使用--config = mkl与--config = cuda时的张量流构建工件之间的差异

Question

我正在从源代码为cpu和gpu构建张量流。

对于GPU 如此处所述https://www.tensorflow.org/install/install_sources

from collections import Counter

words = ["paper", "Plastic", "aluminum", "PAPer", "TIN", " paper", "glass", "tin", "PAPER", "Polypropylene Plastic"]

counts = Counter()
result = []

for word in words:
    caseless = word.casefold()

    if caseless not in counts:
        result.append(word)

    counts[caseless] += 1

result = ['%s (%d)' % (w, counts[w.casefold()]) if counts[w.casefold()] > 1 
                                                else w for w in result]

print(result)

对于CPU 如此处所述https://www.tensorflow.org/performance/performance_guide

['paper (3)', 'Plastic', 'aluminum', 'TIN (2)', ' paper', 'glass', 'Polypropylene Plastic'] 

为什么这里有差异或我的理解不正确？