跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators$599

AMD Radeon RX 9070 XTAMD Radeon RX 9070 XT

$599$599 · 美国起售价,2025 年发布时口径。US launch MSRP.

RDNA 4 的 16 GB 卡:64 CU、4,096 流处理器、640 GB/s GDDR6、64 MB Infinity Cache,FP32 48.7 TFLOPS,FP16 矩阵 195 TFLOPS,FP8 389 TFLOPS,卡功耗 304 W,建议 750 W。ROCm 已能跑主流模型,但 CUDA 专属 kernel、量化库和多卡生态仍更成熟。适合预算优先的 16 GB 本地推理和 Linux/ROCm 用户,不适合必须依赖 CUDA 的工作流。An RDNA 4 16 GB card: 64 CUs, 4,096 stream processors, 640 GB/s GDDR6, 64 MB Infinity Cache, 48.7 TFLOPS FP32, 195 TFLOPS FP16 matrix, 389 TFLOPS FP8, 304 W board power and a 750 W recommendation. ROCm runs mainstream models, but CUDA-only kernels, quantization libraries and multi-GPU support remain more mature. A budget-first 16 GB local-inference card for Linux/ROCm users, not for workflows locked to CUDA.

关键参数Key specs

显存 / memory
16 GB GDDR6, 256-bit
带宽 / bandwidth
640 GB/s
计算单元 / compute units
64 CU / 4,096 流处理器
FP32
48.7 TFLOPS
FP16 矩阵 / FP16 matrix
195 TFLOPS(稀疏 389 TFLOPS)
FP8 矩阵 / FP8 matrix
389 TFLOPS(稀疏 779 TFLOPS)
Infinity Cache
64 MB
卡功耗 / board power
304 W
系统电源 / required PSU
750 W
互联 / interconnect
PCIe 5.0 x16

官方原文怎么说What the source says

Max Memory Size 16 GB
1 https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
Memory Bandwidth Up to 640 GB/s
2 https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
Compute Units 64
3 https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
Typical Board Power (Desktop) 304W
4 https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html

资料与许可Source and licence

资料时间Source checked
2026-11-20
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
类目Category
加速卡Accelerators
状态Status
已出货Shipping
适合Use cases
本地推理Local inference
价格口径Price basis
$599 · msrp

同类目还有More in this category

下一步Next step

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute