HEXA-WAFER: Wafer-Scale Engine with 20736 Streaming Multiprocessors from n=6

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteur principal: Park, Min Woo
Format: Recurso digital
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901114307215360
author Park, Min Woo
author_facet Park, Min Woo
contents We present HEXA-WAFER, a wafer-scale computing engine that tiles \sigma^2 = 144 HEXA-1 compute tiles on a 300 mm silicon wafer, achieving \sigma^4 = 20{,}736 streaming multiprocessors (SMs), \sigma^3 \cdot J_2 = 41{,}472 GB (\sim41.5 TB) of integrated memory, and \sim72 PFLOPS peak FP8 throughput---all derived from the arithmetic functions of the perfect number n = 6. Each tile replicates the HEXA-1 SoC specification (\sigma^2 = 144 SMs, \sigma \cdot J_2 = 288 GB memory), and tiles communicate via an on-wafer mesh interconnect with \tau = 4 nearest-neighbor links per tile at \sim1 TB/s each. A hierarchical optical overlay provides long-range communication across the wafer. The total power envelope of \sim35 kW is managed through \sigma - \tau = 8 power delivery zones, \tau = 4 cooling mani
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19361500
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle HEXA-WAFER: Wafer-Scale Engine with 20736 Streaming Multiprocessors from n=6
Park, Min Woo
wafer-scale
chip-architecture
n=6
perfect-number
large-scale-computing
We present HEXA-WAFER, a wafer-scale computing engine that tiles \sigma^2 = 144 HEXA-1 compute tiles on a 300 mm silicon wafer, achieving \sigma^4 = 20{,}736 streaming multiprocessors (SMs), \sigma^3 \cdot J_2 = 41{,}472 GB (\sim41.5 TB) of integrated memory, and \sim72 PFLOPS peak FP8 throughput---all derived from the arithmetic functions of the perfect number n = 6. Each tile replicates the HEXA-1 SoC specification (\sigma^2 = 144 SMs, \sigma \cdot J_2 = 288 GB memory), and tiles communicate via an on-wafer mesh interconnect with \tau = 4 nearest-neighbor links per tile at \sim1 TB/s each. A hierarchical optical overlay provides long-range communication across the wafer. The total power envelope of \sim35 kW is managed through \sigma - \tau = 8 power delivery zones, \tau = 4 cooling mani
title HEXA-WAFER: Wafer-Scale Engine with 20736 Streaming Multiprocessors from n=6
topic wafer-scale
chip-architecture
n=6
perfect-number
large-scale-computing
url https://doi.org/10.5281/zenodo.19361500