Sapion is launching the 'X330', an AI semiconductor for inference that offers up to four times the performance of its predecessor, the X220, and is making a full-scale attack on the data center market.

SAPEON, a global AI semiconductor company, announced at the 'SK Tech Summit' held on the 16th that it will launch the 'X330', an AI semiconductor for data centers.

Sapion CEO Ryu Su-jeong held a press conference at SKT Tower on the 15th and said, “This is a strategy to accelerate the attack on AI service model development companies and the data center market with the X330, which maximizes the advantages of the X220.” He added, “By increasing the utilization of AI semiconductors across all industrial fields and providing advanced AI technology that anyone can use at a low cost, we aim to contribute to creating a society where everyone can enjoy the benefits of advanced technological advancements.”

AI semiconductors have different required characteristics depending on the implementation purpose. It is for training AI models, and for inference to infer new data based on the trained AI model.

The X330 developed by Sapion this time is an inference NPU that has secured more than 4 times the computational performance and more than 2 times the power efficiency compared to the X220. It was produced through TSMC's 7nm process. Sapion said, "This is a product that targets NVIDIA's L40S, and according to our own performance measurements, the computational performance is about 2 times better and the power efficiency is more than 1.3 times better."

The existing Sapion X220 has proven its superior performance in inference of 'BERT', the first high-performance AI language model in Korea, and the X330 has achieved execution of LLM (Large Language Model) based on 'Transformer', the original technology of ChatGPT.

“Sapion is a smaller market than NVIDIA, so we’re more focused on inference than general-purpose,” said Michael Shevano, CTO of Sapion. “We plan to focus on multiple training areas in the long term.”

Unlike the X220, the X330 is considering all of SK Hynix's GDDR6 products, so the collaboration with SK Hynix is also noteworthy. In particular, Sapion plans to introduce various leading technologies such as HBM, chiplet, and CXL to keep pace with the trend of cost-effectively making its chips lightweight for generative AI models. The X430 is scheduled to be released at the end of 2025, followed by the X530 product. Sapion said, "The 5nm process will be applied to future models, and there will continue to be significant improvements in memory bandwidth."

In addition, Sapion has achieved AI service results through various collaborations. NHN Cloud and Korea's first AI semiconductor-based cloud infrastructure have been built, and in addition to the data center, we are collaborating with SKT through the 'NPU Farm'.

This is a farm built using X220-equipped servers within the SKT and SKB Gasan IDC, showing a processing capacity of 7.6 petaflops and a power-to-performance ratio of up to 7.7 times. CEO Ryu said, “This is the first case in Korea where we have built a large-scale farm and confirmed the possibility of commercializing services such as image analysis, natural language processing, and image quality improvement.” Sapion also announced that it recently signed an agreement with a large domestic hospital.

■ X330 with significantly enhanced performance, “ Expected to be applied across all industries”

Compared to its predecessor, the X330 has a significantly expanded range of applications based on standard technologies, making it applicable to a variety of industries.

The X330 is equipped with four AI cores consisting of 64K MXC and 4NVP, and 16 RISC-V-based CPUs. It has eight 16GB GDDR6 from SK Hynix, and each is controlled by its own GDDR controller. It provides a total of 16GB of memory and 256GB/s of memory bandwidth. It also supports PCIe Gen 5 interface that can communicate with the host processor.

In addition, the X330 has a built-in video codec and video post-processing IP to improve the processing speed of video-related programs. The X330 can process 4-channel 4K 60fps video input through the built-in hardware IP. The X330 supports floating point operations and has high expandability.

In particular, Sapion has rebranded its high-performance lineup with two chips this time as the 'Prime Card'. This ensures the highest level of hardware performance that meets more than twice the performance of compact cards, enabling various calculations with just one card. Meanwhile, power consumption relative to performance has been minimized.

“The Prime card has a bandwidth of 512 GB/s, allowing for the simultaneous movement of numerous parameters between memory and AI cores on models of various sizes,” said CEO Ryu. “It can support all LLMs of that level with a single card, and it can also support larger sizes with multi-cards.”

Sapion supports a software stack based on 'ONNX (Open Neural Network Exchange)' that can optimize performance when mounted on a server together with the X330 semiconductor HW. AI inference platform software and SDK (software development tool) are also provided. In addition, a service is provided to support immediate use when replacing Sapion on existing GPU-based servers such as multiple servers and clusters.

According to Sapion, the X330 is expected to start selling from the end of this year, and significant sales are expected to be generated next year when the X330 goes into full-scale production.

CEO Ryu said, “The global AI data center server size will increase by about four times from the current number to 6.19 million by 2027. In particular, the inference market is expanding. Sapion will continue to focus on developing high-performance processors, and in the future, we plan to introduce IP (semiconductor design assets) for autonomous vehicles and AI NPU for high-performance edge devices such as CCTVs.” For example, the 'X340' is an autonomous driving IP modified from the 300 series, and is expected to be released next year.

Meanwhile, Sapion plans to apply the X330 to the infrastructure area, which is the foundation of the three areas of AI infrastructure, AIX, and AI services in the 'SKT AI Pyramid' strategy for SKT's leap forward as a 'global AI company', and to expand into related businesses that can continuously create new business opportunities.