Asked Questions

Q: Will Rubin PCIe GPUs work in a standard ProLiant DL380 Gen13 (not the ‘a’ model)?
A: No. The standard DL380 has limited power and cooling, and the PCIe slots are not positioned for double-width GPUs. You need the DL380a or a similar GPU-optimized server.

Q: Can I mix Rubin and Blackwell GPUs in the same server?
A: Technically yes, but Nvidia does not support mixing different architectures in the same NVLink domain. It’s best to use all Rubin or all Blackwell for consistent performance.

Q: What is the power consumption of a single Rubin PCIe GPU?
A: Nvidia hasn’t released official specs, but based on the trend, it’s expected to be around 600W to 700W. The server’s power supplies must be sized accordingly.

Q: Do I need liquid cooling for Rubin GPUs in a ProLiant Gen13?
A: For the highest-power variants (600W+), liquid cooling is strongly recommended to avoid thermal throttling. Air cooling may work for lower-power SKUs, but always check HPE’s thermal guidance.

Q: When will Rubin PCIe GPUs be available for purchase?
A: Nvidia has indicated a 2026 release. However, availability may be limited initially due to high demand for AI GPUs.

Q: Will HPE offer Rubin as a factory option in Gen13 servers?
A: HPE typically offers the latest Nvidia GPUs as configurable options in their GPU-optimized servers. It’s likely that Rubin will be offered, but confirm with HPE at the time of purchase.

Related

• Convert Play FLAC in iTunes: Full Guide
• Soda Foundation Charter: What It Means
• Convert Table to Text in Word: Step-by-Step Guide