Thanks to our customers, Quadric has outgrown its headquarters. Effective June 18, our new address is 270 East Lane, Burlingame, CA 94010.
The Chimera SDK
Develop, Simulate, Profile, and Deploy. The power of a code-driven processor with the convenience of no-code graph compilation.
Inference FrameworkPyTorch / TensorFlow
Graph CompilerOptimization
C++ CompilerLLVM
BinaryExecutable
SoCDeployment
Inference Framework
Graph Compiler
C++ Compiler
Binary
SoC
Development Workflow
Complete Development Environment
From AI models to deployment-ready code in a single unified workflow
The Quadric Chimera Software Development Toolkit (SDK) is a comprehensive environment for the development of complex application code targeting the Chimera processor. The Quadric SDK enables the mixing and matching of any data parallel algorithm whether it is expressed as machine learning / AI graph or as traditional C++ code.
Preparing an AI model to run on a Chimera core is simple and tool-driven
- No special network preparation
- No substitution of operators
- No removal of layers
- No reliance on Quadric to prepare models
Quadric users can download a complete Docker image of the Chimera SDK for deployment on private clouds or on-premise systems. Once installed on your compute resources, you can compile ONNX graphs into C++ using the Chimera Graph Compiler, compile your own proprietary C++ code, combine them and profile on the cycle accurate ISS model of the Chimera GPNPU.
SDK Workflow Overview
Machine LearningTraining FrameworksONNX formatTrained NN Graph FP32(for SoC perf. modeling)Trained NN Graph -INT8 (4b/8b weights)(Recommended forproduction deployment)Int8 Auto-QuantizationONNX Runtime / RelayRelay formatChimera Graph Compiler (CGC)Proprietary Chimera I.R. w/ numerous optimization passesC++ Code(using CCL API)C++C++User ApplicationCode (C++)Chimera LLVM C++ CompilerOptimized C++Function LibrariesChimera ISS(Cycle approximate / System C)Pre-packaged SDKDocker Image containsthese ComponentsTarget Silicon withChimera GPNPUOpen SourceQuadric Tooling (binary)Quadric Library (source)
Graph Compilation
Quadric Chimera Graph Compiler
Powerful graph conversion and code optimization for AI inference models
The Quadric Chimera Graph Compiler (CGC) is a powerful graph conversion and code optimization tool that inputs a machine learning inference model created from the leading AI training frameworks, performs numerous optimizations, and outputs an optimized C++ code representation of the AI graph utilizing the Chimera Compute Library (CCL) for later compilation by the Chimera LLVM C++ compiler.
Graph Import
CGC accepts input graphs in ONNX format for networks developed in PyTorch, TensorFlow or other frameworks. A number of optimizations are performed as part of the graph import phase:
- Graph simplification / canonicalization
- Constant propagation - removing operators with purely constant arguments, if possible
- Operator legalization / conversion – converting to GPNPU-specific forms of AI operators
The shape and structure of the network graph is optimized to simplify operators where possible. Compatibility checks are performed to determine if all operators are supported. When an unsupported or custom operator is present in the source graph, the user has the option to partition the graph around the custom operator, write a C++ kernel for the unsupported operator, and reintegrate the new custom operator C++ kernel with code auto-generated by CGC.
Graph and Memory Optimization
CGC creates a full intermediate representation of the input graph and performs multiple passes of optimization with the twin aims of optimizing performance and memory bandwidth utilization.
Memory optimization techniques include:
- Tensor format layout analysis (both L2 and LRM)
- Fusion in Local Memory (FILM) - merging of operations to preserve data within LRM and avoid costly intermediate activation writeback.
- Array-level memory adapters & tensor shape transformations
- Bank conflict minimization, defragmentation
- Predictive weight prefetching
OPERATOR FUSION
Automatic Graph Optimization
The compiler automatically identifies and merges compatible operators. Fused operations keep data in local PE registers, eliminating L2 memory round-trips and dramatically reducing power consumption.
Original GraphConv2DBatchNormL2ReLUL2AddL2Conv2DL2SigmoidL2FusedFusedConvBlock6 ops → 1 kernelData stays in PE registersNo intermediate L2 accessImpactMemory TrafficBaselineL2 Memory Accesses5 round-tripsPower ConsumptionHighUnoptimizedBefore fusion
OPTIMIZATION PASSES
Multi-Stage Transformation
Each pass analyzes and transforms the graph for the Chimera architecture.
Graph Optimization Passes✓Graph ParsingParse ONNX/PyTorch model structureComplete✓Constant FoldingPre-compute static tensor valuesComplete✓Dead Code EliminationRemove unused graph pathsComplete✓Operator FusionMerge compatible adjacent operatorsComplete✓Memory PlanningOptimize tensor layout and L2 allocationComplete✓Code GenerationEmit optimized Chimera instructionsComplete
Compilation & Simulation
Chimera LLVM C++ Compiler & Instruction Set Simulator
Industry-standard LLVM infrastructure with cycle-accurate simulation and profiling
The Chimera LLVM C++ compiler utilizes the industry state of the art LLVM compiler infrastructure with a Quadric-specific code generation back end that emits assembly code specific to the Chimera instruction set.
The Chimera Instruction Set Simulator (ISS) is a pipeline-accurate executable model of the Chimera processor core. The ISS can be utilized both in a standalone mode to profile and tune application code in isolation, or as a callable System C transaction-level model bundled into a more comprehensive virtual prototype of an SoC where more accurate memory model behavior can be used to fine-tune your Chimera code more precisely.
Comprehensive Profiling
The Chimera ISS provides extensive profiling of cycle counts, data bandwidth usage, and power profiles of each unique application workload.
Cycles050,000100,000150,000200,000250,000300,000350,000data_ocmdata_externalmacdata_arraycomputeProfile Categories
Average Power025050075010001250150017502000225025001234567891011Core:SUMInt:SUMQIs:SUMNb:SUMNBRSUMIterator Region
BW (GB/s)05101520253035012345678910ExtBytes_LOAD_BWExtBytes_STORE_BWIterator Region
Cycles010,00020,00030,00040,00050,00060,00070,00080,00090,000100,000computedata_arraymacdata_externaldata_ocm