Exclusive: Nvidia's new China-focused AI chip set to be released at a similar price to Huawei offering.
Nvidia has commenced pre-orders for its new China-specific artificial intelligence (AI) chip, known as the H20, through distributors who are pricing it in alignment with a competing product from Huawei, according to sources familiar with the matter.
Developed in response to the expanded U.S. bans on high-end chip exports to China, the H20 graphics card is the most robust among three Nvidia offerings tailored for the Chinese market. This strategic move aims to impede China's progress in developing AI and advanced computing capabilities, particularly for military applications.
In contrast to Nvidia's flagship H100 AI chip and the H800, the H20 naturally provides lower computing power. However, an examination of the specifications indicates that the H20 may also lag behind Huawei's Ascend 910B in certain critical areas, as suggested by three undisclosed sources.
Nvidia has been quoting prices for H20 orders to distributors in China within the range of $12,000 to $15,000 per card, as reported by two sources. Some distributors are advertising the chips at a significant markup, approaching 110,000 yuan ($15,320), which is slightly below the price of Huawei's 910B, currently available at around 120,000 yuan.
Distributors are also offering H20 servers pre-configured with eight AI chips for 1.4 million yuan. Comparatively, servers equipped with eight H800 chips, launched a year ago, were sold at approximately 2 million yuan.
The delivery timeline for H20 products is expected to begin in small batches during the first quarter of 2024, with larger quantities becoming available in the second quarter, as per sources.
Prior to the U.S. restrictions, Nvidia held a dominant position in China's AI chip market, boasting over 90% market share. However, the company now faces heightened competition from domestic rivals, with Huawei's 910B chip emerging as a leading AI solution within China.
Examining specifications, the H20 appears to trail the 910B in FP32 performance, a critical metric measuring chip processing speed, registering at less than half of its competitor's capability. Nevertheless, the H20 demonstrates an advantage in interconnect speed, crucial for efficient data transfer between chips.
This positions the H20 competitively with the 910B in applications requiring the linking of multiple chips into a cohesive system, as outlined by one source.
In a recent report, Reuters revealed Nvidia's plans for mass production of the H20 in the second quarter of the current year. Originally slated for launch in November of the previous year, the delay was attributed to integration issues faced by server manufacturers.
Nvidia's strategic focus on the Chinese market was underscored by CEO Jensen Huang's visits to the company's offices in Shenzhen, Shanghai, and Beijing last month during the annual Lunar New Year celebrations, as reported by local media.