趋近智
The Arc A770 (16GB) provides 16 GB of memory with 560 GB/s bandwidth, supporting local inference via ROCm and Vulkan acceleration.
总显存
16 GB
可用显存上限
16 GB
显存带宽
560 GB/s
最大推理模型级别
14B Models (Q4)
热设计功耗
225W
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
14B Models
8位精度
8B Models
16位全精度
3B Models
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
14B Models
LoRA
8B Models
全参数微调
1B Models
Guidance for optimal precision and operational ceilings on Arc A770 (16GB).
Inference Sweet Spot
Optimal for 8B models across extended context (32k+). 14B models fit comfortably at Q4 precision with standard context.
Bandwidth & Speed Profile
With 560 GB/s aggregate bandwidth, batch size 1 inference operates in a memory-bandwidth bound regime. Generates approximately 124 tok/s on an 8B Q4 model and 15 tok/s on a 70B Q4 model.
APX AI
在线