Massed Compute

Pros: Cheap price. But they get money from you in other ways. To be honest, there really isn't much else positive to talk about from my experience.

Cons: frequently unavailable, and when I finally deployed a VM with 4× H200 NVL GPUs on Ubuntu 24.04, the system image contained software packages that were significantly older than the versions that normally come with a standard Ubuntu 24.04 installation. In addition, CUDA was preinstalled, but it was also an outdated version. It gave the impression that an intentionally outdated software environment had been prepared, although I cannot confirm that. As a result, I had to spend a significant amount of billed time updating the system packages, CUDA, drivers, and related dependencies before I could begin my work. More importantly, the GPUs did not appear to have working NVLink connectivity, and CUDA peer-to-peer access between GPUs was unavailable. For multi-GPU workloads, this is a major limitation because GPU-to-GPU communication is critical for performance. I also requested a refund for my remaining unused credits, but it was declined. My interactions with support were slow and unhelpful.

Overall: My overall experience was disappointing. I spent a considerable amount of billed time updating the CUDA environment instead of running my workloads. The lack of GPU peer-to-peer access and apparent absence of NVLink support made the instance unsuitable for the distributed training workload I rented it for. Combined with hardware availability issues, the refusal to refund unused credits, and an unsatisfactory support experience, I would not choose this platform again.