D-Matrix Unveils Corsair Chip Claiming 10x Faster AI Inference, Shipping on 6nm This Month

3 articles · Updated · StartupHub.ai · Aug 22

Corsair is entering production now, with D-Matrix saying shipments of the 6-nanometer AI inference chip will begin this month.
SRAM built directly onto the chip is the key pitch: D-Matrix says that avoids scarce external HBM, lifting inference speed to 10 times that of GPUs and cutting power use by up to 5 times.
The chip packs four inference compute dies, 315 PFLOPS, 128 GB of SRAM and 40 PB/s of memory bandwidth, with systems scaling to 256 chips.
Commitments from hyperscalers, Neocloud providers and AI labs suggest early demand as data centers seek alternatives to GPU-heavy architectures constrained by AI memory supply.

Sources

Could D-Matrix's SRAM-centric chip finally shatter the AI memory wall, or is it just a temporary fix until HBM supply recovers?

Will the shift to air-cooled, memory-bound AI co-processors make traditional GPU-only data centers obsolete by the end of 2026?

Why are hyperscalers pairing this new 6nm inference chip with NVIDIA's Blackwell GPUs instead of replacing them entirely?