Paper check About

Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks

IEEE Journal of Solid-State Circuits, 2017 · paper · Computer science

Presents Eyeriss, a reconfigurable accelerator for deep convolutional neural networks that reduces data movement using a "row stationary" dataflow on 168 processing elements. It processes AlexNet convolutional layers at 35 frames/s and 278 mW.

Checked: PDF from courses.cs.washington.edu, fetched 5 Oct 2026. Page numbers are those of that PDF.

1 finding

Substance Sec. IV-A says "In AlexNet, the PE sets are of size 11 × 55 (CONV1), 45 × 27 (CONV2)". The CONV2 PE set is 5 × 27 (R = 5).
  • p. 4 Sec. IV-A, 2-D Convolution PE Set paragraph
    45 × 27 (CONV2)
  • p. 4 Sec. IV-A, PE Set Mapping paragraph
    The 5 × 27 PE set of CONV2
  • p. 4 PE Set Mapping example
    The 5 × 27 PE set of
  • p. 4
    45×27 (CONV2), and 3×13

Judge's note: The page 4 image prints '45 × 27 (CONV2)'. The same page defines the PE set height as R. Later in that section the text says 'The 5 × 27 PE set of CONV2 is divided into ... 5 × 14 and 5 × 13', and Fig. 5 shows the CONV2 segments as 5 rows tall. A height of 45 would also exceed the 12-row array limit stated on the same page. The number should read 5.

Second check: another model agreed

Report a problem with this finding
2 minor (typos and the like; most checked by the judge only)
Wording Eq. (1) bounds the output indices as "0 ≤ y < E, 0 ≤ x < F", but x pairs with the filter-row index i (range R, height) in I[z][k][Ux+i][Uy+j], so x should range over E and y over F.
  • p. 2 Eq. (1)
    0 ≤ y < E, 0 ≤ x < F
  • p. 2 Table I
    ofmap plane height/width

Judge's note: Checked Eq. (1) on the page 2 image. O[z][u][x][y] = ReLU(B[u] + Σ_k Σ_{i<R} Σ_{j<S} I[z][k][Ux+i][Uy+j] × W[u][k][i][j]). Table I defines R/S as filter height/width, H/W as ifmap height/width and E/F as ofmap height/width, and E = (H−R+U)/U. The largest row index is U·x_max + R − 1, which equals H − 1 only when x < E. The printed ranges are therefore swapped. This makes no difference for square layers (E = F), which covers every layer evaluated in the paper, so it is rated minor.

Second check: another model agreed

Report a problem with this finding
Typo Sec. III-B: "Each PE can start its own processing as soon as any fmaps or psums arrives (fmaps or psums)." The parenthetical repeats the preceding words, and the singular "arrives" does not agree with the plural subject.
  • p. 3 Sec. III-B, end of first paragraph
    soon as any fmaps or psums arrives (fmaps or psums).
  • p. 3 Sec. III-B, last sentences of first paragraph
    as soon as any fmaps or psums arrives (fmaps or psums)
  • p. 3 Section III-B
    soon as any fmaps or psums arrives (fmaps or psums)
  • p. 3
    as soon as any fmaps or psums arrives (fmaps or psums).

Judge's note: The crop of page 3 shows the duplicated parenthetical in the typeset text, so this is not an extraction artefact. It is a copy-editing slip.

Second check: none (judge only)

Report a problem with this finding
Something wrong with this page, or a finding we missed?