[News] Alibaba T-Head PPU Lands China Unicom Deal, Reportedly Matches NVIDIA H20 Performance

According to EE Times China, during a report on China Unicom’s Sanjiangyuan data center, China’s CCTV News unexpectedly revealed a table titled “Comparison of Key Parameters Between Domestic and NV Cards.” The table highlighted how homegrown AI chips deployed at the data center — including Alibaba’s T-Head, Huawei’s Ascend 910B, and Biren Technology’s 104P — are making notable progress in closing the gap with international giant NVIDIA in critical performance metrics.

T-Head’s PPU Reportedly Surpasses NVIDIA’s A800, Closing in on H20

As the report highlights, Alibaba’s T-Head Parallel Processing Unit (PPU) has become the centerpiece of attention among those listed in the chart. The report notes that T-Head’s PPU reportedly surpasses NVIDIA’s A800 in key technical metrics. Equipped with 96GB of HBM2e memory, it matches the 96GB HBM3 capacity of NVIDIA’s H20, though the H20 uses the more advanced HBM3 technology. By contrast, NVIDIA’s A800 carries only 80GB of memory, the report adds.

In addition, as the report points out, Alibaba’s T-Head PPU achieves 700GB/s of inter-chip bandwidth, far exceeding NVIDIA’s A800 at 400GB/s and approaching the level of NVIDIA’s H20. The report also mentions that the more critical advantage lies in cost: leveraging domestic 7nm process technology and 2.5D packaging, the PPU’s bill of materials (BOM) per card is 40% lower than that of the imported H20, allowing Alibaba Cloud to cut public cloud inference prices by 50%.

Chinese Tech Companies Deploy Local Chips While Deepening AI Investment

As the report points out, China Unicom’s Sanjiangyuan data center, featured in CCTV news coverage, represents a major large-scale deployment of domestically produced AI chips. The report notes that Alibaba Cloud signed for 1,024 servers equipped with 16,384 T-Head compute cards, delivering 1,945 petaFLOPs of computing power — the largest order among all brands.

China has been stepping up its push to develop domestically designed chips. Reuters, citing The Information, reports that Alibaba and Baidu have reportedly begun training their AI models with in-house chips, partly substituting for NVIDIA hardware.

In parallel, Chinese tech giants are pouring heavily into AI investments. According to Reuters, Alibaba raised $3.2 billion last week through a major convertible bond issuance. The report notes that Alibaba plans to allocate nearly 80% of the proceeds to expanding data centers, upgrading technology, and improving services to meet rising demand for cloud solutions.

Read more

• [News] Alibaba Reportedly Unveils AI Chip Rivaling NVIDIA’s H20, Boosting Taiwan’s Server ODM Inventec
• [News] China Reportedly to Triple AI Chip Output Next Year With Fabs Serving Huawei, Cutting NVIDIA Reliance

(Photo credit: Alibaba)