A Case for Intelligent RAM
1997 · paper · Computer science
Advocates Intelligent RAM, which merges processor and memory on a single chip to reduce memory latency, raise bandwidth and improve energy efficiency. It reviews current processors and DRAMs, discusses opportunities and challenges, and estimates performance and energy of three IRAM designs.
Checked: PDF from mast.stanford.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label Table 1 and the introduction describe the measured system as running "on an AlphaServer 8200 computer", but p.5 says "Recall that the memory latency on the AlphaServer 8400 is 253 ns" and gives "the sustained memory bandwidth of the AlphaServer 8400 ... is 1.2 Gbytes/s". These are the Table 1 values (253.0 ns, 1,200 Mbytes/s).
- p. 1 Introduction, last column
processor running on an AlphaServer 8200
- p. 2 Table 1 caption
Alpha 21164 processor running on an AlphaServer 8200 computer.
- p. 5 Higher bandwidth / Lower latency paragraphs
Recall that the memory latency on the AlphaServer 8400 is 253 ns.
- p. 1 Why there is a problem, text and Table 1 caption
a 300-MHz Alpha 21164 processor running on an AlphaServer 8200 computer
Only the authors can say what is right here.
Judge's note: I checked the Table 1 caption on p.2 (AlphaServer 8200, main memory 253.0 ns and 1,200 Mbytes/s) and the p.5 text (AlphaServer 8400, 253 ns introduced with "Recall", and 1.2 Gbytes/s). The word "Recall" points back to the Table 1 measurement, yet the machine is named differently. The paper never says the two models share a memory system. Only the authors can say which model name is correct.
Second check: another model agreed
Report a problem with this finding
Reference or label The Figure 6 x-axis tick labels read "0.10 0.100 10.00 100.00". The second decade tick is labelled 0.100, which equals the first tick; it should be 1.00.
- p. 9 Figure 6 x-axis
0.10 0.100 10.00 100.00
Judge's note: I viewed the figure on p.9. The gridlines are evenly spaced on a log axis. The labels 10.00 and 100.00 fix the decade spacing, so the tick between 0.10 and 10.00 must be 1.00.
Second check: another model agreed
Report a problem with this finding
9 minor (typos and the like; most checked by the judge only)
Substance "typically 8 to 32 'vector registers,' each with about 32 to 128 sixty-four-bit elements. Thus, they have 32 to 256 Kbits". The stated extremes give a lower bound of 16 Kbits, not 32 Kbits.
- p. 7 IRAM vector processor section
Thus, they have 32 to 256 Kbits of multiported, high-
- p. 7 Vector processor section, register sets paragraph
Thus, they have 32 to 256 Kbits of multiported, high-speed registers.
Only the authors can say what is right here.
Judge's note: I recomputed from the stated ranges. The minimum is 8 x 32 x 64 = 16,384 bits = 16 Kbits. The maximum is 32 x 128 x 64 = 262,144 bits = 256 Kbits, which matches the upper bound. The lower bound does not follow from the same extremes. The words "typically" and "about" soften the claim, so I rate it minor.
Second check: another model agreed
Report a problem with this finding
Wording The text says "halving the retention rate for every increase of 10 degrees" under the bullet heading 'Retention time'; retention time (not rate) halves, which is why refresh rates rise.
- p. 5 Potential disadvantages, retention bullet
ing the retention rate for every increase of 10 degrees
- p. 5 Potential disadvantages, retention bullet
gave a rule of thumb of halving the retention rate for every increase of 10 degrees centigrade
- p. 5 same bullet
refresh rates could rise dramatically
Judge's note: The crops show the wording. 'Rate' contradicts both the bullet heading and the conclusion that refresh rates rise.
Second check: another model agreed
Report a problem with this finding
Wording "Depending on the benchmark, the IRAM advantage is roughly a factor of 2 to 4", repeated in the conclusion, but Table 6 ratios run from 1.0 (perl, 16:1) to 4.5 (compress, 32:1), and five of the ten entries are below 2.
- p. 8 IRAM's energy efficiency paragraph
the IRAM advantage is roughly a factor of 2 to 4.
- p. 8 Table 6
IRAM, 16:1 memory 1.0 1.7 1.8 2.3 1.7
- p. 9 Conclusion
an advantage in energy efficiency of a factor of 2 to 4
- p. 8 IRAM's energy efficiency, text after Table 6
the IRAM advantage is roughly a factor of 2 to 4
Only the authors can say what is right here.
Judge's note: Table 6 on p.8 reads: 16:1 row 1.0, 1.7, 1.8, 2.3, 1.7; 32:1 row 1.5, 2.0, 2.5, 2.5, 4.5. The stated range of 2 to 4 leaves out the low 16:1 values, including 1.0, which means no advantage, and the 4.5 value. "Roughly" softens the claim, so I rate it a wording issue.
Second check: another model agreed
Report a problem with this finding
Typo A word is missing: "10% cheaper than 16-bit-wide part" should read "than a 16-bit-wide part".
- p. 3 DRAM width discussion
10% cheaper than 16-bit-wide part
Judge's note: The crop confirms the article is absent.
Second check: none (judge only)
Report a problem with this finding
Typo Wrong indefinite article on p.7: "a IRAM vector processor could have the same peak performance" and "a eight-pipe vector processor" should both read "an".
- p. 7 Vector processor section
a IRAM vector processor could have the same peak perfor-
- p. 7 Vector processor section
Rather than operating on one element at a time, a eight-pipe
- p. 7 An IRAM vector processor, paragraph on multiple pipelines
a IRAM vector processor could have the same peak performance
- p. 7 Vector processor section
a eight-pipe vector processor
Judge's note: Both phrases appear on the p.7 page image. Elsewhere the paper writes "an IRAM".
Second check: none (judge only)
Report a problem with this finding
Typo The Table 6 benchmark is printed as "hyfsys", which is likely a misspelling of hsfsys.
- p. 8 Table 6 header
hyfsys
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the benchmark in Fromm et al. is the NIST hsfsys. The image does show 'hyfsys'.
Second check: none (judge only)
Report a problem with this finding
Typo The chip is named "Mitsubishi M 32 R/D" in the text but "Mitsubishi 32 R/D" in the Figure 6 label.
- p. 8 How they stack up
Mitsubishi M 32 R/D
- p. 9 Figure 6 label
Mitsubishi 32 R/D
Judge's note: Both images confirm the inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 15 lists the author as "A. Nowatzk", and reference 18's title reads "A 7.68 GIPS 3,84 GB/s 1W", mixing a decimal comma with a decimal point.
- p. 10 Reference 15
A. Saulsbury, F. Pong, and A. Nowatzk
- p. 10 References, item 15
A. Nowatzk
- p. 10 Reference 18
A 7.68 GIPS 3,84 GB/s 1W Parallel Image
- p. 10 Reference 18
A 7.68 GIPS 3,84 GB/s 1W
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the co-author's surname is Nowatzyk. Both strings appear on the p.10 page image. The mixed decimal separators in ref 18 can be seen within the paper itself; the surname spelling needs outside confirmation.
Second check: none (judge only)
Report a problem with this finding
Typo The title of ref. 18 mixes decimal marks: "A 7.68 GIPS 3,84 GB/s".
- p. 10 Reference 18
A 7.68 GIPS 3,84 GB/s 1W Parallel Image
- p. 10 References, item 18
A 7.68 GIPS 3,84 GB/s 1W
Judge's note: The image confirms the inconsistency. The comma may reproduce the original title as published, so either form could be legitimate.
Second check: none (judge only)