Flipping Bits in Memory Without Accessing Them: An Experimental Study of DRAM Disturbance Errors
2014 ACM/IEEE 41st International Symposium on Computer Architecture (ISCA), 2014 · paper · Computer science
An experimental study showing that repeatedly reading one row of commodity DRAM can corrupt data in nearby rows. Errors appeared in 110 of 129 modules, sometimes after as few as 139K accesses; the authors identify a cause and propose a low-overhead fix.
Checked: PDF from users.ece.cmu.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Wording Section 6.4 says "Solid (all '0's) has the fewest errors for all three modules by an order of magnitude or more", but Table 6 gives B11 Solid 12,050 against ColStripe 9,610. For A23, Solid (112,123) is only about 6.8× below ColStripe (763,763).
- p. 9 Section 6.4, right column, para 3
Solid (all ‘0’s) has the fewest errors for all three modules by an order of magnitude or more
- p. 10 Table 6, row B11
12,050
- p. 10 Table 6, rows B11 and A23
12,050 320,095 9,610 302,306
- p. 9 Section 6.4, paragraph after Table 6 introduction
Solid (all ‘0’s) has the fewest errors for all three modules by an order of magnitude or more.
Only the authors can say what is right here.
Judge's note: I checked the page 9 text against Table 6 on page 10. For B11, Solid is not the fewest. For A23 and C19, the gap to ColStripe is less than 10×. The claim holds only against RowStripe/Checkered and RowStripe.
Second check: another model agreed
Report a problem with this finding
11 minor (typos and the like; most checked by the judge only)
Reference or label Page 9 reads "Referring back to Figure 2 (Section 2.1)", but Figure 2 is introduced in Section 2.2 ('For example, Figure 2 shows a 2GB rank').
- p. 9 Section 6.3, Reason 1 (second list)
Referring back to Figure 2 (Section 2.1)
- p. 2 Section 2.2
For example, Figure 2 shows a 2GB rank
Judge's note: I checked page 2. Section 2.1 defines a rank in text only, while Section 2.2 introduces Figure 2. The parenthetical may be meant to point to the rank concept, so either reading is defensible. As written, though, it points the figure to the wrong section.
Second check: another model agreed
Report a problem with this finding
Reference or label Section 5 'Tests.' says "two snippets of pseudocode listed above (Code 2)", but Code 2 is printed below that paragraph in the same column.
- p. 4 Section 5, 'Tests.' paragraph
two snippets of pseudocode listed above (Code 2)
Judge's note: I checked page 4. The Tests paragraph is at the top of the right column and Code 2 sits lower in that column. No earlier page has Code 2, so the direction word is wrong. The fix is 'below'.
Second check: another model agreed
Report a problem with this finding
Wording Page 7 says "The family with the most victims at RI = 64ms is also likely to have the lowest RIth: 8.2ms, 9.8ms, and 14.7ms". In Table 3, though, B's most-victim family is B10-12 (max 2.7×10^6, bold), whose RIth is 11.5ms. The 9.8ms belongs to B14 (9.1×10^5 victims).
- p. 7 Section 6.2, Refresh Interval
The family with the most victims at RI = 64ms is also likely to have the lowest RIth: 8.2ms, 9.8ms, and 14.7ms.
- p. 5 Table 3, B10-12 and B14 rows
9.8
- p. 5 Table 3, rows B10-12 and B14
2.7 × 10^6
Only the authors can say what is right here.
Judge's note: I checked the bold entries in Table 3 on page 5. For A (A19-30: 1.0×10^7, 8.2) and C (C31: 3.3×10^5, 14.7) the claim holds. For B, the bold maximum victims (B10-12) and the bold minimum RIth (B14) are different families. The sentence is hedged with 'likely', but the 9.8ms it cites does not come from the family with the most victims.
Second check: another model agreed
Report a problem with this finding
Wording Page 6 says "As Section 6.4 will show, these two data patterns induce the most errors" without qualification, but in Table 6 Checkered (29,283) beats RowStripe (20,770) for C19. Section 6.4 itself calls RowStripe only 'the second most' for C19.
- p. 6 Section 5, top of left column
As Section 6.4 will show, these two data patterns induce the most errors.
- p. 10 Table 6, C19 row
29,283
Judge's note: I checked page 6 and Table 6 on page 10, plus the Section 6.4 text on page 9. RowStripe gives the most errors for A23 and B11 but not for C19. The forward reference states this more strongly than the section it points to.
Second check: another model agreed
Report a problem with this finding
Wording Section 6.3 says "Most rows in A23, B11, and C19 have at least one cell that experienced an error: 100%, 99.96%, and 41.54%". For C19, 41.54% is under half the rows.
- p. 8 Section 6.3, para 1
Most rows in A23, B11, and C19 have at least one cell that experienced an error: 100%, 99.96%, and 41.54%.
- p. 8 Section 6.3, after Table 5
100%, 99.96%, and 47.04%
- p. 8 Section 6.3 first paragraph
100%, 99.96%, and 41.54%
- p. 8
99.96%, and 41.54%
Only the authors can say what is right here.
Judge's note: I checked page 8. The wording 'most rows' does not fit C19's 41.54%. The suggested mismatch with 47.04% is not established as an error. 41.54% is the share of rows holding victim cells; 47.04% (15414/32768) is the share of aggressor rows found with TestEach on the first bank. These are different sets, and the scope of the first figure (whole module or first bank) is not stated. Only the authors can say whether 41.54% is correct.
Second check: another model agreed
Report a problem with this finding
Wording Page 8 says "Most rows in A23, B11, and C19 cause errors when they are repeatedly opened", yet C19 has 15414 of 32768 aggressor rows (47.04%), which is under half.
- p. 8 Section 6.3, paragraph after Table 5
Most rows in A23, B11, and C19 cause errors when they are repeatedly opened.
- p. 8 Section 6.3
100%, 99.96%, and 47.04%
Judge's note: I checked page 8. The paper's own conclusion a few lines later ('large fractions') is more accurate than 'most rows' for C19.
Second check: another model agreed
Report a problem with this finding
Wording The RowStripe counts in Table 6 (1,318,603; 320,095; 20,770) differ from Section 7's ten-iteration first-bank averages for the same test pair ("1.31M, 339K, and 21.0K ... no iterations that deviated by more than ±0.25%") by more than 0.25%.
- p. 10 Table 6, B11 RowStripe
320,095
- p. 10 Section 7, Errors are Mostly Repeatable
1.31M, 339K, and 21.0K. There were no iterations that deviated by more than ±0.25%
- p. 10 Section 7, Errors are Mostly Repeatable
1.31M, 339K, and 21.0K. There were no iterations that deviated by more than ±0.25% from the average
Only the authors can say what is right here.
Judge's note: I checked page 10. Table 6 does not state its bank scope or run conditions, but its magnitudes match first-bank counts. B11 differs by about 5.6% and C19 by about 1.1%. Only the authors can say whether Table 6 comes from a separate run.
Second check: another model agreed
Report a problem with this finding
Wording Section 7 says the weak cells "(984K, 993K, and 1.22M), which is on par with the number of victim cells", but C19 has 24.4K unique victim cells in the first bank, and 1.4×10^5 module-wide in Table 3.
- p. 10 Section 7, Victim Cells ≠ Weak Cells
(984K, 993K, and 1.22M), which is on par with the number of victim cells
- p. 10 Section 7, Errors are Mostly Repeatable
1.48M, 392K, and 24.4K
Only the authors can say what is right here.
Judge's note: I checked page 10 and Table 3 on page 5. For C19, weak cells outnumber victim cells by about 50× (first bank) or about 8.6× (module). 'On par' does not fit C19 under either reading.
Second check: another model agreed
Report a problem with this finding
Typo Section 8.2 reads "PARA does not guarantee that the adjacent will always be refreshed in time", with the noun missing after 'adjacent'. It should read 'adjacent row'.
- p. 11 Section 8.2, Implementation
guarantee that the adja- cent will always be refreshed in time
Judge's note: I confirmed in the crop image that the noun is absent.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [3] reads "DRAM Fault Analaysis and Test Generation". The fix is 'Analysis'.
- p. 12 Reference [3]
DRAM Fault Analaysis and Test Generation
Judge's note: I confirmed the misspelling in the page 12 crop image.
Second check: none (judge only)
Report a problem with this finding
Typo Section 10 (Conclusion), p. 12: "These errors happen when repeated accesses to a DRAM row corrupts data stored in other rows". The plural subject "accesses" takes "corrupt", not "corrupts".
- p. 12 Section 10, Conclusion
repeated accesses to a DRAM row corrupts data stored in other rows
Judge's note: Checked the crop of PDF page 12, Section 10, first paragraph. The sentence reads as quoted. The head of the subject is the plural noun "accesses" ("to a DRAM row" only modifies it), so the verb should be "corrupt". This is a cosmetic grammar error.
Second check: none (judge only)