Floorplan and architecture analysis
Full-die imaging after custom delayering, block identification and area measurement, generation-over-generation and cross-foundry comparison
Die shots of devices that defeat conventional delayering, including the first products with backside power delivery, with every block identified, measured, and compared with its predecessor and with other foundries’ nodes.
What it reveals: Die and block areas, core counts, cache sizes and densities, where the node’s density gain went, and what the silicon costs.
Panther Lake vs Lunar Lake block areas from the floorplan
Intel 18A compute tile against its TSMC N3B predecessor
| Block | Change | Note |
|---|---|---|
| P-core (Cougar Cove) | ~0% | L2 2.5 → 3 MiB in the same area |
| Shared P-core L3 | −14.8% | |
| LP E-core cluster (Darkmont) | −5.0% | 1.5 MiB L2 region −14.9% |
| NPU (NPU 5 vs NPU 4) | −36.9% | Same INT8 MACs, 6 scratchpads vs 12 |
| NCE envelope | +22.6% | 3 engines vs 6, larger MAC arrays |
| GPU L1/SLM | +5% | 192 → 256 KiB, density +27% |
Source: STEEL Panther Lake teardown, September 2026
Kirin 9030 vs Kirin 9020 block areas from the floorplan
SMIC N+3 against N+2, same ~140 mm² die
| Block | Change | Note |
|---|---|---|
| Prime core (TaiShan) | −7.6% | L2 1 → 2 MiB; −21% excluding L2 |
| Middle core | −22% | 3 → 4 cores |
| Big CPU cluster | ~0% | L3 10 → 12 MiB |
| Tiny CPU cluster | slightly larger | Shared L2 2 → 4 MiB |
| GPU compute unit | −28% | Ray tracing added; 4 → 6 CUs |
| GPU cluster | +~10% | Non-CU area +33% |
| NPU | 1 Lite + 2 Tiny | From 1 Lite + 1 Tiny |
Source: STEEL SMIC N+3 teardown, June 2026
Cache arrays measured on the die shot Kirin 9030
Achieved density of SRAM arrays as placed, against the previous generation
| Array | Area | Achieved density |
|---|---|---|
| SLC, 128 KiB | 0.0392 mm² | 25.5 Mib/mm² |
| SLC array vs Kirin 9020 | −18% | 0.0477 → 0.0392 mm² |
| L3, 64 KiB | 0.0210 mm² | 23.8 Mib/mm² |
| Prime-core L2, 32 KiB | 0.0142 mm² | 17.6 Mib/mm² |
| L2 array vs Kirin 9020 | −17% | 0.0171 → 0.0142 mm² |
| SLC capacity | 3 MiB per bank | From 2 MiB, 24 arrays vs 16 |
Source: STEEL SMIC N+3 teardown, June 2026
Xe3 across two foundries GPU tile areas
Panther Lake GT1 (Intel 3) and GT2 (TSMC N3E) against Lunar Lake Xe2
| Metric | GT1, Intel 3 | GT2, TSMC N3E |
|---|---|---|
| Xe cores | 4 | 12 |
| Xe core area vs Lunar Lake | +~69% | GT1 is ~55% larger than GT2 |
| L2 capacity | 4 MiB | 16 MiB |
| L2 macro | 8 KiB | 16 KiB, 54% larger |
| L2 macro density | 18.3 Mbit/mm² | 23.7 Mbit/mm² |
| L2 bank-level density | 10.4 Mbit/mm² | 16.9 Mbit/mm² |
Source: STEEL Panther Lake teardown, September 2026


