D-Matrix is developing Raptor, a second-generation chip for AI inference that stacks 3D DRAM beneath its compute logic. Company fellow Satyam Srivastava said at SC26 that a test chip, Pavehawk, is validating the technology ahead of Raptor, which he said is due next year.
Srivastava claims the stacked memory will offer 10 times the bandwidth and 10 times the energy efficiency of high-bandwidth memory (HBM), with capacity between HBM and SRAM. Those figures are company claims; the interview does not provide independent test results. D-Matrix also says pairing its Corsair accelerator with Nvidia Hopper GPUs can raise utilization from about 20% to as much as 80%.
Why it matters
For companies serving AI models, the claimed bandwidth gains could help keep inference hardware busy, but independent results and Raptor’s actual performance are not yet documented.