The Case for a Single-Chip Multiprocessor
1996 · paper · Computer science
Argues that a multiprocessor of simpler processors on one chip is a better use of silicon than a wide-issue superscalar processor. In a same-area comparison, the superscalar is up to 30% better on sequential code, while the multiprocessor is 50–100% better on highly parallel workloads.
Checked: PDF from ece.ucdavis.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
3 findings
Substance Section 4.1: "the actual execution units only occupy about 70 mm2 — just 18% of the die"; Table 2 gives integer 31 mm2 (7%) + FP 37 mm2 (9%) = 68 mm2, 16% of 430 mm2, and 70/430 is also about 16%.
- p. 4 Section 4.1, first paragraph
only occupy about 70 mm2 — just 18%
- p. 6 Table 2, Integer/FP Functional Units rows
31
- p. 4 Sec 4.1 para 2
70 mm 2 --just 18% of the die
- p. 6 Table 2
Integer Functional Units 31 7%; FP Functional Units 37 9%
Judge's note: I checked the p5 text and Table 2 on p7. The die total is 430 mm2. Neither 68 nor 70 mm2 is 18% of 430, and the table's own percentages sum to 16%. Nothing in the paper redefines 'execution units' to include other rows.
Second check: another model agreed
Report a problem with this finding
Reference or label Section 2 says of the banked cache's added access time "We investigate this issue in more detail in Section 4.2." The banked-cache analysis is in Section 4.1.
- p. 3 Section 2, end of execution-phase paragraph
We investigate this issue in more detail in Section 4.2.
- p. 5 Section 4.1, cache paragraph
banked eight ways into eight small, single-ported, independent 4 KB cache banks
Judge's note: On p4 the pointer reads 'Section 4.2'. On p6, Section 4.1 holds the 8-bank L1 analysis: bank control and crossbar add a cycle, area grows 25%, and hit time is 2 cycles. Section 4.2 covers the MP's single-banked caches and only notes that no arbitration overhead is incurred.
Second check: another model agreed
Report a problem with this finding
Reference or label Ref [7] dates the 22nd ISCA to 1996 ("Proceedings of the 22nd Annual International Symposium on Computer Architecture, pp. 333–344, Santa Margherita Ligure, Italy, June, 1996."), but ref [23] dates the same conference to 1995 and ref [16] dates the 23rd ISCA to 1996.
- p. 10 Reference [7]
Margherita Ligure, Italy, June, 1996.
- p. 10 Reference [23]
22nd Annual Int. Symp. Computer Architecture, Santa Margherita, Italy, June 1995.
- p. 10 Reference [7]
Proceedings of the 22nd Annual International Symposium on Computer Architecture, pp. 333–344, Santa Margherita Ligure, Italy, June, 1996.
Judge's note: Checked p.10 image and crops. [7]: 22nd ISCA, Santa Margherita Ligure, June 1996. [23]: 22nd Annual Int. Symp. Computer Architecture, Santa Margherita, Italy, June 1995. [16]: 23rd Int. Symp. Computer Architecture, Philadelphia, 1996. The paper's own references show the inconsistency, so 1995 is the evident fix for [7].
Second check: another model agreed
Report a problem with this finding
8 minor (typos and the like; most checked by the judge only)
Substance In Table 2 the Reorder Buffer row reads "17 | 9 | 300% | 34 | 9%". 34 of 430 mm2 is 7.9%, the % Area column then sums to 101%, and 9 mm2 grown by 300% gives 36 rather than 34.
- p. 6 Table 2, Reorder Buffer row
34
- p. 6 Table 2, Reorder Buffer row, % Area
9%
- p. 6 Table 2, Reorder Buffer row
Reorder Buffer 17 9 300% 34 9%
Only the authors can say what is right here.
Judge's note: I read Table 2 on page image p7. The other rows follow size/430 to within rounding: 38 gives 9%, 31 gives 7%, 37 gives 9%, 50 gives 12%. The reorder buffer row does not. With 8% the column would sum to 100. The unrounded extrapolation 17*(0.25/0.35)^2*4 is about 34.7. Only the authors can say whether the size or the percentage is the intended value.
Second check: another model agreed
Report a problem with this finding
Wording "the actual IPC increases by less than a factor of 1.6 for all of the integer and multiprogramming applications", but m88ksim's tabulated IPC goes from 1.4 (Table 5) to 2.3 (Table 6).
- p. 8 Section 6, first paragraph
increases by less than a factor of 1.6 for all of the integer and multi- programming applications
- p. 8 Table 5 / Table 6, m88ksim IPC
2.3
- p. 8 Tables 5 and 6, m88ksim
m88ksim 1.4 / m88ksim 2.3
- p. 8 Section 6 text vs Tables 5 and 6
increases by less than a factor of 1.6 for all of the integer and multi-
Only the authors can say what is right here.
Judge's note: I checked Tables 5 and 6 on p9. The other integer applications are below 1.6: compress 1.33, eqntott 1.38, MPsim 1.5, pmake 1.4. m88ksim's printed IPCs give 1.64. In Fig. 6 the m88ksim SS bar sits at about 1.6. Because the IPCs are rounded to one decimal, the true ratio could be just under 1.6, so this is marginal.
Second check: another model agreed
Report a problem with this finding
Typo Ungrammatical: "it will be very difficult for this parallelism could be extracted with a superscalar execution model" (should be "to be extracted").
- p. 3 Section 3, paragraph on fine-grained thread-level integer applications
it will be very difficult for this parallelism could be extracted with a superscalar execution model
- p. 3 Sec 3, para on manually parallelized applications
it will be very difficult for this parallelism could be extracted
Judge's note: The page image (p4, right column) shows this wording exactly.
Second check: none (judge only)
Report a problem with this finding
Typo Benchmark name misspelled "Mk88ksim, MPsim and pmake"; it is m88ksim everywhere else.
- p. 8 Section 6, Figure 4 discussion
Mk88ksim, MPsim and pmake
- p. 8 Section 6 first paragraph
increases by less than a factor of 1.6 for all of the integer and multi-
- p. 8 Tables 5 and 6, m88ksim
m88ksim
Judge's note: The crop of p9 shows 'Mk88ksim'. Reviewer note 6 in this cluster, on the 1.6 factor, belongs with C011.
Second check: none (judge only)
Report a problem with this finding
Typo Doubled full stop: "2.4 for swim.."
- p. 8 Section 6, first paragraph
2.4 for swim..
Judge's note: Both periods are visible in the crop of p9.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [1] gives the venue as "Seventh SIAM Conference on Parallel Processing for Scientific Compiler"; the word 'Compiler' appears to be a slip.
- p. 10 Reference [1]
Conference on Parallel Processing for Scientific Compiler
- p. 10 Reference [1]
Parallel Processing for Scientific Compiler
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the SIAM conference is titled 'Parallel Processing for Scientific Computing'. The page image (crop C017-a) clearly prints 'Scientific Compiler', so this is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [9] spells the author's name "Don Drappper," with a tripled 'p'.
- p. 10 Reference [9]
Don Drappper
- p. 10 Reference [9]
Don Drappper,
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the author of the ISSCC 1996 'The interconnect nightmare' is Don Draper. The page image prints 'Drappper', and a tripled consonant is very unlikely to be the intended spelling.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [14] lists the co-author as "S. Naffzinger"; the name appears to be Naffziger.
- p. 10 Reference [14]
S. Naffzinger
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the PA-8000 ISSCC 1996 co-author is Samuel Naffziger. The page image shows 'Naffzinger'.
Second check: none (judge only)