Updated
Updated · Wccftech · Aug 24
d-Matrix Unveils 100 TB/s Raptor 3D DRAM as 0.37 pJ/bit Challenges HBM4
Updated
Updated · Wccftech · Aug 24

d-Matrix Unveils 100 TB/s Raptor 3D DRAM as 0.37 pJ/bit Challenges HBM4

3 articles · Updated · Wccftech · Aug 24

Summary

  • Raptor pairs a TSMC 4nm logic die with custom 3D DRAM to deliver 32 GB and 100+ TB/s per card, which d-Matrix says targets the AI inference memory bottleneck.
  • At 0.37 pJ/bit, the design uses about one-tenth of HBM energy and avoids HBM’s PHY and beachfront limits, while pitch-matching 256 tensor engines to DRAM banks to cut data movement.
  • Against a 192 GB HBM4 setup at 18 TB/s and 2-3 pJ/bit, d-Matrix claims 5.6x higher bandwidth, 5-8x lower energy and 7x denser I/O; versus Nvidia’s Rubin R200 platform, it cites 23.4x higher bandwidth density.
  • The approach still faces thermal, power-delivery and yield hurdles: logic-on-top requires liquid cooling, I/O alone can draw about 300 W at 100 TB/s, and the DRAM refreshes every 4 ms with only 1.37% bandwidth loss.
  • Presented at Hot Chips 2026, Raptor extends d-Matrix’s push for an alternative to HBM as AI model sizes and KV-cache demands keep rising.

Insights

Could a startup's radical 3D memory stack finally break the AI bottleneck and dethrone traditional HBM infrastructure?
As AI memory shortages threaten server deployments through 2030, is bringing compute directly to memory our only viable escape route?
With extreme thermal limits, will this new stacked-DRAM technology actually survive the harsh realities of modern data centers?