Rethinking Silicon Math
for the Next Era of AI.
Cassia.ai's patent-pending approach to digital arithmetic circuits speeds up the essential functions of AI chips. Tensor functions like Matrix Multiplication (MatMul) using Multiply-Accumulate (MAC), and vector activation functions like SoftMax, TanH and GELU are accelerated by over 10x.
Keep the same GPU, NPU, TPU, CPU or xPU architecture you're using now. Keep the same data types, keep the same pretrained models. Cassia.ai is an easy RTL upgrade for your lowest level foundational mathematics IP; everything else about your chip remains the same while getting faster, smaller and saving power.
We Do Math Better.
Fundamental Innovation at the Gate Level.
While conventional accelerators rely on ancient mathematical techniques, Cassia.ai redesigns the foundational logic of computer math for incredible performance gains.
Order-of-Magnitude MAC & Activation Scaling
Multiply-Accumulate (MAC) operations make up over 90% of deep neural network execution cycles, while activation functions can consume 50% or more of delays. Cassia.ai's mathematical acceleration techniques decimate gate logic, allowing chip architects to pack more arithmetic density into identical die boundaries while shrinking the delays of the most time-consuming activation operations.
No Loss of Precision
Aggressive quantization typically destroys transformer accuracy. Cassia.ai's high-precision, deterministic and monotonic outpus permit sparsity and training that are impossible with quantization. Cassia.ai allows you to keep the precision of your choosing while giving you performance and gate count benefits.
Cloud to Edge Applications
Cassia.ai is synthesizable RTL, meaning that the acceleration scales from performance to power according to your needs - from incredibly fast datacenter performance to power-sipping edge applications.
Resolving the AI Energy Crisis
Cassia.ai's extreme efficiency means that you can deploy AI even in power-constrained environments. Did you meet your green requirements? Is your local grid maxed out? Transformers on backorder? Ran out of cooling before you ran out of floor space? Floors not strong enough to support the weight of liquid cooled equipment? Installing on-prem AI in existing server rooms? We can help.
Accelerating Real-World Silicon.
Where high mathematical density and ultra-low power unlock completely new hardware product categories.
AI Datacenter Processors
Improve performance and reduce power in AI Inference and Training - from RAGs to Frontier Training.
Edge LLMs
Shift multi-billion parameter language reasoning directly onto consumer client hardware with zero cloud latency.
On-Premises AI
Run private AI workloads within your own infrastructure, on your own site, within your electrical and cooling constraints.
Industrial and Medical Sensors
Sub-millisecond localized telemetry monitoring and prediction without round-tripping to remote servers.
Semiconductor Veterans.
Decades of combined IC tape-out experience, patent authorship, and front-end ASIC verification.
James Tandon, Ph.D.
10+ chip tape-outs. Developed Cassia.ai's computer arithmetic algorithms. Author of multiple patents across IC design and machine learning.
Mike Reznikov
15+ years experience optimizing RTL and front-end digital design across high-performance FPGA and ASIC architectures.
Nehal Patel
18+ years leading module-level and full-chip verification cycles for mission-critical enterprise semiconductor products.
Marc Greenberg
30+ years industry leadership in semiconductor product marketing, IP licensing, and silicon hardware engineering.
All the Math You Need, in One RTL Package.
Cassia.ai delivers foundational arithmetic building blocks that integrate cleanly across modern silicon stacks and custom AI accelerators.
| Product | Description | Non-NDA Datasheet |
|---|---|---|
| Adder-Subtractor | Adder-subtractor for AI | Download |
| Cosine | Fast and low-power cosine function | Download |
| Divider | Highly Accelerated Divider for AI | Download |
| Exponent (Base 10) | Fast and low power Base 10 Exponentiation, highly accelerated, for standalone use or integration into higher-order functions. This function is fast even in a single cycle, or may be pipelined for even higher speed. | Download |
| Exponent (Base 2) | Fast and low power Base 2 Exponentiation, highly accelerated, for standalone use or integration into higher-order functions. This function is fast even in a single cycle, or may be pipelined for even higher speed. | Download |
| Exponent (Base K) | Fast and low power Base K Exponentiation, highly accelerated, for standalone use or integration into higher-order functions. This function is fast even in a single cycle, or may be pipelined for even higher speed. | Download |
| Exponent (natural) | Fast and low power Base e Exponentiation, highly accelerated, for standalone use or integration into higher-order functions. This function is fast even in a single cycle, or may be pipelined for even higher speed. | Download |
| Floating Point Multiply-Accumulate | Floating Point Multiply-Accumulate unit, highly accelerated, for standalone use or integration into higher-order AI functions | Download |
| Gaussian Error Linear Unit (GELU) | This highly accelerated function implements the Gaussian Error Linear Unit (GELU) activation function that is popular in transformers. This function is fast even in a single cycle, or may be pipelined for even higher speed. | Download |
| Logarithm (Base 10) - Log10 | Base-10 Log generator, highly accelerated, for standalone use or integration into higher-order functions | Download |
| Logarithm (Base 2) - Log2 | Base-2 Log generator, highly accelerated, for standalone use or integration into higher-order functions | Download |
| Log (Base K) | Base-K Log generator, highly accelerated, for standalone use or integration into higher-order functions | Download |
| Logarithm (natural) | Natural Log generator, highly accelerated, for standalone use or integration into higher-order functions | Download |
| Mish Activation Function | This fast and low-power mish function is a great improvement over simple activation functions | Download |
| Multiplier | This highly accelerated, low gate count multiplier for AI can be integrated into MAC and MatMul applications or used standalone where needed | Download |
| Reciprocal Function | Fast and low power reciprocal function | Download |
| Reciprocal of Square Root | Fast and low power reciprocal of the square root function | Download |
| Sigmoid activation function | This fast and low-power sigmoid function is a great improvement over simple activation functions | Download |
| SiLU-SwiGLU Activation Function | This fast and low-power SiLU function is a great improvement over simple activation functions, can be used alone or together with a multiplier to create the SwiGLU function. Similar to Swish but with Beta fixed to 1. | Download |
| Sine - Cosine | Fast and low-power selectable sine or cosine function | Download |
| Sine | Fast and low-power sine function | Download |
| SoftMax Accelerator | Cassia.ai SoftMax functions are incredibly fast and low power, and may be piplined as needed. Our SoftMax functions are built to order to meet each customer's exact requirement, and use this function as a key primitive. | Download |
| Square Root | Square Root function | Download |
| Swish-SwiGLU activation funcion | This fast and low-power Swish function is a great improvement over simple activation functions, can be used alone or together with a multiplier to create the SwiGLU function. Similar to SiLU but with the addition of a Beta parameter. | Download |
| Tangent | Fast and low-power tangent function | Download |
| Tangent (Hyperbolic) | Fast and low-power hyperbolic tangent function, also suitable for use as an activation function | Download |
Cassia.ai in the News.
Accelerate Your Silicon Roadmap.
Contact our architecture team to evaluate RTL IP integration, benchmark data on your workloads, or discuss licensing.