US chip giant Nvidia is gearing up to mass produce a range of AI chips including its H20 model for China in the second quarter of 2024, with Taiwanese electronics manufacturer Wistron picked as the major supplier of these GPU (graphics processing unit) substrate orders, according to US media outlet Wccftech.

Why it matters: Despite the US government’s ban on high-performance AI chip exports, Nvidia has developed modified AI chips for the Chinese market and is ready to ship limited sales of less powerful AI chips to China as long as they adhere to regulatory standards.

Details: Nvidia is modifying its latest AI chips specifically for the Chinese market, including the HGX H20, L20 PCle, and L2 PCle. The three chips are based on the Nvidia H100, but allow the company to comply with the latest US export control policies announced last October.

• Nvidia originally planned to start selling a modified version of these AI chips by the end of 2023, but postponed the launch to early 2024 as the China-US chip war escalated, the Wccftech report said. 
• Since then, the US government has adopted a softer stance on commercial sales in China, and Nvidia has assured the government that their chips will fully meet compliance guidelines established by US trade and commerce authorities. Last week, Nvidia launched its China-exclusive GeForce RTX 4090D, a less powerful version of the flagship RTX 4090 GPU it sells elsewhere.
• The H20 AI chip is a scaled-down version of the H100 GPU, equipped with 96 GB memory capacities that operate at speeds of up to 4.0 Tb/s, according to Nvidia. It offers computing power of 296 TFLOPs and a performance density of 2.9 TFLOPs/die, in contrast to the H100’s 1,979 TFLOPs and 19.4 TFLOPs/die.

Context: Last October, the US government issued a new ban, further restricting the export of high-performance AI chips by Nvidia. Subsequently, the chip giant announced an immediate halt to the shipment of its A100, A800, H100, H800, and L40S products.

• In December 2023, during an interview with Reuters, US Secretary of Commerce Gina Raimondo stated that the government was in discussions with Nvidia about allowing the chip firm to sell AI chips to China under certain conditions. Raimondo emphasized that the government would stop Nvidia from exporting its most complex and powerful AI chips, but would not curb its chip sales to China altogether. 

On Tuesday, US chip giant Nvidia revealed that the US government has instructed it to immediately halt the export of certain high-end artificial intelligence chips to China, as regulators have expedited the enforcement of new restrictions, which were originally scheduled to take effect from Nov. 16.

Why it matters: The US withdrew Nvidia’s 30-day exemption period for chip exports to China on Oct. 23, implementing the new regulation 24 days earlier than expected. Currently, Chinese customers, such as Tencent and ByteDance, can no longer obtain any AI-related products from Nvidia. 

Details: The US government notified Nvidia of the immediate implementation of export restrictions on AI chips starting Oct. 23, as per Nvidia’s announcement to the US Securities and Exchange Commission. The affected products include five GPUs (graphics processing units): A100, A800, H100, H800, and L40S.

• On Oct. 17, the US Department of Commerce announced a series of new restrictions on chip exports, addressing loopholes identified after the US imposed export limitations on chips last October. These new regulations broaden the definition of advanced AI chips and impose additional licensing requirements on chip products destined for over 40 countries and regions, aiming to prevent resale to China. According to the regulations, the new restrictions will take effect from Nov. 16.
• Following the implementation of the new regulations, Nvidia must cease shipping A800 and H800 chips to China unless it has express permission from the US government. A800 and H800 chips are alternative solutions Nvidia offered in place of the originally prohibited A100 and H100 chips, following the initial AI chip export restrictions imposed by the US last October. The Nvidia L40S, an advanced GPU for data centers, will also be affected by the new restrictions.
• Due to the widespread global demand for these products, the sudden acceleration of US restrictions is not expected to have a short-term impact on the company’s financial performance, Nvidia said in its announcement. 
• However, Nvidia CEO Jensen Huang earlier stated that the new ban is expected to significantly impact Nvidia’s sales in the Chinese market, though he added that the company remains committed to complying with US regulations. 
• Several AI industry professionals in China have voiced concerns and doubts over the new measures, according to a report by local media outlet TMTPost. They worry about the potential impact of the new restrictions on the future training of large-scale AI models in China, which could result in a tech gap compared to US-based AI companies like OpenAI.

Context: In August, Nvidia reported a fourfold increase in its Data Center revenue over the last two years, establishing itself as a leader in AI chips with a market share of over 70%. Nvidia’s stock continues to rise, as the company achieved a market cap of $1 trillion earlier in 2023.

• The data center sector comprises processors such as central processing units (CPUs), data processing units (DPUs), and graphic processing units (GPUs). GPUs are favored for AI applications due to their ability to handle multiple tasks simultaneously.

TSMC is urgently seeking equipment suppliers from whom it can buy CoWoS (Chip on Wafer on Substrate) machines, as Nvidia, AMD, and Amazon expand orders for AI chips, local media outlet Economic Daily News exclusively reported on Monday. TSMC has increased its equipment orders for CoWoS by 30% to meet growing AI-fueled demand, the report claimed.

Why it matters: The AI boom has reshaped the semiconductor landscape, positioning chip-making companies as critical enablers of the tech revolution while presenting complex challenges related to supply chain resilience and technological advancement. 

Details: CoWoS is a high-density advanced packaging technology developed by TSMC for high-performance chips. The current shortage of CoWoS packaging capacity has become the main bottleneck in the production chain for AI chip orders.

• Nvidia is TSMC’s largest customer of CoWoS advanced packaging, accounting for 60% of the production capacity, according to the report. Recently, in response to the strong demand for AI computing, Nvidia has increased its orders from TSMC, while AMD, Amazon, and Broadcom have also been placing urgent orders.
• Due to its shortage of CoWoS machines, TSMC has sought the assistance of local equipment suppliers including Scientech Corporation, AllRing-Tech, Grand Process Technology, E&R Engineering Corporation, and Group Up Industrial. TSMC has increased its original equipment orders by up to 30%, with delivery expected in the first half of 2024. Mass production will commence in the second half of 2024.
• Currently, TSMC’s monthly capacity of CoWoS advanced packaging is approximately 12,000 units, sources familiar with the matter told Economic Daily News. To accommodate the proposed production expansion, the monthly capacity of CoWoS will be increased to between 25,000 and 30,000 units.
• Major customers such as Nvidia and AMD have increased their orders of TSMC wafers in the third quarter, boosting TSMC’s 7nm and 5nm advanced process capacity utilization.
• Pressure on CoWoS production capacity is expected to ease after next summer, TSMC’s president C. C. Wei stated during a recent earnings call. The Economic Daily News report revealed that TSMC has been expanding its facilities at plants including Zhuke, Zhongke, Nanke, and Longtan, so as to increase CoWoS production capacity.

Context: On September 24, Chinese media outlet IThome reported that Qualcomm’s next generation Snapdragon 8 Gen 4 may have been manufactured using TSMC’s N3E process technology, as indicated by leaked documents from Qualcomm.

• On September 8, TSMC announced net revenue for August 2023 of approximately NT$188.69 billion ($5.87 billion), representing an increase of 6.2% from the previous month.

On Monday, leading chip manufacturer TSMC started pre-production operations on its 2nm process, according to Taiwan-based media outlet Economic Daily News. Sources revealed that the chip maker will use an advanced AI system to improve energy efficiency and accelerate efficiency. Apple and Nvidia are expected to be among the first batch of customers for the Taiwan-headquartered company’s 2nm production, putting significant pressure on its competitors such as Samsung.

TSMC did not comment on specific details in response to the report, but stated that the development of 2nm technology is progressing well and aims to hit mass production by 2025.

Why it matters: TSMC’s 2nm plan could bring it into head on competition with Samsung. Samsung, which beat TSMC to be the first chip manufacturer to widely employ 3nm processes, also announced last year that it expects to have 2nm chips in mass production by 2025.

Details: According to the report, TSMC is estimated to begin test producing hundreds of 2nm chips this year, laying the foundation for mass production in 2025. The 2nm production base will be established at TSMC’s Fab 20 at Hsinchu Science Park, with subsequent expansion to Taichung Science Park, encompassing a total of six phases of engineering. 

• TSMC has already begun to dispatch engineers to the Hsinchu Science and Industrial Park for the 2nm process, building a research and development team of over 1,000 employees, according to Economic Daily News. 
• In the initial stages, TSMC will set up a small trial-production line at its Hsinchu headquarters, with a target of taping out around 1,000 2nm chip pieces by the end of this year, the report said. 
• On April 26, TSMC showcased its latest technology developments at its 2023 North America Technology Symposium, including in the fields of 3nm and 2nm technology. 
• The 2nm process will be the first process node at TSMC to employ gate-all-around (GAA) transistors, nanosheet transistors which TSMC says will boost performance, energy efficiency, and transistor densities. The company claims this will offer an improvement in speed of up to 15% over 3nm at the same power, and up to 30% power reduction at the same speed.
• Nvidia’s CEO Jensen Huang previously said at GTC 2023 that TSMC will bring the AI system that Nvidia and TSMC and other partners worked on into 2nm trial production this June. The system promises a computation time that is 1/40th of its predecessor.

Context: Samsung expects to have 2nm chips in mass production by 2025 and plans to deliver mass production using a 1.4nm process by 2027.

• TSMC reported record revenue for 2022 of NT$2.26 trillion ($74.25 billion), a rise of 42.6% from 2021. It was the first time the world’s largest contract chip manufacturer reached NT$2 trillion in revenue for a year.
• The Apple iPhone 15 Pro will be powered by the A17 Bionic chip based on TSMC’s 3nm technology, according to product review site Tom’s Guide. The chip is also expected to debut in this year’s upcoming MacBook models.

Top GPU maker Nvidia told Caixin (in Chinese) on Wednesday that they will provide alternative GPU products to Chinese clients as many of them face imminent supply issues after the US government recently announced export bans on cutting-edge chips. 

Why it matters: The export ban from the US was aimed at limiting China’s expansion in AI and other tech fields that need high-performance GPU chips. 

• California-headquartered Nvidia is one of two major GPU makers (the other being AMD) that are particularly affected by the ban. 

Details: Jensen Huang, CEO of Nvidia, told Caixin and other media outlets that the alternative GPU chips will be built with their new Hopper architecture for Chinese clients, which enables them to sell the hardware without violating the export ban.

• The Hopper architecture is adopted in the coming 4nm H100 GPU, which is nine times more powerful in large-scale training than the previous generation, the A100, according to Nvidia. However, the company has yet to reveal what it will do to bypass the ban and ensure the product can meet the new regulations.
• The alternative chips can “fulfill most of the needs in China,” while for those clients who absolutely need the banned products, Nvidia could attempt to apply for licenses, Huang added.
• Huang emphasized that China is an important market to Nvidia and many important partners and clients are based in the country, saying he believed “the market will still bring increased opportunities to the firm.”

Context: The mainland Chinese market accounted for 26% – or $7.11 billion – of Nvidia’s global revenue during its 2022 financial year, which ended on Jan. 30, 2022.

• Nvidia’s revenue from the mainland Chinese market is more than 1.5 times that of the US market during the same period.
• Last week, Taiwanese media outlet UDN reported that Nvidia had sent urgent orders to TSMC for banned GPU chips, including A100 and H100.

Nvidia has unveiled a line of new products and announced a ten-year collaboration with the Taiwanese government at Computex 2018 pre-show press conference.

“AI is the most powerful technology force of our time,” Jensen Huang, founder and CEO of Nvidia, said. “Its first phase will enable new levels of software automation that boost productivity in many industries. Next, AI, in combination with sensors and actuators, will be the brain of a new generation of autonomous machines.”

Among the most exciting news is the launch of Nvidia ISACC robotics platform, which Huang said will be used to power the new generation of autonomous machines, bringing AI capabilities to manufacturing, logistics, agriculture, construction and other industries.

Nvidia also introduced Jetson Xavier, which is said to be the world’s first intelligent machine processor designed for robotics. With the size not much bigger than the palm, Jetson Xavier will be priced at $1299. A workstation equipped with GPUs that has roughly the same processing power usually cost around $10,000. The Jetson Xavier Devkit will begin early access in this August.

Nvidia also announced today that it is deepening its collaboration with Taiwan’s Ministry of Science and Technology (MOST) to advance Taiwan’s AI capabilities. The company said the ten-year partnership aims is to build up local deep learning and AI-related technologies. The two will collaborate in multiple areas to advance AI application in local industries and accelerate research and development on the island. Huang said the collaboration “starts with building a supercomputing infrastructure in Taiwan,” adding that the two will also be working on self-driving cars as well as intelligent machines.

Last year, Nvidia unveiled the collaboration with the Taiwanese government to accelerate the development of AI in the island’s commercial sector in support of the “AI Grand Plan,” which aims to help foster domestic AI-related industries.

Computex, Asia’s largest tech expo that opens on June 5th in Taipei, is a springboard for some of the year’s most exciting tech launches and announcements. During the five-day expo, global tech leaders will take the opportunity to flaunt their new chips, gaming gear, VR, laptops and more.

Baidu has launched an autonomous driving ecosystem with 50 partners at its first AI developers conference in Beijing. At its heart is the US-developed code for controlling the vehicles, but the scope of Apollo 1.0 is to create an entire ecosystem encompassing research universities, components makers such as NVIDIA, navigation developers such as TomTom, and car manufacturers including Ford, Daimler and FAW (Volkswagen’s joint venture partner in China).

The federation approach is radically different to that of the traditional manufacturers and is expected to allow more companies to participate. “It even allows a range of different business models to operate within the ecosystem,” said Baidu COO Lu Qi.

The code for Apollo 1.0 is completely open-source and will be available on Github. Documentation will be updated weekly and the code fortnightly with overhauls planned for September and December—when fully autonomous urban driving is expected to be achievable.

Baidu co-founder and CEO, Robin Li, introduced the new ecosystem via a live link up to his driverless car as he headed to the conference along Beijing’s fifth ring road. The 4- to 5,000-strong audience was also shown a world-first: a video of two autonomous cars driving in the same test pen (which we later experienced for ourselves).

Another important part of the plan is Apollo’s Simulator Engine. The program uses real data about roads and junctions to create a simulation for virtual cars running on Apollo. A demonstration at the conference showed a simulation of a car crashing at an intersection and then how the code would be fixed and uploaded for the Apollo team to check before being added into the overall source code. This way “. . . Apollo can be tested over millions of kilometers every day,” said Lu, who estimated around 10 billion kilometers of testing is needed for an autonomous vehicle system, meaning Apollo’s R&D will soon accelerate beyond that of the competition.

Development

“We partnered with Baidu through a mutual client of ours in Silicon Valley and Baidu talked to us about creating the base platform for the Apollo project,” Josh Whitley, a software engineer at California-based AutonomouStuff who had come to Beijing to install his company’s software on the vehicles at the conference, told TechNode. “The Lincoln MKZ that they have here, the computing platform, they’re all provided by AutonomouStuff.”

Whitley managed to install the software on the Lincoln’s drive-by-wire system and test and tune it in just three days, a process that would normally take a dozen workers six months.

“The Apollo software is very flexible, made to accommodate different vehicles very easily. The feature set is mainly for recording a GPS-based route. [Apollo] is definitely better at a specific set of things [than other platforms],” said Whitley. “Part of the core infrastructure is a safe run-time environment—a real-time operating system—that won’t skip any commands or be delayed waiting for vehicle catch up.”

“The intent, for the Chinese market and eventually for other markets, is to make it a unified software platform for all the Chinese automakers and then others to use,” added Whitley.

Application

“China is very much about one solution in general. Think of WeChat – there’s one solution. Didi – one solution,” Lei Ma, a senior product manager of autonomous driving at Baidu, based in Silicon Valley, told TechNode. “We’re hoping that Apollo becomes that one solution for autonomy.”

According to Ma, Baidu will make no claim on any use of its source code, however, it is used: “People are free to take Apollo, modify it or not, put it on a car and say ‘we’re selling autonomous vehicles’. Baidu does not lay claim to revenue, data, intellectual property. They can take it and commercialize it, anywhere in the world… Of course, if you work with Baidu, we can make things move a lot faster.”

The nature of establishing an ecosystem rather than a closed garden set up means the system is expected to accelerate, according to Lu Qi: “In 3 to 5 years China will lead the world in autonomous driving.”

Baidu’s AI operating system, DuerOS, will be fundamental to the application of Apollo 1.0. “DuerOS means that Apollo could be compatible with different cars from different manufacturers, or you can build your own,” explained Lu.

Ma explained the ecosystem’s development within the China context. “Back to the ‘one solution,’ whoever creates that ecosystem—the biggest, the fastest—is going to be the single player. I personally think it’s going to be a winner takes all solution. There’ll be a first place, a second perhaps, and maybe only a very different third.”

Speaking of Didi as a ‘one solution,’ the ride-hailing company’s logo was absent from the display of partners at the launch, so we asked Ma if we can expect to hear anything soon. “I think Didi is interested, but they’re taking a wait-and-see approach. A lot of companies are. The companies we announced today are not the only companies we talked to.”

Test ‘Drive’

The two Lincoln MKZ’s running on Apollo 1.0 were available for us to take a ride. But before that, we took a spin in a Haval running on Baidu’s software- and hardware-based advanced driver assistance system (ADAS).

The ADAS is Level 3 in terms of autonomy, which means it’s assisted driving rather than a fully autonomous Level 4 system such as Apollo 1.0. The Haval SUV had been programmed to run a particular route through the tires but when a helper put a sign in the middle of the road it changed course. The ride was jiddery, as though to take a bend the car breaks the curve into a series of short segments.

“Comfort is, of course, going to be a very important factor in terms of commercialization, but it’s not the biggest priority [at the moment],” explained Lei Ma.

“There are limits to what the car will let us do. Going forward we’re hoping to work with a lot of car makers and have access to their drive-by-wire interfaces so we can calibrate those controls and make things comfortable, basically the same as a human driver,” said Ma.

“Put your seatbelt on,” said the otherwise redundant human driver when we switched to the Lincoln MKZ. Running on Apollo 1.0, the difference to the ADAS was palpable. It felt much more like a human was driving, though the car still went into corners at quite a pace like the Haval (perhaps we’re just more cautious on the corners).

“We’re trying to smooth out all the movements,” explained the engineer in the back, who explained he was there “to press the start button.”

The other MKZ made its own loops in the test pen and we can happily report that the two vehicles, though driving around the same tracks and coming close together, did not even come close to any sort of collision.

Conversations around virtual reality are quickly centering around one of the biggest bottlenecks in the industry’s development: high-quality content.

The two-day Global AR/VR China Summit, which kicked off on Wednesday, featured speakers from different corners of the industry, from virtual reality analytics company Touch Virtual to Neobear, an early education startup developing augmented reality products. The hype around VR hardware seems to have petered off with most presentations focused on building exciting and useful VR content in marketing, education, travel, and social applications.

“We have to choose PC-based VR [over mobile-based VR] because our client’s expectations for high quality content,” said Yuan Yuan, the CEO of AR/VR software company Ugion, on a panel discussing the current and future adoption of VR and AR technology.

“The most important thing is there’s no killer-app,” he said. ” There’s no killer app to attract users to the headset.”

At the moment, VR content still leaves a lot to be desired. While a multitude of VR headset and hardware startups have sprung up in China, the same cannot be said about quality VR content and software. Still, as this year’s conference shows, there’s plenty of interest and potential in the area – it’s only a matter of time before VR content hits its stride.

Here are some highlights from Global AR/VR China Summit 2016:

Multiple users can interact with and import 3D models into MiddleVR‘s virtual meeting space software aimed at designers and architects. These two guys are measuring a CAD model.

Plex VR creates custom 360 degree content for shops, museums, and real estate developers.

Conference attendees lined up to try Microsoft’s HoloLens AR headset.

Scott Lai, the marketing and sales director of Realmax, explains the company’s AR training software.

Behavioral research firm Noldus is jumping on the VR bandwagon with eye tracking and VR interaction analytics.

A conference attendee tries Nvidia’s Funhouse app where users can play a range of carnival games, like whack-a-mole and arrow shooting.

Baidu is hoping to take its autonomous car project to the next level through a partnership with high performance chip maker Nvidia.

Baidu CEO Robin Li and Nvidia CEO Jen-Hsun Huang spoke together on stage at a Baidu event in Beijing last week. According to a blog post by NVIDIA, Baidu will contribute their cloud platform and mapping technology while NVIDIA will offer their self-driving computing platform and HD mapping solutions.

The partnership also renews the Chinese search engine’s competitive edge against local internet firm LeEco, which has been working on a similar cloud-based ecosystem for autonomous cars.

“We’re going to bring together the technical capabilities and the expertise in AI and the scale of two world-class AI companies to build the self-driving car architecture from end-to-end,” said NVIDIA in the blog post.

Together, the pair are hoping to develop autonomous parking and ‘level three’ autonomous vehicle control, which is the last level before fully autonomous (there are five levels all together).

Baidu has been rapidly expanding their testing grounds for autonomous cars both locally and abroad. The tech giant plans to establish ten local testing locations in China by the end of the year (current locations include Beijing, Wuzhen and Wuhu cities). Baidu recently received the go ahead from motor vehicle authorities in the U.S. to test vehicles alongside Google’s autonomous cars in California.

It’s not the first time Baidu and Nvidia have joined forces. According to Nvidia, the chip company has contributed components for other projects undertaken by Baidu’s artificial intelligence research unit, which is headed by Andrew Ng and based in the U.S.

Baidu recently joined automaker Ford to invest $150 million USD in detection technology for autonomous cars by Velodyne LiDAR Inc., as they continue to leverage partnerships in their goal to sell fully autonomous cars by 2018.

Alibaba Cloud, the cloud computing arm of Alibaba Group, announced on Wednesday a collaboration agreement with Accenture, a global professional services company and SAP China, a German multinational software corporation.

The partnership with Accenture, revealed during their cloud computing conference in Shenzhen, shows Alibaba Cloud’s global ambition to become a leader in service technology and business solutions for clients in China and in ASEAN markets.

“Together, [Alibaba Cloud and SAP] will help enterprises to simplify their IT infrastructure, realize business value and accelerate digital transformation,” Mark Gibbs, President of SAP Greater China said.

Through the collaboration Accenture, Alibaba Cloud will combine their infrastructure as a Service (IaaS) solutions with Accenture’s industry and technology consulting capabilities. The two companies will also work together to jointly bring cloud-based third-party solutions onboard the Alibaba Cloud platform.

“We are glad to work with Accenture to strengthen the offerings of reliable cloud-based, analytics-driven enterprise solutions in China as well as in ASEAN markets,” said Yu Sicheng, Vice President of Alibaba Cloud in a statement.

“Applying our broad industry expertise, analytic capabilities and technology integration services along with the IaaS functions of Alibaba Cloud will enable clients to accelerate their adoption of ‘as-a-Service’ strategies that prevail in today’s digital economy,” said Yu Yi, Managing Director and Lead of Accenture Digital in Greater China.

With the announcement, Alibaba Cloud adds two more multinational corporates to a slew of current partners, including Foxconn, who signed a strategic partnership with Alibaba Cloud in October 2015, and NVIDIA, the Graphic Processing Unit provider that signed a strategic partnership with Alibaba Cloud earlier this year.

Other Chinese tech giants have also extended into cloud services. Baidu Yun, a cloud computing arm of Baidu, jointly revealed China’s first internet data center (IDC) optical transport network that supports T-SDN last week, followed by a report that Tencent is seeking additional loan worth $2 billion USD to invest in gaming, intellectual property rights for entertainment and cloud computing.

Image Credit: Alibaba Cloud

One Stop “Big Data Platform”