Skip to main content

PEZY-SC2

Vendor: PEZY Computing

Category: ASIC Dedicated Accelerator

Architecture: MIMD Many-Core Processor

Introduction​

PEZY-SC2 is a MIMD (Multiple Instruction Multiple Data) many-core processor developed by Japan's PEZY Computing, optimized for high-performance computing (HPC) and scientific computing. The chip integrates 2,048 processing elements (PEs) using a hierarchical architecture (Prefecture→City→Village→PE), supporting 16,384 concurrent threads. PEZY-SC2 is deployed in the Gyoukou supercomputer, combining immersion liquid cooling for high-density deployment.

Specifications​

ModelCompute (Peak)MemoryInterfaceTDPProcess
PEZY-SC24.1 TFLOPS (FP64) / 8.2 TFLOPS (FP32) / 16.4 TFLOPS (FP16)DDR4 128GB (100 GB/s)PCIe 3.0/4.0 x16 ×2130W16nm FinFET

Architecture Highlights​

FeatureDescription
Core Architecture2,048 PE, hierarchical interconnect (8 Prefecture → 128 City → 512 Village → 2048 PE)
Thread Concurrency16,384 hardware threads (8 threads per PE)
SIMD Support64-bit SIMD, single PE simultaneously executes 1×DP / 2×SP / 4×HP
Cache HierarchyL1 D-cache 4MB + L2 D-cache 8MB + LLC 40MB + Atomic Cache 16KB
Onboard CPUMIPS64 R6 (P6600) 6 cores, for host-side management
InterconnectPrefecture-level X-bar crossbar, 6 levels of synchronization granularity

Performance Benchmarks​

BenchmarkPerformanceNotes
FP64 Peak4.1 TFLOPSDouble precision floating point
FP32 Peak8.2 TFLOPSSingle precision floating point
FP16 Peak16.4 TFLOPSHalf precision floating point
Energy Efficiency~31.5 GFLOPS/W (FP64)130W TDP
Gyoukou HPL1.67 PFLOPSNov 2018 TOP500

Quick Installation​

# 1. Install PEZY SDK (requires contacting vendor)
# Download: https://www.pezy.co.jp/en/products/

# 2. Set environment variables
export PEZY_HOME=/opt/pezy
export PATH=$PEZY_HOME/bin:$PATH
export LD_LIBRARY_PATH=$PEZY_HOME/lib:$LD_LIBRARY_PATH

# 3. Compile sample program
pzcc -o hello_pz hello_pz.c

# 4. Run
./hello_pz

Code Examples​

// PEZY-SC2 OpenCL example: vector addition
#include <CL/cl.h>
#include <stdio.h>

int main() {
    // Get PEZY-SC2 OpenCL platform
    cl_platform_id platform;
    clGetPlatformIDs(1, &platform, NULL);

    cl_device_id device;
    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ACCELERATOR, 1, &device, NULL);

    cl_context ctx = clCreateContext(NULL, 1, &device, NULL, NULL, NULL);
    cl_command_queue queue = clCreateCommandQueue(ctx, device, 0, NULL);

    printf("PEZY-SC2 device ready\n");
    printf("Max compute units: query device info for PE count\n");

    clReleaseCommandQueue(queue);
    clReleaseContext(ctx);
    return 0;
}

Pricing​

ProductPrice RangeNotes
PEZY-SC2 ModuleContact vendorRequires contacting vendor for quote
PEZY SDKFreeNDA signing required

Model Compatibility​

Model TypeSupportNotes
Scientific Computing✅ GoodPrimary use case (CFD, molecular dynamics, etc.)
Traditional HPC✅ GoodLinpack, HPCG and other benchmarks
Deep Learning Inference⚠️ LimitedNot a primary design goal, requires adaptation
Deep Learning Training❌ Not supportedNo CUDA/Tensor Core matrix acceleration
LLM❌ Not supportedNo targeted optimization

OS Support Matrix​

OSSupportNotes
Linux (CentOS/RHEL)✅Primary support platform
Linux (Ubuntu)✅Supported
Windows❌Not supported

Version History​

VersionDateDescription
PEZY-SC22017First release, 2,048-core MIMD architecture
Gyoukou Launch2017-11TOP500 #11, HPL 1.67 PFLOPS
Gyoukou Expansion2018-11Expanded to 19,856 nodes, TOP500 list

Large-Scale Cluster Deployments​

Based on global AI supercomputing cluster statistics, PEZY-SC2 has accumulated over 11,600 chips deployed across 1 publicly disclosed cluster.

Chip Model Statistics​

Chip ModelTotal DeployedCluster Count
PEZY-SC211,6001

Notable Deployment Clusters Top 10​

#Cluster NameTotal ChipsChip ModelOperator
1JAIST Gyoukou11,600PEZY-SC2 ×11,600JAIST, Japan

If you're evaluating alternatives, the following products may also fit your scenario: