提议的假设（“通用数字”）实现 Noreferrer“>似乎很有趣。
是否有人执行任何基准来将他的实施（在软件）与现有硬件浮子速度进行比较？
我很好奇是否在实施此速度方面是否有任何速度优势，用于使用现有硬件支持用作默认通用数字系统VS的编程语言。这似乎令人怀疑，但当然值得好奇。
The proposed posits ("universal numbers") implementation by Michael Feldman seems quite interesting.
Has anyone performed any benchmarks to compare his implementation (in software) against existing hardware floats speeds?
I am curious as to whether there's any speed advantage at all in implementing this for a programming language to use as a default universal number system vs using existing hardware support. It seems doubtful, but certainly worthy of curiosity.
发布评论
评论（1）
我绝对远离专家，但是如果有的话，我自己对该主题的研究使我进入 https ：//ieeexplore.iee.org/document/8892116 和（除其他），显然有几个权衡会影响这种基准的结果：对于某些人来说，假定需要30％的硅面积才能获得类似的硅？功能。对于其他人来说，应在性能方面将其与float64进行比较，因此可以预期获得35％的胜利。但是，无论您是在硅中实施问点还是仅在硅中的一部分，也将严重影响性能。我个人选择研究另一种用例：在Arduino板上发现的Atmega328使用16位（和8位）假设，以替换简化的RL算法的FLOAT32计算。因此，简而言之，您的里程很可能会根据您感兴趣的领域而有所不同。
哦，顺便说一句，费尔德曼先生确实写了一篇关于Poitit的文章，他没有实施。
I'm definitely far from an expert, but if anything, my own research on the subject has led me to https://ieeexplore.ieee.org/document/8892116 and https://arxiv.org/abs/2305.06946 (among others), and there are clearly several trade-offs that will influence the result of such a benchmark : for some, Posit would need 30% more silicon area for similar??? functionality. For others, Posit32 should be compared to Float64 in terms of performance, so a win of 35% could be expected. But whether you implement a quire or only part of it in silicon will also be seriously impacting the performance. I personally chose to investigate another use case : using 16-bit (and 8-bit) Posits for the ATmega328 found in Arduino boards to replace float32 calculations for simplified RL algorithms. So in short, very likely your mileage will vary according to your domain of interest.
Oh BTW, Mister Feldman did write an article on Posit, he didn't make an implementation.