Denglin Goldwasser (GPU+)
Product Overview
Denglin Technology (Denglin / GLead) was founded in 2017; its core team members mostly came from GPU companies such as Vivante, and it is dedicated to building high-performance, high-efficiency general-purpose AI acceleration solutions with its proprietary GPU+ architecture. Its first product, the Goldwasser series, achieved Full Mask mass production on the TSMC 12nm process in June 2020 and completed production introduction at industry-leading companies in smart security, the internet, smart cities, and more.
Goldwasser's core innovation is GPU+ (a software-defined in-chip heterogeneous architecture): by organically combining a "programmable GPGPU engine" and an "efficient Tensor engine" within a single chip, it resolves the tension between generality and efficiency in a dataflow-driven manner, and simultaneously supports inference and training (training-inference integrated). It is compatible with the CUDA / OpenCL programming models, lowering customers' migration cost from the NVIDIA ecosystem — a key differentiator the company emphasizes.
Goldwasser offers multiple models covering edge to data center: edge Goldwasser UL (25-35W, 32-64 TOPS), half-height half-length server card Goldwasser L (40-70W, 128-256 TOPS), and full-height full-length Goldwasser XL (512 TOPS).
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | GPU+ in-chip heterogeneous (programmable GPGPU engine + efficient Tensor engine, dataflow-driven, CUDA/OpenCL compatible) |
| Process Node | TSMC 12nm |
| FP16 / BF16 Compute | Not disclosed (training-inference integrated; specific floating-point figures not listed by the vendor) |
| INT8 Compute | 32-512 TOPS (by model: UL 32-64 / L 128-256 / XL 512 TOPS) |
| FP32 Compute | Not disclosed |
| Memory Capacity | Not disclosed |
| Memory Type | Not disclosed |
| Memory Bandwidth | Not disclosed |
| TDP | 40-70 W (typical range for Goldwasser L; UL 25-35W, XL undisclosed) |
| Interconnect | Not disclosed |
| Interface | PCIe (specific version not listed by the vendor) |
| Launch | 2020-06 |
| Mass Production/Availability | 2020-06 (Full Mask mass production) |
Note: In testing, Goldwasser delivered 128 TOPS at a 40W TDP (i.e., the lower bound of Goldwasser L); the vendor claims 3-10× computational efficiency gains over mainstream products on different neural networks at the same process node, with reduced dependence on external memory throughput. FP16/BF16, FP32, and memory specifications are not officially disclosed, hence "Not disclosed".
Key Features
- GPU+ in-chip heterogeneity = generality + efficiency: schedules the GPGPU engine and Tensor engine in a dataflow manner, balancing the CUDA-compatible ecosystem with high compute density.
- Training-inference integrated: the same architecture supports both AI inference and training, unlike most inference-only ASICs.
- Software-defined: software definition lets different neural networks maximize hardware performance and energy efficiency, avoiding the data-exchange overhead of system-level heterogeneity.
- CUDA/OpenCL compatible: significantly lowers customer migration cost; existing ecosystem assets can be reused directly.
- Full-scenario coverage: from 25W at the edge to 512 TOPS in the data center, with a complete model gradient.
Vendor Information
| Item | Details |
|---|---|
| Company | Suzhou Denglin Technology Co., Ltd. (GLead) |
| Headquarters | Suzhou, China (R&D/office sites in Shanghai, Beijing, Hangzhou, Chengdu, Xi'an, Shenzhen, and elsewhere) |
| Founded | 2017 |
Use Cases
- ✅ Cloud/edge AI inference and training (training-inference integrated, covering security, internet, smart cities)
- ✅ Migration of existing CUDA-ecosystem models (hardware-compatible with CUDA/OpenCL)
- ✅ High-density data center accelerator cards (Goldwasser XL 512 TOPS)
- ❌ Ultra-low-power edge (the UL starts at 25W, still above most edge NPUs)
- ❌ Fully proprietary instruction set ecosystem (CUDA compatibility is the priority; the proprietary ecosystem is still maturing)
Related Cards
- TsingMicro TX81 — Chinese reconfigurable computing innovative architecture
- Vastai VA10 — Chinese cloud inference accelerator card
- Lightmatter Envise — International compute-in-memory route comparison