趋近智
The NVIDIA RTX 6000 Blackwell (96GB) is equipped with 96 GB of dedicated VRAM and 1792 GB/s memory bandwidth. It provides 100% addressable VRAM for CUDA and TensorRT-LLM runtimes, supporting full precision and quantized inference.
总显存
96 GB
可用显存上限
96 GB
显存带宽
1792 GB/s
最大推理模型级别
70B+ Models (Q4)
热设计功耗
500W
随上下文增长模拟的显存需求。当曲线超过参考线时将导致显存不足(OOM)。
量化精度:
不同量化精度下可运行的最大模型参数级别。
上下文长度:
4位量化
70B+ Models
8位精度
70B Models
16位全精度
32B Models
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
70B+ Models
LoRA
70B Models
全参数微调
7B-8B Models
Guidance for optimal precision and operational ceilings on NVIDIA RTX 6000 Blackwell (96GB).
Inference Sweet Spot
Suitable for 70B parameter models at Q4/Q8 with extended 32k context windows, or high-throughput batching of smaller dense architectures.
Bandwidth & Speed Profile
With 1792 GB/s aggregate bandwidth, batch size 1 inference operates in a memory-bandwidth bound regime. Generates approximately 398 tok/s on an 8B Q4 model and 47 tok/s on a 70B Q4 model.
APX AI
在线