Tensor Cores
- Tensor engines
- 1,024 matrix engines
- Precision
- FP8, BF16, FP16, TF32
AI Accelerator
A1
Purpose-built intelligence at infrastructure scale.
A dense tensor accelerator for high-performance model training, low-latency inference, and multimodal AI systems.
Request Dev AccessTensor compute
4.8 PFLOPS FP8
Thermal design power
700 W
Process node
2 nm
Memory
256 GB HBM4
Technical specifications
Memory
- Capacity
- 256 GB HBM4
- Bandwidth
- 10 TB/s
Interconnect
- Fabric
- RSVC Link 3.0
- Scale
- Up to 1,024 accelerators
Power
- TDP
- 700 W
- Cooling
- Liquid cooling optimized
Software Stack
- Compiler
- RSVC Forge compiler
- Runtime
- Distributed training runtime
Built for demanding workloads
LLM Training
Train frontier-scale language models with efficient distributed compute.
Inference at Scale
Serve responsive models across high-volume production endpoints.
Multimodal AI
Bring vision, language, audio, and sensor data into one system.
Benchmark comparison
Normalized performance index. Higher is better.
FP8 training
RSVC
100
Competitor A
77
Competitor B
70
Inference throughput
RSVC
100
Competitor A
81
Competitor B
74
Interconnect scale
RSVC
100
Competitor A
72
Competitor B
66
Software ecosystem
PyTorchJAXTensorFlowONNX
Request an engineering sample
Talk with our platform team about A1 validation and deployment planning.