News
BackNew GPUs – Double the VRAM, More Power
cloudscale's GPU servers power your AI workloads. In addition to dedicated CPU cores, 1 to 4 GPUs per virtual server deliver the performance needed to run even demanding applications. Now we are taking it to the next level: effective immediately, we are offering NVIDIA RTX PRO 6000 Max-Q GPUs instead of the L40S. But we are not stopping at more powerful GPUs: the new "GPU2" flavors also come with more memory – and are even more affordable.
The next generation
The NVIDIA L40S in our previous GPU servers came with 48 GB of VRAM, and up to four of the GPUs could be used in parallel in a virtual server. The new NVIDIA RTX PRO 6000 Max-Q boasts twice that amount: 96 GB of VRAM per GPU, and here too, you can tap into the power of up to four GPUs in a single server. The physical GPUs are passed through to the virtual server undivided, so that the full performance is dedicated to your use.
Speaking of performance: The RTX PRO 6000 not only offers more VRAM, but also significantly more computing power than the L40S, and we were also impressed by the energy efficiency of the "Max-Q" version. In line with the larger VRAM, we are also equipping our new GPU2 flavors with more memory, which you can combine in various ratios with 16 to 96 dedicated CPU cores. Nothing has changed with the scratch disk: up to 1600 GB of lightning-fast NVMe SSD storage is available locally to minimize latency.
Migrate with care
To switch from a GPU1 server with L40S to a GPU2 server with RTX PRO 6000 Max-Q, in principle, it is sufficient to scale the server to one of the new flavors via the cloud control panel or API. Scaling GPU servers can take a moment, as moving to different physical hardware also requires transferring the contents of your scratch disk.
We recommend, however, that you create a second, new server as a precaution and first make sure that everything works as desired (including the open-source GPU kernel module required for the Blackwell architecture, in Debian/Ubuntu for example the package nvidia-open). You can then migrate the workload to the new server – if you use a Floating IP and/or a load balancer, the IP address will remain the same for your users.
The new, and in some cases considerably cheaper, GPU2 flavors with NVIDIA RTX PRO 6000 Max-Q GPUs have been available for a few days now, and early customer feedback on "real" workloads indicates a significant improvement in performance. We are confident that your application will benefit from our GPU servers too – why not try it out for yourself?
Step it up a notch,
Your cloudscale team