‍

Quick Look on RTX 5090

The NVIDIA GeForce RTX 5090 introduces significant advancements over the RTX 4090, particularly in architectural enhancements and memory capacity:

1. Blackwell Architecture with FP4 Support

Built upon NVIDIA's Blackwell architecture, the RTX 5090 introduces support for 4-bit floating point (FP4) precision in its 5th Genertation Tensor Cores. This enhancement significantly accelerates AI-related tasks by doubling performance for specific operations, thereby improving efficiency in AI model training and inference. 

2. Memory Capacity and Bandwidth

In comparison to the 24 GB GDDR6X in RTX 4090, RTX 5090 is equipped with 32GB of high-speed GDDR7 memory, , offering increased capacity to handle larger models and more complex workloads.

Workstation Configuration

To get our RTX 5090 GPU up and running quickly, we visited a local Micro Center to pick up in-store accessories. Our configuration serves as a reference design for a dual-GPU setup. We are happy to share the hardware list of our workstation setup:

‍

Due to stock limitations, some component choices are not ideal. If you're planning to build a similar workstation we recommend the following adjustments:

1. SSD Storage: Opt for the standard SN850X instead of the PS5-optimized SN850P, as the latter's heat sink is too thick and causes installation issues.
2. Power Supply: Choose a newer model with a 12VHPWR interface for easier installation and improved power stability. A recommended option is the Seasonic PRIME TX-1600.
3. Memory: Ensure the RAM is DDR5 ECC Registered, as the motherboard supports only 4 slots. We recommend using the highest-capacity single modules available to maximize memory capacity.

Overcoming Configuration Challenges

Setting up a brand-new GPU model like the RTX 5090 is never without its difficulties. We encountered a few key issues along the way. Initially, our workstation failed to recognize the GPU vialspci, even after upgrading the driver to version 570 (with CUDA 12.8 support) and adjusting the PCIe slot settings to PCIe gen4 and gen3. None of these attempts resolved the issue. The display remained completely black after exiting the BIOS. Through debugging, we identified the critical steps required to make the GPU compatible with our workstation:

1. Enable the Following PCIe Settings:

Resize BAR Support: A PCIe feature that allows the system to dynamically adjust the amount of GPU memory (VRAM) that is directly accessible to the CPU. Traditionally, GPUs expose only 256MB of VRAM at a time due to legacy 32-bit addressing limitations. Resize BAR enables access to the full VRAM

SR-IOV Support: A PCI Express (PCIe) feature that allows a single physical PCIe device (e.g., a GPU or network card) to be shared among multiple virtual machines or containers.

‍

2. Disable CSM

CSM (Compatibility Support Module) is a BIOS/UEFI setting that enables legacy boot mode for older operating systems and devices that do not support UEFI 

‍

3. UEFI Mode Setting

Choose Other OS instead of default Windows UEFI mode

4. Install the Latest Driver Manually

i. Do NOT install via apt install nvidia-570-driver, as this package is incompatible with our configuration. The workstation still failed to recognize the GPU after successful driver installation.

ii. Important: Download the latest driver from nvidia.com/en-us/drivers. Select the MIT/GPL Version instead of the standard NVIDIA Proprietary version on the first interface.

‍

‍

‍

‍