Customers can purchase RTX 4090 servers from us and choose to host them in a standard IDC in Taiwan or in a self-built data center, depending on their budget. We can also assist customers in planning UPS systems or large-scale GPU clusters with thousands or tens of thousands of GPUs (in Southeast Asia). Our team has extensive experience in data center construction and robust system integration capabilities.
Chinese companies are collecting RTX 4090 GPUs from around the world to extract chips and memory, then reassemble them into turbine fan models. Servers for inference widely use 4060 Ti and 4070 GPUs for large-scale clusters with tens of thousands of cards. This configuration currently offers one of the best Tokens/$ performance ratios, sacrificing some precision while maintaining usable performance, which could be key to making large model inference profitable.
With PCIe 5.0, the increased bandwidth allows for PCIe extensions to support more devices and additional RTX 4090 GPUs, while maintaining the benefits of multiple GPUs. This type of inference or training server is significantly more cost-effective compared to cloud-based solutions, with a single unit available for just a few thousand dollars. Even if our customers purchase just one unit, we can provide excellent hosting services.
Although GPU manufacturers cite reasons like cooling and noise reduction to justify producing oversized graphics cards, aiming to differentiate enterprise from consumer products, advancements in each generation's chip fabrication process have endowed these GPUs with impressive FP32 computing power. Various AIC manufacturers still strive to produce similar products for key clients. We assist our clients in deploying software and AI applications to the most suitable GPU environment. For further assistance, feel free to contact A20.
Since schools and development teams require substantial GPU resources to validate AI ideas, cloud service providers sell containerized services at prices up to ten times higher than the hardware cost. Purchasing servers directly or renting bare-metal servers is currently a more cost-effective solution.
This type of machine is large and not well-suited for standard IDC hosting. We can adjust configurations according to customer needs, even omitting UPS and redundant power supplies. However, due to Taiwan's relatively stable power supply compared to Southeast Asia, this approach allows for very affordable hosting costs. Our facility also maintains excellent humidity control, ensuring the machines remain stable. These servers can accommodate any standard consumer GPUs, such as the RTX 4090, and future models like the 5090 or 5090 Titan.