News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
Corsair

Corsair

AI inference accelerator with in-memory compute

Corsair is an AI inference chip by d-Matrix, launched in 2024. It uses in-memory computing and block floating point formats to run large language models with low latency in data centres.

Share on WhatsApp

  • Uses Digital In-Memory Compute cores
  • Supports MXINT4, MXINT8, and MXINT16 formats
  • 256 GB LPDDR5 memory per card
  • Runs Llama3-70B models on a single card
  • No support for FP32 or BF16 data types

What it is

Corsair is a specialised processor for artificial intelligence inference. Designed by d-Matrix, it targets data centre environments where speed and efficiency matter. The chip avoids standard floating-point maths, focusing instead on low-precision formats. This approach allows it to handle large language models without the high energy costs of moving data. It integrates into existing server racks via PCIe connections.

How it came about

d-Matrix introduced Corsair as its first product in 2024. The company positioned it as a memory-centric platform for scale. It was presented at Supercomputing 2024. The design aims to solve the bottleneck of data movement in traditional AI hardware. By placing computation inside memory cells, the team sought to reduce latency for real-time applications.

Inside the chip

The architecture relies on Digital In-Memory Compute cores. Each core embeds a multiplier within the memory bit cell. This eliminates the need to shuttle data between separate memory and processing units. Corsair supports Microscaling formats like MXINT4, MXINT8, and MXINT16. These block floating-point types share exponents across groups of numbers. The chip does not support IEEE 754 standards such as FP32 or BF16. A single package contains four identical rectangular dies.

How it was made

Each Corsair package includes 1 GB of SRAM and connects to 128 GB of LPDDR5 memory. A full card holds two packages, offering 256 GB of LPDDR5 bandwidth. The design uses a chiplet-based approach to scale SRAM capacity. d-Matrix claims this structure supports models with up to 100 billion parameters.

Where it went

Corsair cards fit into standard PCIe slots in data centres. They can run multiple instances of large models like Llama3-70B on a single card using MXINT8 precision. Smaller models like Llama3-8B fit entirely in the fast SRAM across eight cards. The system supports rack-scale deployments through JetStream interconnects. It works alongside GPUs in heterogeneous pipelines to offload inference tasks.

Versions and family

The primary configuration involves a card with two packages. Each package contains four dies. The system offers two memory modes: one using only fast SRAM for low latency, and another accessing the larger LPDDR5 pool for capacity. There are no mentioned predecessors or successor chips in the current documentation.

Why it matters

Corsair introduces block floating-point formats to the AI hardware market. This aligns with trends seen in Nvidia Blackwell and Microsoft MAIA chips. By prioritising memory-centric compute, it challenges traditional GPU-dominated inference workflows. The design demonstrates how specialised formats can maintain accuracy while reducing data movement overhead.

Problems and reception

Technical analysts note the resemblance of the four-die layout to AMD Naples CPUs, though the function differs entirely. The exclusion of standard floating-point types limits its use to specific low-precision workloads. Reviewers highlight the pragmatic use of block floating point for energy efficiency.

Questions people ask about Corsair

What type of maths does Corsair use?

It uses Microscaling formats such as MXINT4, MXINT8, and MXINT16. These are block floating-point types that share exponents. It does not support standard IEEE 754 formats like FP32, FP16, or BF16.

How much memory does a Corsair card have?

A single card contains two packages. Each package has 1 GB of SRAM and 128 GB of LPDDR5. This gives a total of 256 GB of LPDDR5 memory per card, with 2 GB of combined SRAM.

Can it run large language models?

Yes. A single card can host two instances of Llama3-70B using MXINT8 precision. For smaller models like Llama3-8B, eight cards can store the model entirely in SRAM for faster access.

How does it connect to servers?

Corsair integrates into existing infrastructure using PCIe connections. It supports rack-scale deployments through JetStream technology for communication between accelerators.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Corsair yet, so the facts on this page come from the sources listed above: d-matrix.ai/product/, chipsandcheese.com/p/d-matrix-corsair-256gb-of-lpddr-for. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.