Breaking the Bottleneck | Issue 59
[9/22/2024] Semiconductors, Steel, & Solidworks!
Breaking the Bottleneck is a weekly manufacturing technology newsletter with perspectives, interviews, news, funding announcements, manufacturing market maps, and a startup database!
💥 If you are building, operating, or investing in manufacturing, hardware, or robots, please reach out. My email is aditya@schematicventures.com – I’d love to connect!
🏭 If you were forwarded this and found it interesting, please sign up!
Content I Enjoyed Last Week 🏭🗞️🔬 📚
News:
Challengers Are Coming for Nvidia’s Crown [IEEE Spectrum]
NVIDIA is the dominant force in the AI chip market. Revenues have skyrocketed from approximately $12 billion in its 2019 fiscal year to $60 billion in 2024. NVIDIA's cutting-edge AI chips, such as the H100 and H200 GPUs, are in high demand and scarce supply, prompting an antitrust investigation by the U.S. DOJ. The company's success is due to its advanced hardware and robust CUDA software ecosystem. Released in 2006, it enables developers to harness the full potential of its GPUs. "They made sure every computer science major coming out of university is trained and knows how to program CUDA," says Matt Kimball. This extensive ecosystem makes NVIDIA hardware the path of least resistance for AI development, creating a significant barrier for competitors. Despite NVIDIA's dominance, several competitors are emerging with strategies to challenge its position.
AMD is releasing its upcoming Instinct MI325X GPU to potentially rival NVIDIA's H100 hardware. Microsoft, Meta, and Dell have deployed AMD's MI300X GPU in various applications. However, its open-source platform, ROCm, lacks the widespread adoption of NVIDIA's CUDA, and AMD is currently acquiring companies like Silo AI and ZT Systems to bolster its software capabilities.
Intel has a robust software platform, OneAPI, which spans CPUs, GPUs, and FPGAs, but its hardware lags behind. Intel's Gaudi AI accelerators, acquired from Habana Labs, have shown promise, with the latest Gaudi 3 offering performance competitive with NVIDIA's H100. However, uncertainty looms as Intel's next major AI chip, Falcon Shores, is not expected until late 2025.
Cerebras specializes in AI for supercomputers, offering the Wafer Scale Engine 3 (WSE-3)—a colossal chip containing 4 trillion transistors. This is significantly more than NVIDIA's largest GPU, the B200, which has 208 billion transistors. Cerebras's technology is niche and suited for large-scale AI training tasks like next-generation large language models.
SambaNova's SN40L chip features a reconfigurable dataflow architecture, allowing developers to optimize hardware configurations for different AI models. The company's software stack, SambaFlow, analyzes machine learning models and reconfigures hardware to accelerate performance.
Groq focuses on maximizing AI inference performance with its Language Processing Unit Inference Engine. Despite utilizing older 14-nanometer technology, Groq's chip demonstrates impressive inference speeds, exceeding 1,250 tokens per second when running Meta's Llama 3 8-billion parameter model.
Qualcomm focuses on AI inference and energy efficiency. Its Cloud AI 100 servers have significantly improved benchmarks like ResNet-50, increasing performance from 180 to 240 samples per watts.
The hyperscalers are developing in-house AI chips to reduce the reliance on NVIDIA and maintain tailored hardware for specific needs. Google's Tensor Processing Units (TPUs) power much of its AI infrastructure, with the latest Trillium announced in May. Other companies have introduced AI accelerators: Amazon's Trainium and Inferentia, Microsoft's Maia, and Meta's MTIA.
Despite the burgeoning competition, NVIDIA's current dominance is unparalleled. However, as AI technology advances and demands evolve, the landscape suggests a potential market diversification in the coming years.
Chinese Overcapacity is Crushing the Global Steel Industry [Economist]
China, producing approximately 1 billion tonnes of steel annually—as much as the rest of the world combined—has experienced a significant surge in steel exports due to its struggling economy and a downturn in the property sector. In 2023, Chinese steel exports reached 90 million tonnes, a 35% increase from the previous year, surpassing the annual steel production of countries like the United States or Japan and enough to build a thousand Golden Gate bridges. This has distressed global competitors: Nippon Steel in Japan saw an 11% drop in net profit, while Europe's ArcelorMittal reported a 73% decline. In response, countries including Canada, the United States, India (imposing tariffs up to 30%), Brazil, Mexico, Thailand, Turkey, and Vietnam have initiated protective measures against Chinese steel imports. To circumvent trade barriers, Chinese steelmakers are expanding production overseas—China Baowu Steel doubled its investment in a Saudi Arabian plant, and Tsingshan Group began production in Zimbabwe—and are shifting focus to domestic industries like electric vehicle manufacturing, which are also seeking international markets. As James Campbell of CRU Group observes, "Steel will always find a home, whether the world's politicians like it or not," highlighting the ongoing impact of China's steel exports on global trade dynamics and prompting further economic and political responses worldwide.
What’s New in SOLIDWORKS 2025 [Dassault]
Dassault Systèmes launched SOLIDWORKS 2025, the latest release of its portfolio of 3D design and product development applications. SOLIDWORKS 2025 features enhanced collaboration and data management, streamlined workflows for parts, assemblies, drawings, 3D dimensioning and tolerancing, electrical and pipe routing, ECAD/MCAD collaboration, and rendering. The updates also include additions to the SOLIDWORKS PDM, SOLIDWORKS Simulation, SOLIDWORKS Electric Schematic, SOLIDWORKS Electrical Schematic Designer, and DraftSight applications that enable faster design.
What Scared Ford’s CEO in China [WSJ]
Jim Farley, the CEO of Ford Motor Company, recently returned from a visit to China and has since expressed his concerns. He stated that Chinese carmakers are "moving at light speed," utilizing artificial intelligence and other technologies in ways not seen in the U.S. For example, the Xiaomi EV features an infotainment system that connects to home devices, such as turning on lights or the air conditioner when the car approaches. The $77,000 electric minivan from Li Auto boasts plush seats with heated arm and leg rests and large multimedia screens controlled by hand gestures. During a test drive of an electric SUV from Ford's joint venture partner, Changan Automobile, Farley and CFO John Lawler were impressed by its smooth ride and advanced features. Lawler remarked, "Jim, this is nothing like before. These guys are ahead of us." These experiences led Farley to shift Ford's focus in China to commercial vehicles and to reconsider the company's EV strategy globally. In response, Farley is exploring partnerships with the same low-cost parts suppliers used by Chinese automakers and is pivoting toward smaller, more affordable EVs. This strategic shift resulted in the cancellation of a planned electric SUV comparable to the Ford Explorer. Ford is developing a low-cost mechanical platform for future EVs, including a midsize pickup truck expected in 2027. However, this transition will be tough to navigate as the company is projected to lose about $5 billion on EVs this year, which could be up to half of its projected operating profit.
Synthetic Reality for ML Training [Digital Engineering]
Synthetic data to augment real-world data has become a standard practice in training autonomous driving systems. According to Foretellix, "The industry’s long-standing reliance on-road test drives ... has grown over the past decade into an estimated average annual testing cost of $800 million to $2.7 billion per program (largely depending on the mix of physical world testing and simulations)". AV manufacturers often employ a combination of real-world, simulation, and synthetic data to train their navigation algorithms. Carmakers use drive-simulation software like NVIDIA Drive to generate virtual drive data that resembles real-world physics. Simulation offers significant advantages by allowing developers to focus on rare or challenging scenarios without requiring extensive real-world driving. Ansys, for instance, provides tools like Ansys AVxcelerate to generate synthetic data for AV training. By simulating variations in tunnel shapes, vehicle speeds, and environmental conditions, developers can create comprehensive datasets for training. While synthetic data plays a crucial role, there is consensus that it should not replace real-world data entirely, especially in safety-critical applications. For instance, using data from dry, dusty, and gravel roads to predict vehicle behavior on wet roads is risky. The strategic use of synthetic data, supported by advanced simulation tools and careful validation, will be vital to advancing autonomous vehicle technology and help automotive companies manage costs and development time.
Virginia’s Bold Plan to Turn Old Coal Mines into Clean Data Centers [Canary Media]
Two former Virginia state energy office officials turned private-sector consultants are spearheading Energy DELTA Lab to revitalize Southwest Virginia by repurposing 65,000 acres of former coal mine lands into test sites for solar-powered data centers. The first phase of the project, known as Data Center Ridge, involves persuading tech companies to construct data centers that can tap into underground mine water to cool their servers. The mine water, estimated to be between 6 billion to 10 billion gallons at a consistent temperature of 55 degrees, offers a sustainable cooling solution. The data center industry has taken notice. Josh Levi, president of the Data Center Coalition based in Loudoun County, acknowledged the credibility of the efforts. "What they’re doing is credible," he said, mentioning that data centers have begun expanding into "tertiary markets" like Southwest Virginia. Furthermore, the economic impact could be significant. The projects could generate over 1,600 jobs, add 1 gigawatt of new power, and induce $8.25 billion in private investments. A 36-megawatt data center could create about 50 jobs with annual salaries averaging $134,300. They aim to expand Data Center Ridge to 1,000 megawatts in the long term. However, challenges remain. Access to reliable electricity and high-speed broadband are critical for data center operations. The project would require significant infrastructure upgrades, including substations and transmission lines. By pushing forward with this initiative, they hope to prove the doubters wrong and usher in a new era of prosperity for Southwest Virginia.
Research:
Industrial Engineering #3 - With Ashwini Balasubramanian, President of Automotive Women's Alliance
A Robotic Foundation Model for Production [Nature]
Japanese roboticists are advancing artificial intelligence to endow industrial robots with human-like dexterity, enabling them to manipulate irregular, deformable, or transparent objects—tasks that robots traditionally struggle with but humans perform intuitively. The researchers employ deep neural networks—the same technology behind large language models—but instead of text, their models process simulated real-world data like sounds and images. Training the AI system on randomly generated 3D shapes, both familiar and unusual, enhanced the robot's ability to perform picking tasks. The system also estimates the force needed to grasp various products, including soft ones. It generates 'stiffness maps' or 'force maps' to guide the robot's grasp—determining, for example, that it's better to pick up a cleaning brush by its handle rather than its bristles. The initiative seeks to help Japan regain its leadership in industrial robotics. At the same time, Japanese robots held an 80% global market share 30 years ago. That figure has declined to 50% due to overseas competition and the demand for more intelligent robots.
Podcasts:
Humanoid Robotics w/ Melanee Wise [Manufacturing Happy Hour]
Manufacturing Deals🏭💵
TeamBridge - A company building a HR platform for hourly workers
$28 million [Series B] - Led by Mayfield and joined by General Catalyst and Abstract Ventures
Inbolt - A French provider of real-time robot guidance solutions
€15 million [Series A] - Led by Exor Ventures and joined by Bpifrance, Yann Fleureau, MIG Capital, SOSV, and BNP Paribas Développement
Planned Downtime 🏭🧑🔧
Interview With The Head of Apple Silicon