Model when splitting a giant monolithic accelerator die into chiplets pays off
— then run it: the simulation executes on the ChipFoundryServices compute pool. Defect-limited
yield is super-linear in die area — random particle defects at density D0 kill any
die they land on, so a reticle-sized SoC (the clustered negative-binomial law
Y = (1 + A·D0/alpha)-alpha) can yield below 50%. Disaggregation cuts
the design into N smaller chiplets that each yield far better — but you now pay for a
package that wires them back together, lose parts at assembly, and burn energy shuttling
bits across die boundaries. Chiplets are tested first (known-good-die): coverage t catches
most bad dies, but escapes get built into a package and scrap the whole unit. System yield =
KGD × assembly. Each packaging class — organic MCM, EMIB silicon bridge, CoWoS
interposer, 3D hybrid bond — sets the die-to-die bandwidth density (GB/s per mm of shoreline) and
the energy per bit. The node returns the monolithic-vs-chiplet cost per good part, the
system yield, the beachfront bandwidth and interconnect power, the optimal chiplet count,
and the operating regime. Reduced-order educational model. See also the
speculative-decoding, Mixture-of-Experts,
FlashAttention, quantization,
KV-cache, HBM bandwidth,
systolic array, 3D-parallelism,
power & thermal, transistor I-V,
thermal, die-yield and
lithography simulators and the
compute-pool status.
curl -X POST https://www.chipfoundryservices.com/edge/chiplet \
-H "Content-Type: application/json" \
-d '{"die_area_mm2":800,"num_chiplets":4,"defect_density_d0":0.10,
"clustering_alpha":3,"packaging_type":3,"kgd_coverage":0.95,
"assembly_yield_join":0.998,"d2d_traffic_tbps":2,"wafer_cost_usd":16000}'
Returns JSON with outputs (format, packaging, chiplet_area_mm2, gross_die_mono,
gross_die_chiplet, mono_yield_percent, chiplet_yield_percent, kgd_yield_percent, assembly_yield_percent,
system_yield_percent, mono_cost_per_good_usd, silicon_cost_usd, package_cost_usd,
chiplet_cost_per_good_usd, savings_x, shoreline_mm, avail_d2d_bw_tbps, bw_utilization_percent,
beachfront_limited, d2d_energy_pj_bit, d2d_power_w, link_power_penalty_w, stack_thermal_rise_c,
optimal_n, optimal_savings_x, verdict), the full profile
(16-point savings_vs_n and yield_vs_n sweeps, 48-point
savings_vs_d0 sweep, plus num_chiplets, optimal_n,
savings_x, system_yield_percent, mono_yield_percent,
chiplet_yield_percent, bw_utilization_percent, d2d_power_w),
the serving node, and compute_ms. Endpoint aliases
/edge/chiplets, /edge/packaging, /edge/disaggregation,
/edge/cowos, /edge/interposer, /edge/d2d.