Despite the enthusiasm around AMD's offering, there are concerns about its performance compared to Nvidia. To address this, TensorWave plans to launch its MI300X nodes using RDMA over Converged Ethernet (RoCE). The company will fund its expansion by using its GPUs as collateral for a large round of debt financing, a strategy used by other datacenter operators. TensorWave's COO, Piotr Tomasik, hinted at a major announcement regarding this financing later in the year.
Key takeaways:
- TensorWave, a startup specializing in AI infrastructure, is opting to use AMD's Instinct MI300X accelerators instead of Nvidia GPUs, citing cost-effectiveness and availability.
- By the end of 2024, TensorWave plans to have 20,000 MI300X accelerators deployed across two facilities, and will introduce additional liquid-cooled systems in the following year.
- The startup is facing challenges in terms of confidence in AMD's performance and supply chain issues, particularly with rear door heat exchangers (RDHx), a cooling technology for denser GPU clusters.
- TensorWave plans to fund its infrastructure build by using its GPUs as collateral for a large round of debt financing, similar to strategies used by other datacenter operators like Lambda and CoreWeave.