Specifications
Architecture
Kalray MPPA (Massively Parallel Processor Array), KVX VLIW cores
Coolidge v2 Processor
MPPA Coolidge v2 @ 1 GHz, 80 KVX cores (standalone design-in SoC)
Coolidge v2 Acceleration
Integrated AI accelerator, cryptography, FEC and data compression
Coolidge v2 Interfaces
Ethernet, PCIe Gen 4, DDR4
K200-LP Processor
1× MPPA Coolidge v1 @ 1 GHz, 80 cores
K200-LP Memory
8 GB DDR4 @ 3200 MT/s
K200-LP Network
2× 100G Ethernet
K200-LP Host
PCIe Gen 4, 16-lane
K200-LP Offload
NVMe-oF (RoCE / TCP), encryption, compression, erasure coding
K300 Processor
1× MPPA Coolidge v2 @ 1 GHz, 80 cores
K300 Memory
16 GB DDR4 @ 3200 MT/s
K300 Network
4× 10/25G Ethernet
K300 Performance
40 TOPS (INT8) · 20 TFLOPS (FP16)
K300 Offload
FEC (LDPC), cryptography, data compression
TC4 Processor
4× MPPA Coolidge v2 @ 1 GHz, 320 cores
TC4 Memory
32 GB DDR4, PCIe Gen 4 16-lane host
TC4 Performance
160 TOPS (INT8) · 80 TFLOPS (FP16)
Turbo Server TS1
2U, 2× AMD EPYC 9004 (168 cores), 8× TC4 cards
Turbo Server Performance
1280 TOPS (INT8) · 640 TFLOPS (FP16)
Turbo Server Memory
24 slots of 12-channel RDIMM DDR5
Turbo Server Expansion
8× PCIe Gen 5 ×16 FHFL slots
Software
Kalray SDK with 80+ AI libraries, Docker, Triton Inference Server API
Overview
Kalray's MPPA (Massively Parallel Processor Array) DPUs are production-deployed accelerators built around large arrays of KVX VLIW cores. Unlike GPUs, they execute very large numbers of parallel, asynchronous operations — an architecture that suits mixed classical and AI processing workloads such as smart vision, storage offload and network packet processing.
The family covers a range of deployment shapes: the Coolidge v2 processor as a standalone design-in SoC; the K200-LP for storage, security and networking offload with NVMe-oF over RoCE/TCP; the K300 for AI and smart vision at 40 TOPS INT8; and the TC4 with four Coolidge v2 processors delivering 160 TOPS INT8 for demanding industrial AI inference.
For data-intensive smart vision, the 2U Turbo Server TS1 combines two AMD EPYC 9004 CPUs with eight TC4 cards for 1280 TOPS INT8 across eight PCIe Gen 5 ×16 FHFL slots, and ships with the open Kalray SDK — standard toolchains, Docker and a Triton Inference Server API, with no proprietary framework lock-in.
Key Benefits
Open programmability: standard toolchains, 80+ AI libraries, Docker and a Triton Inference Server API — no vendor lock-in. Complementary to GPUs: massive parallel asynchronous execution suits vision, storage and packet workloads alongside GPUs. Production-ready: deployed today in OEM and infrastructure vendor systems with pre-validated performance.
Applications
AI inference at the edge and in industrial vision · NVMe-oF storage acceleration · Network and security offload · Industry 4.0 machine vision · Telecom infrastructure processing
Request a Quote — KALRAY MPPA DPU CARDS — COOLIDGE V2, K200-LP, K300 & TC4 ACCELERATORS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →