Production of DeepX's next-generation neural processing unit (NPU), a South Korean AI chip developer that also uses the same foundry as Tesla, is expected to face delays due to changes in production schedules at Samsung Electronics' foundry.

DeepX second-generation NPU DX-M2

Industry sources revealed on Sunday that multi-project wafer (MPW) production of DeepX's second-generation NPU DX-M2, originally scheduled for April, has been delayed by approximately six months. 

MPW is a service that allows the production of multiple customer semiconductor prototypes on a single wafer, enabling companies to spread manufacturing costs and reduce prototyping expenses.

DeepX initially planned to begin mass production of the DX-M2 in the second quarter of next year. Due to delays in MPW production, the mass production timeline has also been postponed. Quality testing is now expected to begin no earlier than the third quarter, with full-scale sales potentially not until the fourth quarter of next year. Therefore, revenue contributions from already mass-produced chips (such as the DX-M1) are particularly crucial to filling this gap.

Industry sources indicated that this delay stems from Tesla's production plans. DeepX's DX-M2 is the first external customer chip to utilize Samsung's 2nm process. Tesla is also using the same 2nm process at Samsung's foundry to develop its AI6 AI chip. According to industry sources, the delay in Tesla's MPW chip production has also impacted DeepX's schedule.

Tesla's MPW project delayed

The specific reasons for the delay in Tesla's MPW project have not been disclosed. Industry insiders speculate that factors such as the mass production timelines for self-driving cars and humanoid robots, as well as the progress of investment in supercomputers, may have played a role.

A Samsung Electronics official declined to comment, stating that the company could not confirm matters related to its clients.

Tesla recently sent a senior purchasing executive to Samsung Electronics to discuss expanding its AI6 chip foundry capacity. The previously agreed capacity was approximately 16,000 wafers per month, while the newly discussed capacity is approximately 24,000 wafers per month. If the new agreement is finalized, the total capacity is expected to reach approximately 40,000 wafers per month.

The DX-M2 is DeepX's generative AI accelerator. Its predecessor, the DX-M1, was optimized for visual AI applications, while the DX-M2 is designed specifically for AI data center workloads. This chip can process models with up to 100 billion parameters.

The processor is expected to consume up to 5 watts of power and deliver a computing performance of 80 TOPS (trillion operations per second). It supports the low-power DRAM standard LPDDR5X memory.

According to industry insiders, DeepX has not yet received orders for the DX-M2. NPU chips are typically ordered only after customers have tested sample chips. Therefore, substantial revenue is not expected until mass production begins next year.

DeepX's customers include Samsung Electronics (Exynos 2100 processor), Hyundai Motor Group's Robotics Lab (MobED), and Intel (some Core i5 processors). The company recently supplied Baidu with 40,000 chips (DX-M1) and modules (DX-M1+M.2).

The DX-M1 chip is priced at approximately $20 to $50, while the module is priced at $50 to $100. Based on these prices, this shipment is expected to generate between $800,000 and $4 million in revenue, representing approximately 5% to 24% of DeepX's projected $17 million in revenue this year.

Tesla gambles heavily on Samsung

Tesla is reportedly in talks with Samsung Electronics to expand production of its next-generation AI6 artificial intelligence chip, according to sources familiar with the matter.

Industry insiders say Tesla's senior purchasing executive is expected to visit Samsung this week to discuss increasing foundry output for the 2-nanometer AI6 chip.

Tesla has requested increased capacity from Samsung. The original contract signed last year stipulated a supply of approximately 16,000 wafers per month. Tesla is now seeking to increase capacity by approximately 24,000 wafers. This visit aims to discuss specific supply terms based on this request.

If the deal is finalized, total monthly production is expected to reach approximately 40,000 wafers. Such high production would significantly improve the initial capacity utilization of Samsung's new wafer fab in Taylor, Texas.

Tesla signed a foundry agreement with Samsung last year to manufacture the AI6 chip. Under the contract, chip supply will continue until December 31, 2033, and the contract is valued at approximately 22.8 trillion won (about US$17 billion). Given the eight-year contract, analysts previously estimated Samsung would earn approximately 2 trillion to 3 trillion won annually. The total contract value could be even higher, as the additional orders under discussion would exceed the initial order.

The AI6 chip is expected to power multiple Tesla platforms. Industry insiders say it could be used not only in Tesla's Autopilot system but also in its humanoid robot Optimus and internal AI data centers.

This chip cluster is also expected to replace the role originally envisioned for Tesla's Dojo AI supercomputer project, which was effectively suspended last year. The AI6 processors will likely be deployed in a cluster configuration, with multiple chips mounted on a single server motherboard.

As Tesla expands its Autopilot service and robotics development, the company has been significantly increasing its investment in AI infrastructure. In its January earnings call, Tesla stated it planned to increase capital expenditures this year to over $20 billion, a record high. This figure is roughly double its recent annual investment expectations ($8 billion to $11 billion).

Tesla's semiconductor collaboration with Samsung

Tesla's semiconductor collaboration with Samsung dates back to 2019. At that time, Samsung's System LSI division participated in the design of Tesla's custom application-specific integrated circuit (ASIC), while its foundry manufactured the chip using a 14-nanometer process. This chip was later named HW3, also known as AI3.

The HW4 (AI4) chip currently used in Tesla vehicles is also manufactured using Samsung's 5-nanometer process. Tesla previously outsourced the production of its AI5 chip to both TSMC and Samsung, but the company again chose Samsung for the production of its AI6 chip.

Tesla has consistently pursued a strategy of "no sourcing from China, no sourcing from Taiwan" for key components, explicitly urging partners to exclude suppliers from these regions. This reflects the increasingly tense geopolitical situation between the US and China, and concerns about risks related to the Taiwan Strait. Samsung's factory in Taylor, Texas - a state-of-the-art manufacturing facility in the US - aligns perfectly with this strategy.

If this additional order is finalized, Samsung's foundry division will gain significant access to the production of cutting-edge semiconductor products. An industry insider stated that while final confirmation and production plans are still subject to change, this deal could mark a significant recovery for Samsung's foundry business.

The collaboration between the two companies extends beyond chip manufacturing. Samsung's System LSI division has completed the development of a 5G modem for Tesla vehicles and plans to begin shipping it in the first half of this year. The initial products are expected to be used in Tesla's self-driving taxi fleet in Texas, before being rolled out to more models. The scope of the collaboration between the two companies continues to expand.

Source: Compiled from thelec

View more at EASELINK