D-Matrix to Launch Raptor Chip With 3D DRAM
The 2027 accelerator stacks a new memory tier beneath its logic die, sitting between SRAM and HBM in capacity.
Updated on Nov. 19, 2026 in Semiconductors

D-Matrix plans to bring its second-generation Raptor chip to market in 2027, with 3D DRAM stacked beneath the logic die. The company's Pavehawk test chip is validating the stacked-memory technology for Raptor.
Why it matters
The memory design targets bandwidth constraints as AI workloads shift toward inference. D-Matrix has also adopted Nvidia's NVLink Fusion rack-scale infrastructure to connect its accelerators to Nvidia AI factories.
D-Matrix claims its planned memory tier delivers 10 times HBM's bandwidth and 10 times its energy efficiency, with the Corsair accelerator lifting GPU utilization from about 20% to as much as 80% in mixed fleets. These are company projections, and independent validation for the 2027 Raptor chip has not been reported.
The players
D-Matrix
D-Matrix is a semiconductor company developing AI inference accelerators and stacked-memory technology.
Nvidia
Nvidia is a leading designer of GPUs and AI infrastructure whose NVLink Fusion rack-scale platform D-Matrix has adopted.
Groq
Groq is an AI chip company that has explored 3D-stacked DRAM for a future accelerator.
The details
Satyam Srivastava, a D-Matrix fellow, said the Pavehawk test chip is validating stacked-memory technology for Raptor. He described an approach in which software coordinates GPUs, CPUs and specialized accelerators to serve different workloads. Groq has also weighed 3D-stacked DRAM for a future AI chip.
Timeline
2027: Planned market launch of the Raptor chip with 3D DRAM
The Tech Race
D-Matrix's adoption of NVLink Fusion places its accelerators inside Nvidia's rack-scale ecosystem rather than against it. The planned 3D DRAM tier, paired with Groq's separate exploration of stacked DRAM, shows smaller AI chip makers converging on memory-bandwidth fixes as inference demand grows.
For users of AI services, the chip's targeted efficiency gains could eventually lower the cost of running inference at scale. The timeline is 2027, so no immediate change to hardware choices is implied by the announcement.
The takeaway
Memory bandwidth, not just raw compute, is becoming the limiting factor for AI inference, and that is shaping chip roadmaps. Buyers evaluating accelerators should watch independent benchmarks once Raptor reaches market.
Further reading
Explore more on Semiconductors to see how chipmakers are tackling memory bandwidth for AI.
Source note: This article includes information reported by TokenPost.



