d-matrix.ai

d-Matrix is a semiconductor company focused on AI inference, and its core claim is architectural: it moves compute into memory rather than keeping the two far apart. The company describes this as "memory-centric compute" built for ultra-low latency, greater efficiency, and scalable AI infrastructure. Its main product, Corsair, is a PCIe inference card that can be deployed on its own for fixed-memory workloads or paired with GPUs in a heterogeneous setup.

The problem d-Matrix is targeting

d-Matrix's site frames the market around an "inference crunch": inference demand is growing faster than today's hardware can serve it, because the technology that trained AI models was not designed to deliver inference at current scale. The company's argument is that traditional architectures keep compute far from memory, and even approaches that move compute "closer" still leave distance on the table. d-Matrix says it brings compute directly into memory itself, which it claims lowers latency, improves efficiency, and gives a better path to scaling inference.

That is the company's positioning, not an independently verified benchmark. Treat the latency and efficiency claims as vendor statements until you can test them against your own workload.

Corsair: the product

Corsair is described as purpose-built for AI inference, with these stated characteristics:

The practical takeaway from the form factor is that Corsair is designed to slot into standard PCIe-based systems rather than requiring a proprietary enclosure. The site does not publish pricing, availability dates, or detailed specifications on the pages reviewed, so those details would need to come from a sales conversation.

Standalone or alongside GPUs

d-Matrix positions Corsair in two deployment modes:

• Standalone: a complete inference solution for fixed-memory workloads.
• Heterogeneous: paired with GPUs, with each workload placed on the architecture best suited to run it. d-Matrix specifically mentions offloading fixed-memory tasks to Corsair.

The heterogeneous pitch is the more distinctive one. Rather than asking buyers to replace GPUs, d-Matrix argues that some inference work fits its memory-first design better and can be offloaded, leaving GPUs for what they do best. Whether that split makes sense depends on how much of your inference load is fixed-memory and latency-sensitive.

Software and ecosystem

Two software pieces appear on the site:

• Aviator SDK: described as offering maximum flexibility in inference optimization, with a PyTorch-like interface.
• JetStream: described as codesigned with Aviator.

The PyTorch-like framing matters if your team already works in that ecosystem, since it suggests a lower porting burden than a fully custom toolchain. The site does not detail supported model formats, frameworks beyond the PyTorch comparison, or quantization support, so validate those against your stack before committing.

d-Matrix also announced a partnership with NVIDIA on NVLink Fusion, stating that NVLink Fusion will connect next-generation d-Matrix XPUs to the NVIDIA AI platform, including NVIDIA Vera, ConnectX-9, BlueField-4 DPU, and Spectrum-X Ethernet. For buyers, this signals an intent to integrate with NVIDIA-based infrastructure rather than compete with it outright — consistent with the heterogeneous positioning.

Who should look closer

d-Matrix is worth investigating if you run latency-sensitive inference at scale, have a meaningful share of fixed-memory workloads, and are open to a mixed accelerator fleet rather than an all-GPU deployment. The PCIe form factor and PyTorch-like SDK lower the evaluation cost.

It is a weaker fit if you need published pricing and specs to build a budget now, if your workloads are dominated by tasks that already run well on GPUs, or if you require a fully documented software stack before trialing hardware. In those cases, the next step is a direct conversation with d-Matrix sales rather than a self-serve evaluation.

d-Matrix builds AI inference hardware around a memory-centric architecture: instead of keeping compute and memory apart (or merely moving them closer), it places compute directly into memory itself. According to d-Matrix, collapsing that distance is what lowers latency and improves efficiency for inference workloads. The approach is embodied in its Corsair product line, which the company positions as either a standalone inference solution for fixed-memory workloads or as a complement to GPUs in heterogeneous systems.

The core idea: from "far" to "near" to "in"

d-Matrix describes the evolution of compute-memory relationships in three stages:

• Far — traditional architectures keep compute physically distant from memory, so data must travel back and forth.
• Near — some designs move compute closer to memory, shortening the trip but not eliminating it.
• In — d-Matrix moves compute into memory itself, which the company says fundamentally changes how data moves and how quickly inference results are delivered.

The stated rationale is that inference is exploding in scale, and hardware designed primarily for training ("the brains of AI") was not built to deliver intelligence at the volume and speed now demanded. By working "with physics rather than fighting it," d-Matrix argues it can reduce latency and improve efficiency without compromising scalability.

What Corsair actually integrates

Corsair is the product that implements this architecture. Per d-Matrix's description, a single Corsair card combines:

The card is described as built to scale up and scale out to rack-level systems, so the same architecture can serve both single-card deployments and larger inference clusters.

Standalone or alongside GPUs

d-Matrix explicitly frames Corsair as "built to stand alone" and "built to work with GPUs":

• Standalone: Deploy Corsair as a complete inference solution for fixed-memory workloads.
• Heterogeneous: Pair Corsair with GPUs and route each workload to the architecture best suited to run it. d-Matrix says its memory-first technology is suited to offloading fixed-memory tasks and certain advanced inference techniques.

This matters if you already run GPU infrastructure: the pitch is not wholesale replacement but workload placement, putting memory-bound inference on Corsair while leaving other work on GPUs.

Software and ecosystem

• Aviator SDK — d-Matrix's interface for next-generation AI, codesigned with its JetStream offering. It is described as offering maximum flexibility in inference optimization, with a PyTorch-like interface.
• NVIDIA NVLink Fusion partnership — d-Matrix has announced that NVLink Fusion will connect next-generation d-Matrix XPUs to the NVIDIA AI platform, including NVIDIA Vera, ConnectX-9, BlueField-4 DPU, and Spectrum-X Ethernet. This signals an intent for d-Matrix accelerators to interoperate within NVIDIA-based data center environments.

What this means in practice

If your inference workloads are latency-sensitive and memory-bound — where the cost is dominated by moving data between memory and compute — the memory-centric design is the relevant claim to evaluate. The practical questions to bring to d-Matrix are which fixed-memory workloads Corsair supports today, how the Aviator SDK maps existing PyTorch-style models onto the hardware, and what a heterogeneous deployment with your current GPUs looks like. Pricing, availability, and configuration details are not specified on the site; the published path is to contact sales.

d-Matrix Corsair is a card built specifically for AI inference that puts high-bandwidth memory and compute together on the same fabric. According to d-Matrix, it combines an inference-optimized compute engine, high-speed chip-to-chip connectivity, and an industry-standard PCIe form factor, and it is designed to scale up and out to rack-level systems. It can be deployed as a complete inference solution for fixed-memory workloads or paired with GPUs in a heterogeneous setup.

The design idea behind Corsair

d-Matrix describes its approach as moving compute into memory rather than only bringing the two closer together. The company frames the evolution as "from far, to near, to in": traditional architectures keep compute far from memory, other designs move it nearer, and d-Matrix places compute directly into memory itself. The stated result is lower latency, greater efficiency, and a better path to scaling inference.

This matters because, in d-Matrix's framing, inference demand is growing faster than today's hardware can support, and chips originally designed to train AI models were not built to deliver inference at current scale.

Key features at a glance

Standalone or alongside GPUs

Corsair is positioned as usable in two ways. As a complete inference solution, it targets fixed-memory workloads. Paired with GPUs, it lets each workload run on the architecture best suited to it — d-Matrix describes offloading fixed-memory tasks and newer inference techniques to Corsair while GPUs handle what they do best. The company calls this heterogeneous computing.

Software: JetStream and Aviator

d-Matrix pairs Corsair with JetStream, which it describes as codesigned with Aviator, its interface for next-generation AI. The Aviator SDK is presented as offering flexibility in inference optimization, including a PyTorch-like interface.

Ecosystem: NVIDIA NVLink Fusion

d-Matrix has announced a partnership with NVIDIA on NVLink Fusion, which would connect next-generation d-Matrix XPUs to the NVIDIA AI platform. The announcement lists NVIDIA Vera, ConnectX-9, BlueField-4 DPU, and Spectrum-X Ethernet as part of that integration.

What the page does not cover

The source material does not include pricing, availability dates, detailed performance benchmarks, or full technical specifications for Corsair. Readers evaluating a purchase should request those details directly through d-Matrix's sales channel rather than inferring them from the product overview.

The Aviator SDK is d-Matrix's software interface for optimizing AI inference on its hardware. According to d-Matrix, it is codesigned with JetStream and offers "maximum flexibility in inference optimization," with a PyTorch-like programming model. In practical terms, it is the layer that lets you tune how models run on Corsair and related d-Matrix XPUs rather than accepting a fixed runtime. The details below reflect what d-Matrix has published; specific APIs, supported operators, and performance numbers are not covered in the available material, so confirm those with d-Matrix before planning a deployment.

Where Aviator Sits in the d-Matrix Stack

d-Matrix's pitch is memory-centric compute: compute is moved into memory itself, collapsing the distance between memory and compute to cut latency and improve efficiency for inference. Corsair is the product built on that idea, combining high-bandwidth memory and compute on the same fabric with an inference-optimized compute engine, high-speed chip-to-chip connectivity, and a standard PCIe form factor.

Aviator is the software counterpart. The company describes it as "our interface for next-generation AI," codesigned with JetStream. That codesign matters: because the SDK and the runtime are developed together, optimization work in Aviator is intended to map onto how JetStream actually executes workloads, rather than being a generic layer bolted on afterward.

What "PyTorch-Like" Means for Optimization

The published description says Aviator offers a PyTorch-like experience. For teams already building in PyTorch, the implication is a shorter path from an existing model to an optimized inference deployment: familiar model definitions and programming patterns, with optimization controls exposed through the SDK.

What the material does not specify is the exact degree of compatibility — whether that means direct import of PyTorch models, a similar tensor/operator API, or a graph-level interface. Treat "PyTorch-like" as a signal of developer ergonomics, not a guarantee of drop-in compatibility, and verify the supported model formats and operators directly.

How It Connects to Inference Performance

The optimization story rests on matching workloads to the right architecture. d-Matrix positions Corsair two ways:

• Standalone: Corsair as a complete inference solution for fixed-memory workloads.
• Heterogeneous, alongside GPUs: pair Corsair with GPUs and put each workload on the architecture best suited to it, offloading fixed-memory tasks and newer inference techniques.

Aviator is the interface through which that tuning happens. If your goal is to maximize inference performance, the practical workflow implied by the materials is: identify which workloads are fixed-memory and latency-sensitive, target those at Corsair, and use Aviator to optimize them — while leaving other workloads on GPUs.

Who Should Look at Aviator

Aviator is most relevant if you are evaluating Corsair for latency-sensitive inference and want control over how models are optimized, rather than a closed appliance. It is less useful as a standalone software product — it exists to drive d-Matrix hardware.

Two caveats before you commit engineering time: the available material gives no pricing, licensing, or access details for the SDK, and no concrete benchmark figures. Both should be settled in a direct conversation with d-Matrix, ideally with a proof-of-concept on your own model rather than a generic demo.

Yes to both. According to d-Matrix, Corsair is "built to stand alone" as a complete inference solution for fixed-memory workloads, and it is also "built to work with GPUs." In a GPU-paired setup, you put each workload on the architecture best suited to run it, using Corsair to offload fixed-memory tasks and certain inference techniques while GPUs handle the rest.

What Corsair Is

Corsair is d-Matrix's inference card. Per the company's site, it is purpose-built for AI inference and combines high-bandwidth memory and compute on the same fabric. Its stated components:

• An inference-optimized compute engine
• High-speed chip-to-chip connectivity
• An industry-standard PCIe form factor

It is designed to scale up and scale out to rack-level systems. The architecture is described as memory-centric: d-Matrix says it moves compute into memory itself, rather than keeping compute far from memory or merely closer to it, to reduce latency and improve efficiency.

Option 1: Corsair Standalone

Use Corsair by itself when your workload fits a fixed-memory profile — the model and its working data fit within the card's memory, and the task is inference rather than training.

What you get: a complete inference solution on one architecture, without GPUs in the path. d-Matrix positions this for fixed-memory workloads where its memory-centric design delivers lower latency and greater efficiency.

What to check before committing: whether your model's memory footprint fits the card, whether your serving stack can target it, and whether your throughput and latency targets are met without GPU acceleration. The site does not publish per-card memory capacity, model support lists, or benchmark numbers, so those specifics need to come from d-Matrix directly (the site routes interested buyers to "Contact Sales").

Option 2: Corsair Paired With GPUs

Use the heterogeneous setup when your work mixes workload types — for example, some inference tasks that are memory-bound and fixed in size, alongside tasks that benefit from GPU compute.

How the split works: d-Matrix says you pair Corsair with GPUs and "put each workload on the architecture best suited to run it." The company describes its memory-first technology as enabling offloading of fixed-memory tasks and "cutting-edge inference techniques" to Corsair.

The interface layer: the Aviator SDK is d-Matrix's interface for next-generation AI, described as offering maximum flexibility in inference optimization with a PyTorch-like interface. JetStream is codesigned with Aviator. If your team already works in PyTorch-style tooling, that is the relevant integration surface to evaluate.

Ecosystem note: d-Matrix has announced a partnership with NVIDIA on NVLink Fusion, connecting next-generation d-Matrix XPUs to the NVIDIA AI platform, including NVIDIA Vera, ConnectX-9, BlueField-4 DPU, and Spectrum-X Ethernet. This signals an intended path for Corsair to sit alongside NVIDIA infrastructure rather than replace it — but treat the announcement as a roadmap signal, not a shipping configuration you can assume today.

Choosing Between Them

A practical way to decide: start by classifying your inference workloads. If they are uniformly fixed-memory and latency-sensitive, standalone Corsair is the simpler architecture to validate. If you have a mix — or you already run GPU infrastructure you don't want to idle — the heterogeneous path lets you assign each workload type to the hardware d-Matrix says suits it, with Aviator as the optimization layer.

Because d-Matrix does not publish pricing, capacity figures, or benchmark results on the pages reviewed, confirm fit, availability, and commercial terms through a sales conversation before planning a deployment.