Revolutionizing Cloud Computing with AMD Accelerators

by Annabelle Stephenson
in *AR, *CH, *DE, *ES, *FR, *JP, *KO, *PT, *RU, *TR, *UA, News, PC, xBG, xCS, xDA, xEL, xET, xFA, xFI, xHE, xHR, xHU, xLT, xLV, xNL, xNN, xRO, xRS, xSI, xSK, xSV
on 16 April 2024

Najnowsze akceleratory MI300X firmy AMD zyskują popularność wśród operatorów chmury

Cloud operators specializing in managing energy-intensive graphics processing units (GPUs) and other artificial intelligence infrastructures are increasingly turning to AMD accelerators instead of Nvidia’s offerings.

One such operator, TensorWave, has recently started integrating systems with the new AMD Instinct MI300X accelerators. The company plans to rent these chips at lower prices compared to Nvidia’s accelerators.

The latest AMD accelerators have garnered significant interest due to their advantages. They are readily available for purchase compared to Nvidia’s competing products, giving TensorWave access to a large quantity through their negotiation skills.

By the end of 2024, TensorWave aims to install 20,000 MI300X accelerators in their two facilities. Additionally, the company plans to introduce liquid-cooled systems next year for enhanced performance.

The AMD integrated circuits are also faster than the highly sought-after Nvidia H100. The MI300X outperforms the H100 in specifications, boasting greater memory capacity and data throughput. It was unveiled during AMD’s Advancing AI event in December 2024 and is reported to be 32% faster than the Nvidia H100.

In comparison to the rival H100, the AMD chip features larger 192 GB HBM3 memory capacity, enabling a data throughput of 5.3 TB/s. In contrast, the H100 has an 80 GB memory capacity and 3.35 TB/s throughput.

While AMD accelerators are gaining popularity, some users still question their performance compared to Nvidia’s products. TensorWave plans to deploy MI300X nodes using RoCE (RDMA over Converged Ethernet) technology to expedite deployment processes.

Long-term plans for TensorWave include implementing a more advanced resource management solution, connecting up to 5,750 GPUs and petabytes of high-throughput memory using GigaIO’s PCIe 5.0-based FabreX technology. This project will be funded through a secured GPU accelerator credit, a method also used by other data center companies.

Similar initiatives have been undertaken by other industry players, with Lambda securing a $500 million credit and CoreWeave obtaining $2.3 billion for infrastructure expansion. TensorWave intends to announce similar news later this year.

For more information on AMD accelerators and their applications in artificial intelligence, visit the AMD website.

FAQ

What are AMD accelerators?
AMD accelerators are advanced processing units designed to enhance the performance of graphics processing and artificial intelligence tasks.
How do AMD accelerators differ from Nvidia accelerators?
AMD accelerators offer advantages such as availability for purchase and superior specifications, including greater memory capacity and data throughput compared to Nvidia counterparts.
What technologies does TensorWave plan to implement for optimizing AMD accelerator performance?
TensorWave plans to utilize RoCE (RDMA over Converged Ethernet) technology for accelerating deployment processes and assessing the efficiency of AMD accelerators.