- 精华
- 1
- 帖子
- 1079
- 威望
- 1 点
- 积分
- 1106 点
- 种子
- 34 点
- 注册时间
- 2012-11-3
- 最后登录
- 2025-2-26
|
发表于 2021-6-16 23:29 · 河北
|
显示全部楼层
说真的,XSX和XSS的CU真是太太太不像RDNA2的了,微软自己在ISSCC上的演讲中说,为实现12T浮点性能,测试了[email protected]和[email protected]两个方案,测试结果是[email protected]的功耗比[email protected]高20%,但鉴于保证芯片中56CU全部可用的良率成本,微软最终屏蔽4个CU,承受着额外20%功耗的代价选择了[email protected]。
原文:
Paul Paternoster explained that from chips coming off the production line, a substantial number could run with all 28 WGPs enabled. The goal of the graphics was to provide 12 TFLOPs of performance, and so by some simple math, Microsoft could do either of the following to hit that number:
28 WGPs enabled at 1675 MHz
26 WGPs enabled at 1825 MHz
Both of these configurations enable 12 TFLOPs. Because the frequency of the 28 WGP design is lower, this also enables a lower voltage, combined for an overall power saving of 20% if all 28 WGPs are used.
Of course, a 20% power saving is quite substantial, as it would either enable better performance per watt, or enable higher performance. But the issue is that not enough processors were coming off of the production line with all 28 WGP running at this frequency. The variability of the processors, due to both transistor performance and defects, meant that 28 WGP versions didn’t make sense financially.
从[email protected]到[email protected],CU少了4个(少了7.7%),频率提升150MHz(提升了9%),功耗就涨了20%?要知道RDNA2的一大重点改进就是CU的高频能力,可以在很高的频率下功耗仍不失控,XSX这个GPU的频率功耗曲线真的是RDNA2?
另外按微软自己在ISSCC上说的,XSX最高功耗270瓦,6700XT是整卡230瓦,降点频率把浮点控到12T的话功耗也就210瓦,再去掉infinity cache,连200瓦都不见得有,再捆上8颗Zen2 CPU核心,把显存位宽扩到320bit,加4颗GDDR6显存,功耗会比270瓦更高吗?既然微软顾及到良率问题宁可付出20%功耗代价也选择了52CU,那navi22这40CU的高良率小核心为什么不选呢?6700XT飙到2.58GHz也没见有什么功耗代价啊? |
本帖子中包含更多资源
您需要 登录 才可以下载或查看,没有帐号?注册
x
|