D-Matrix Plans Stacked-Memory AI Chip to Launch in 2027
The Raptor chip's 3D DRAM tier is still in test-chip validation, and its claimed 10-times-HBM bandwidth has not been independently verified.
Updated on Nov. 19, 2026 in Semiconductors

D-Matrix plans to bring its second-generation Raptor chip to market in 2027, with a stacked 3D DRAM memory tier beneath the logic die. The memory tier sits between SRAM and high-bandwidth memory (HBM) in capacity. The development is research-stage: D-Matrix's Pavehawk test chip is validating the stacked-memory technology, and no shipped product exists yet.
Why it matters
The memory design targets memory bandwidth needs as AI workloads shift toward inference, a bottleneck that has pushed chipmakers to add memory tiers between on-chip SRAM and HBM. Groq has also weighed 3D-stacked DRAM for a future AI chip, suggesting the approach is drawing interest across the accelerator field.
The claimed figures, 10 times HBM's bandwidth and 10 times HBM's energy efficiency, come from D-Matrix's own projections for the planned memory tier. The 20% to 80% utilization gain for the Corsair accelerator is a company claim measured against an existing Nvidia Hopper GPU fleet, and no independent benchmark is cited in the source.
The players
D-Matrix
An AI chip startup whose Corsair accelerator and second-generation Raptor chip target inference workloads, with its Pavehawk test chip validating the stacked-memory design.
Nvidia
The dominant maker of AI GPUs whose NVLink Fusion rack-scale infrastructure D-Matrix has adopted to connect its accelerators to Nvidia AI factories.
Groq
An AI inference chip company that has weighed 3D-stacked DRAM for a future AI chip.
The details
D-Matrix proposes stacking DRAM, a dense memory type, vertically beneath the logic die, the layer that performs computation, so the memory sits physically close to the processing circuits. This shortens the distance data must travel, which is how stacked designs raise bandwidth and cut energy per bit moved. Srivastava described a broader approach in which software coordinates GPUs, CPUs and specialized accelerators to serve different workloads. D-Matrix connects its accelerators to Nvidia AI factories using Nvidia's NVLink Fusion rack-scale infrastructure, a high-speed interconnect for linking many chips into one system.
Timeline
2027: D-Matrix's planned market launch of the Raptor chip with 3D DRAM.
The Tech Race
D-Matrix's planned tier sits between on-chip SRAM and HBM in capacity, positioning it against HBM as the standard stacked memory used with leading AI GPUs. Groq's reported interest in 3D-stacked DRAM suggests a parallel push by a second inference-focused chipmaker toward the same memory architecture.
The Raptor chip is planned for 2027, so no hands-on availability exists yet. Data-center operators running mixed fleets of Nvidia GPUs could eventually see higher utilization if the Corsair approach is adopted, though pricing and availability have not been announced.
The takeaway
The trajectory points toward memory placement becoming a central lever in AI chip design as inference workloads grow. Watch whether D-Matrix's Raptor chip reaches market in 2027 and whether independent benchmarks confirm the claimed bandwidth and efficiency gains.
Further reading
Explore more Semiconductors coverage to see how chipmakers are tackling memory bottlenecks in AI hardware.
Source note: This article includes information reported by TokenPost.


