Tigris microprocessor
By Cornami
The Tigris microprocessor from Cornami is a production-proven morphable silicon solution designed for workloads requiring deterministic execution, sub-millisecond reconfiguration, and elastic computation.
Publisher review
The Tigris microprocessor from Cornami is a production-proven morphable silicon solution designed for workloads requiring deterministic execution, sub-millisecond reconfiguration, and elastic computation. It targets industries where these capabilities are critical, such as AI, signal processing, and cryptography. The microprocessor is part of Cornami's end-to-end compute solutions, which include custom silicon, system software, and workload mapping, engineered for specific outcomes rather than sold as off-the-shelf products.
Cornami's approach is particularly suited for environments where AI-generated code demands hardware that can adapt at the speed of software. The Tigris microprocessor features a FracTLcore fabric with 2,048 reconfigurable compute units, each with private local memory, and supports full IEEE-754 floating point and integer operations. It is built on a 16 nm process technology and includes 8 GB of HBM2 memory.
The MX64 system, which integrates 32 Tigris silicon units, scales to 65,536 FracTLcores and 256 GB HBM2, offering rack-scale acceleration. This architecture is stream-centric, following the TruStream computing model, which emphasizes dataflow graphs over thread-centric approaches. Cornami positions the Tigris microprocessor against competitors like Hewlett Packard Enterprise, Groq, and Cerebras by emphasizing its unique combination of deterministic execution, morphable architecture, and elastic computation.
The company has raised $212M in funding from investors like SoftBank Vision Fund and Applied Materials, underscoring its market potential. However, the Tigris microprocessor's niche focus on morphable silicon may limit its appeal to broader markets that do not require such specialized capabilities. Additionally, the lack of disclosed pricing information makes it difficult to assess its cost-effectiveness compared to alternatives. The reliance on a 16 nm process, while sufficient for current needs, may also pose challenges as industry standards move toward more advanced nodes.
How it works
-
Deterministic execution
Every compute element operates on a compile-time fixed cycle, ensuring predictable latency and performance under load.
-
Sub-millisecond reconfiguration
Compute, memory, and interconnect can reconfigure in approximately 1 ms, adapting to workload changes dynamically.
-
Elastic computation
Power consumption tracks active compute units, eliminating waste from idle regions and global clock domains.
-
FracTLcore fabric
A reconfigurable fabric of 2,048 compute units with private local memory and four layers of interconnect.
-
TruStream computing model
Stream-centric programming model that describes dataflow graphs for continuous execution with compiler-guaranteed reuse.
-
IEEE-754 support
Full floating point and integer operation support, enabling diverse applications from AI to cryptography.
-
Rack-scale acceleration
MX64 system scales to 65,536 FracTLcores and 256 GB HBM2, delivering turnkey rack-scale performance.
Strengths and trade-offs
Strengths
- The Tigris microprocessor offers deterministic execution with p99.99 latency equal to p50, ensuring predictable performance under load.
- Sub-millisecond reconfiguration allows the hardware to adapt dynamically to changing workloads, a unique capability in the market.
- Elastic computation ensures power efficiency by tracking active compute units, reducing energy waste from idle components.
- The FracTLcore fabric provides 2,048 reconfigurable compute units with private local memory, enabling high parallelism and flexibility.
Trade-offs
- The niche focus on morphable silicon may limit adoption in markets that do not require such specialized reconfiguration capabilities.
- Lack of disclosed pricing information makes it difficult to compare cost-effectiveness against competitors.
- Built on a 16 nm process, the Tigris may face challenges as industry standards shift to more advanced nodes.
- The stream-centric TruStream model may require significant adaptation for developers accustomed to thread-centric programming.
Pricing context
Not specified in the sources
Getting started with Tigris microprocessor
-
Contact Cornami
Reach out to Cornami directly to inquire about purchasing or licensing the Tigris microprocessor for your specific use case.
-
Define workload requirements
Identify and document the specific workload requirements, such as AI, signal processing, or cryptography, to ensure compatibility with the Tigris microprocessor.
-
Configure system software
Set up and configure the system software provided by Cornami to map your workloads onto the Tigris microprocessor's architecture.
-
Deploy MX64 system
Integrate the MX64 system, which includes 32 Tigris silicon units, into your infrastructure to achieve rack-scale acceleration.
-
Monitor and optimize
Continuously monitor the performance of the Tigris microprocessor and optimize configurations to maintain efficient and deterministic execution.
Frequently Asked Questions
What is the Tigris microprocessor?
The Tigris microprocessor is Cornami's morphable silicon solution designed for deterministic execution and elastic computation. It features 2,048 reconfigurable compute units with private memory, targets AI and cryptography workloads, and supports sub-millisecond hardware reconfiguration. Built on 16nm technology, it emphasizes dataflow over traditional thread-centric approaches.
What industries use Tigris microprocessors?
Tigris microprocessors serve industries requiring deterministic performance and hardware adaptability, including AI model acceleration, real-time signal processing, and cryptographic applications. Its sub-millisecond reconfiguration suits dynamic workloads where traditional fixed-architecture processors struggle to maintain efficiency and predictable latency.
How does Tigris compare to Groq and Cerebras processors?
Tigris differentiates with morphable silicon that reconfigures in 1ms, deterministic execution guarantees, and elastic power scaling. Unlike Groq's static tensor streaming or Cerebras' wafer-scale approach, Tigris combines reconfigurable FracTLcore fabric with compiler-managed dataflow for adaptive performance in changing workloads.
What are the key features of Tigris architecture?
Tigris features 2,048 reconfigurable FracTLcores with local memory, IEEE-754 floating point support, and TruStream dataflow programming. Its MX64 system scales to 65,536 cores with 256GB HBM2. Unique capabilities include deterministic execution, 1ms reconfiguration, and power tracking that scales with active compute units.
What are the limitations of Tigris processors?
Tigris' 16nm process may lag behind industry shifts to advanced nodes. Its niche morphable architecture requires workload specialization, and undisclosed pricing complicates cost comparisons. Developers must adapt to TruStream's dataflow model, differing from conventional thread-based programming approaches.
How does Tigris handle AI workload acceleration?
Tigris accelerates AI through its reconfigurable fabric that adapts to model changes in 1ms, with deterministic latency for inference tasks. The 2,048-core fabric supports floating-point operations essential for neural networks, while elastic computation optimizes power use during variable AI workload phases.
Alternatives
How Tigris microprocessor compares
Direct head-to-head against 2 competitors. Picked by 7wData.
Tigris microprocessor
- Pricing
- Not specified in the sources
- Target
- The Tigris microprocessor from Cornami is a production-proven morphable silicon solution designed for workloads requiring deterministic execution, sub-millisecond reconfiguration, and elastic computation.
- Strength
- The Tigris microprocessor offers deterministic execution with p99.99 latency equal to p50, ensuring predictable performance under load.
- Watch for
- The niche focus on morphable silicon may limit adoption in markets that do not require such specialized reconfiguration capabilities.
Duality
- Pricing
- Custom/Contact sales
- Target
- Enterprises needing FHE for regulated data
- Deployment
- On-prem, cloud
- Strength
- Pure-software FHE solutions with cross-platform support
- Watch for
- Performance overhead vs hardware-accelerated approaches
Cerebras
- Pricing
- $2M+ per CS-3 system
- Target
- HPC and AI workloads at wafer-scale
- Deployment
- On-prem rack
- Strength
- Wafer-scale engine for massive parallelism
- Watch for
- Limited FHE-specific optimizations
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.