d-Matrix is a semiconductor startup focused on building computing systems for artificial intelligence inference, particularly for large language models. The company develops a chip architecture that integrates memory and computation to reduce the energy and latency associated with running trained neural networks. Founded in 2019, d-Matrix has positioned itself in the competitive field of AI accelerators, targeting data center deployments where inference workloads dominate.
The core technology behind d-Matrix's products is digital in-memory compute (DIMC), which differs from traditional von Neumann architectures that separate processing units from memory. By performing multiply-accumulate operations directly within memory arrays, the design aims to minimize data movement, a major bottleneck in AI inference. The company's first commercial chip, the Nighthawk, was announced in 2024, with a successor, the Corsair, slated for 2025.
Founding and Early History
d-Matrix was founded in 2019 by Sid Sheth, who previously worked at Intel and AMD, and Sudeep Bhoja, a former engineer at Broadcom. The two met while working on high-speed interconnects and recognized the growing demand for efficient inference hardware. The company initially operated in stealth mode, raising seed funding from venture capital firms including Wing Venture Capital and M12, Microsoft's venture fund.
In 2021, d-Matrix emerged from stealth with $44 million in Series A funding, led by Playground Global. The funding round also included participation from SK Hynix, a major memory manufacturer, signaling interest from the memory industry in compute-in-memory technologies. By 2023, the company had raised an additional $110 million in Series B funding, with investors such as Temasek and Innovation Endeavors, bringing total funding to over $154 million.
The company's early focus was on developing a chip that could handle the specific demands of transformer-based models, which had become the dominant architecture in Natural language processing and Generative AI. The founders believed that conventional GPUs, while powerful for training, were inefficient for inference due to their high memory bandwidth requirements.
Technology: Digital In-Memory Compute
The fundamental innovation at d-Matrix is its digital in-memory compute architecture. Unlike analog in-memory computing approaches, which use physical properties like resistance or capacitance to perform calculations, d-Matrix uses digital logic embedded within memory cells. This approach aims to combine the benefits of reduced data movement with the precision and reliability of digital computation.
In a traditional Neural network inference, weights are stored in memory and must be fetched to a separate compute unit for each operation. This constant data transfer consumes significant energy and time. d-Matrix's DIMC architecture places computational elements directly adjacent to or within the memory array, allowing multiply-accumulate operations to occur where the data resides. This reduces the need for extensive data movement, potentially lowering latency and power consumption.
The company's chips are designed to support Large language model inference, including models with billions of parameters. The architecture is optimized for the matrix multiplication operations that dominate transformer models, which are the basis for systems like OpenAI's GPT series and Anthropic's Claude. d-Matrix claims its approach can achieve higher throughput per watt compared to traditional accelerators, though independent benchmarks have been limited as of 2025.
Product Line: Nighthawk and Corsair
d-Matrix announced its first product, the Nighthawk, in March 2024. The Nighthawk is a PCIe card designed for inference acceleration, featuring 32 GB of on-chip memory and supporting models up to 70 billion parameters in a single card. The card is built using a 5-nanometer process from TSMC, a leading semiconductor foundry. The Nighthawk is intended for use in servers, where multiple cards can be combined to handle larger models.
In October 2024, d-Matrix unveiled its second-generation chip, the Corsair, which is expected to ship in 2025. The Corsair is designed to deliver higher performance than the Nighthawk, with improved memory capacity and compute density. The company has stated that the Corsair will support models exceeding 100 billion parameters on a single card, targeting enterprise deployments of Generative AI applications.
Both chips are accompanied by a software stack called the d-Matrix Compiler, which translates models from frameworks like PyTorch into optimized code for the hardware. The software includes support for common model architectures, including transformers and Residual Network (ResNet) variants, and aims to simplify the deployment process for developers.
Market Position and Competition
d-Matrix operates in a crowded market for AI inference accelerators. Major competitors include Groq, which offers a language processing unit (LPU) optimized for sequential workloads, and SambaNova, which uses a reconfigurable dataflow architecture. Established players like NVIDIA dominate the broader AI accelerator market, while AMD and Intel have also introduced inference-focused products.
The company differentiates itself through its focus on in-memory computing, which it argues is more efficient than approaches that rely on high-bandwidth memory (HBM) like those used by Nvidia's GPUs. d-Matrix's technology is particularly aimed at reducing the cost per token for serving large models, a key metric for cloud providers and enterprises.
As of 2025, d-Matrix has not disclosed major public customers, though it has announced partnerships with Amazon Web Services for potential integration into its cloud offerings. The company is also working with Oracle Cloud Infrastructure and Microsoft Azure to explore deployment options, though these relationships are in early stages.
Funding and Investors
d-Matrix has raised a total of $154 million across multiple rounds. The seed round, undisclosed in size, was followed by the $44 million Series A in 2021 and the $110 million Series B in 2023. Key investors include Playground Global, Wing Venture Capital, M12, Temasek, and Innovation Endeavors. The participation of SK Hynix, a memory chip manufacturer, is notable given the company's compute-in-memory focus.
The funding has been used to develop the hardware and software stack, as well as to hire engineering talent. As of 2025, the company employs approximately 100 people, with offices in Santa Clara, California, and Bangalore, India. The Bangalore office focuses on software development and compiler optimization.
Applications and Use Cases
d-Matrix's primary target market is data center inference for Large language model applications. These include chatbots, code generation, and document summarization, which are increasingly deployed by enterprises. The company's hardware is designed to handle the high throughput requirements of serving many users simultaneously, a scenario common in cloud environments.
Beyond language models, d-Matrix's chips can accelerate other Deep learning workloads, including computer vision and recommendation systems. The architecture is general enough to support various neural network types, though the company's marketing emphasizes transformer-based models. The low latency of in-memory computing could benefit real-time applications like autonomous driving, though d-Matrix has not announced partnerships in that sector.
The company also targets edge inference scenarios where power efficiency is critical. However, its current products are designed for server environments, and edge deployment would require further miniaturization and power reduction.
Challenges and Future Outlook
The AI accelerator market is highly competitive, with rapid technological advancements and significant capital requirements. d-Matrix faces challenges in scaling production and convincing customers to adopt a new architecture. The company must also compete with the ecosystem and software maturity of Nvidia's CUDA platform, which has become the industry standard.
Technical hurdles remain in proving the reliability and performance of in-memory computing at scale. While digital approaches mitigate some issues of analog compute, such as noise and precision loss, they still require careful integration with existing server infrastructure. The success of the Corsair chip in 2025 will be a critical test of the company's technology.
Looking ahead, d-Matrix aims to expand its product line and potentially enter the training market, though training is more compute-intensive and may not benefit as much from in-memory techniques. The company is also exploring partnerships with TSMC for advanced manufacturing and with Arm Holdings for potential integration into ARM-based servers. As of 2025, d-Matrix remains a private company with no announced plans for an initial public offering.
Leadership and Team
Sid Sheth, the CEO, has over two decades of experience in the semiconductor industry, having held engineering and management roles at Intel and AMD. Sudeep Bhoja, the CTO, previously worked at Broadcom on high-speed interconnect technologies. The leadership team includes veterans from companies like Qualcomm and Samsung Electronics, bringing expertise in chip design and manufacturing.
The company's technical advisory board includes academics and industry experts in machine learning and computer architecture, though specific names have not been publicly disclosed. The team has published several papers on in-memory computing techniques, contributing to the broader research community.