- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes EAPlace-VLM for macro placement. It builds on a WireMask-style evolutionary search where a replace-set of macros is removed and greedily reinserted. Most iterations choose the replace-set with a scalar heuristic (HPWL, RUDY, overlap). Every tau iterations, a VLM (GPT-4o or Qwen) looks at a rendered placement, in which each macro is coloured by its highest-fanout net, and returns up to K macro IDs. After the search, a local search in continuous coordinates removes the whitespace created by ceiling-based grid quantization. On ISPD2005, the full method has the best average rank in macro-only HPWL (Table 1) and, followed by DREAMPlace, in mixed-size HPWL (Table 2). When EGPlace gets the same refiner, the average gain drops to about 3.1%, with wins on 6 of 8 benchmarks (Table 4).
Critical strengths
Strength 1: The VLM only picks macro IDs, and the deterministic mask-guided insertion handles coordinates and legality. This is a sensible way to use a VLM in a placement loop, since a wrong answer can only waste one iteration and never produces an illegal layout. The invalid-output fallback also keeps the search robust.
Strength 2: The continuous refiner is simple and cheap (3 to 24 minutes against a 1000-minute search, Table 9). It helps both search methods: an 8.7% average HPWL reduction for EGPlace and 7.4% for EAPlace-VLM (Table 4). I also appreciate that the paper reports the matched-refiner comparison and the congestion increase openly instead of hiding them.
Critical weaknesses
Weakness 1: The central claim, that VLM diagnosis improves the search, is not isolated. There is no run of the same framework with the VLM branch replaced by the heuristic, or by a random K-subset on the same schedule. The comparison against EGPlace does not settle this, because the paper does not state whether its heuristic branch, parent selection and population settings are identical to EGPlace. GPT-4o and Qwen give almost the same numbers, and the Appendix A rationale describes a "same-color group" that does not match the image. So a fair reading is that any periodic perturbation might give a similar effect. In addition, tau is never reported, so the reader cannot tell how many VLM calls a run actually makes.
Weakness 2: Most of the gain comes from the refiner, not the VLM. Table 1 reports up to 25.1% against EGPlace, but against EGPlace with the same refiner the mean gain is about 3.1%. That mean is driven by bigblue3 (14.2%) and adaptec2 (9.5%), and EAPlace-VLM loses on adaptec3 and bigblue2. On adaptec1, adaptec4 and bigblue4, the differences are smaller than one standard deviation over 5 seeds, and no significance test is given. The refiner also raises peak RUDY by up to 32% (Table 6), and the mixed-size gains are not uniform (Table 2).
Questions for authors
Clarification requests:
- What value of tau was used, how many VLM calls happen in a 1000-minute run, and is VLM latency counted inside the 1000-minute budget?
- Table 10 reports a constant 337 prompt tokens for every benchmark. The prompt is said to include a macro map with 543 to 1,329 entries plus an image. How does that fit in 337 tokens? Is the full macro map actually sent?
- Eq. 17 computes HPWL from macro lower-left coordinates and does not mention pin offsets or I/O ports. Is this the metric behind Tables 1 to 4? If so, are the copied baseline numbers from BBOPlace-Bench computed the same way? Were WireMask-EA and EGPlace rerun with the same 1000-minute budget and hardware?
- Several Table 6 values differ from Tables 1 and 4. For example, EGPlace adaptec2 is 37.80 vs 39.92, and EAPlace-VLM bigblue3 is 36.02 vs 39.74. Which runs does Table 6 use?
- In Table 2, MaskPlace + DP has the identical entry 26.14 ± 0.07 on both adaptec3 and adaptec4. Is this a copy error?
- K = 6 and N = 5 were chosen on the same benchmarks used for the final comparison (Tables 7 and 8). K = 6 is clearly better than K = 5 and K = 7 on adaptec3 and bigblue3, which looks more like seed noise than a smooth trend. Can you comment?
- Which Qwen model and version is used by default? Also, the Figure 3 caption refers to "Appendix ??".
Suggestions for future work:
A random-subset control and a text-only control (macro map without the image) would make the VLM claim much more convincing. Results on a benchmark with timing or routing evaluation would also strengthen the practical case, since ISPD2005 HPWL alone says little about downstream quality.