M100: Galactic Spiral & AI Dataflow Architecture

• M100 is the designation for both a nearly face‐on, grand-design spiral galaxy in the Virgo cluster with active star-forming regions and Li Auto’s dataflow AI computing architecture.
• Astrophysical studies of M100 use multi-wavelength observations to reveal detailed dust profiles, hierarchical filament structures, and diverse star formation regimes across its disk.
• The computing architecture M100 employs compiler-architecture co-design and dataflow synchronization, achieving significant speedups in AI inference tasks through cache elimination and optimized data streaming.

M100 most commonly denotes Messier 100 (NGC 4321), a nearly face-on, grand-design intermediate spiral—normally classified as S bc—in the Virgo cluster and rich in star forming regions along its spiral arms. Across recent work it has been analysed with Spitzer, Herschel, ALMA, VLA, GALEX, HST, and X-ray data to constrain dust structure, molecular and atomic gas, hierarchical filaments, star-formation ages, halo models, and transient environments. In a distinct engineering usage, “M100” is also the name of Li Auto’s dataflow AI computing architecture (De et al., 2020, Pan et al., 2017, Zhou et al., 2024, Xie et al., 20 Apr 2026).

1. Identification and astrophysical setting

M100 is NGC 4321, a grand-design spiral in the Virgo cluster with numerous H II regions and a two-armed optical morphology. The optical radius is $R_{25} = 3.71' \simeq 17.8$ kpc. Multiband imaging shows that the blue and UV appearance is dominated by clumpy recent star formation along the arms, whereas the near-infrared is smoother: 2MASS $K_s$ imaging shows a smooth bulge of radius $\sim 30''$, hints of a weak oval distortion, and bulge isophotes that are rounder than the optical arms, with $\epsilon \approx 0.10$ in the bulge. Optical imaging also shows dust lanes in the inner arms and a ring-like structure at $r \sim 20''$. X-ray imaging reveals point sources, including SN 1979C, together with diffuse hot gas in the central $\sim 1'$ and a halo extending over $\sim 3'$ ($\sim 15$ kpc) (Marasca et al., 12 Nov 2025, Pohlen et al., 2010).

A notable feature of the M100 literature is the range of adopted fiducial distances. The SN 2019ehk analysis adopts $d_L = 16.2$ Mpc with $\mu \simeq 31.05$ mag and notes a published Cepheid range of $K_s$0–$K_s$1 Mpc; the multiband study adopts $K_s$2 Mpc; the SPIRE dust-continuum mapping uses $K_s$3 Mpc; the PHANGS-ALMA CO(2–1) analysis adopts $K_s$4 Mpc; the ALMA 12CO(1–0) GMA study uses $K_s$5 Mpc; and the DIISC H I analysis uses $K_s$6 Mpc. These differences are methodological rather than taxonomic: all refer to the same Virgo spiral. Metallicity estimates place the galaxy near or slightly above solar in the inner disk, with $K_s$7–$K_s$8, declining to $K_s$9 near the outer spiral arms; on the Pettini & Pagel 2004 scale, one radial fit is

$\sim 30''$0

The galaxy is also described as actively star-forming, with a global SFR of order a few $\sim 30''$1 (De et al., 2020, Eales et al., 2010, Zhou et al., 2024).

2. Dust, far-infrared structure, and ISM mapping

Herschel/SPIRE established that cool, submillimetre-emitting dust in M100 extends to at least the optical radius and follows a broken-exponential radial profile with a clear break at

$\sim 30''$2

The submm colour indices decline monotonically with radius: $\sim 30''$3 falls from $\sim 30''$4 in the centre to $\sim 30''$5 at $\sim 30''$6, while $\sim 30''$7 drops from $\sim 30''$8 to $\sim 30''$9. Under a single-temperature modified blackbody with $\epsilon \approx 0.10$0, these ratios imply a dust temperature decreasing from $\epsilon \approx 0.10$1 K in the centre to $\epsilon \approx 0.10$2 K at $\epsilon \approx 0.10$3. A separate greybody analysis of 70, 250, and 350 $\epsilon \approx 0.10$4m data yields $\epsilon \approx 0.10$5 across the disk. The gas-to-dust ratio rises outward: a proxy based on $\epsilon \approx 0.10$6 increases from $\epsilon \approx 0.10$7 at $\epsilon \approx 0.10$8 to $\epsilon \approx 0.10$9–$r \sim 20''$0 by $r \sim 20''$1, with the corresponding physical trend described as $r \sim 20''$2–$r \sim 20''$3 in the inner disk rising to $r \sim 20''$4 in the outskirts (Pohlen et al., 2010).

The same Herschel program explored dust continuum as an alternative ISM tracer. The hydrogen mass was written as

$r \sim 20''$5

with $r \sim 20''$6 taken at 350 $r \sim 20''$7m, $r \sim 20''$8, $r \sim 20''$9, and an initial $\sim 1'$0. For M100, the $\sim 1'$1-$\sim 1'$2 relation from CO(1–0)+21 cm gives $\sim 1'$3, while dust-continuum mapping gives $\sim 1'$4; the point-by-point scatter is comparable, $\sim 1'$5 dex in both cases. At fixed $\sim 1'$6, however, the dust-inferred $\sim 1'$7 is lower by $\sim 1'$8 dex than the CO+H I estimate. Recalibrating at $\sim 1'$9 gives $\sim 3'$0, suggesting either a lower opacity or a temperature bias in the single-$\sim 3'$1 fits (Eales et al., 2010).

A complementary two-component decomposition of the FIR profile describes M100 as a disk plus a compact core. In the notation of the Herschel analysis,

$\sim 3'$2

with an exponential disk and Gaussian core, or more generally $\sim 3'$3. The core is hotter than the disk and its flux fraction is strongly wavelength dependent: $\sim 3'$4 at 24 $\sim 3'$5m, $\sim 3'$6 at 70 $\sim 3'$7m, $\sim 3'$8 at 160 $\sim 3'$9m, and $\sim 15$0–$\sim 15$1 from 250 to 500 $\sim 15$2m. The core heating is attributed to the modest LINER/Seyfert 2 (“T2”) nucleus and/or a mild nuclear starburst, probably bar-driven (Sauvage et al., 2010).

3. Filaments, “beads on a string,” and hierarchical gas inflow

Spitzer/IRAC revealed a network of long, narrow dust-emission filaments in M100 that is almost entirely invisible in optical images. In the four IRAC bands—3.6, 4.5, 5.8, and 8.0 $\sim 15$3m—these filaments are dotted with compact clumps, producing a “beads on a string” morphology. The most obvious 27 filaments contain 147 marked clumps. Using a scale of $\sim 15$4 per pixel, corresponding to 59 pc at 16.2 Mpc, the histogram of adjacent clump separations peaks strongly at 7 pixels, or $\sim 15$5 pc. A second diagnostic shows a strong peak near zero in the relative separation between successive gaps, indicating nearly equal spacings, and a secondary peak near $\sim 15$6, corresponding to an occasional “missing” clump. The average clump colours,

$\sim 15$7

indicate diffuse gas, PAH emission, and local heating from star formation. Neighboring clumps on the same filament have similar magnitudes, and the dispersion of differences between neighboring 8 $\sim 15$8m clumps is only $\sim 15$9 mag compared with $d_L = 16.2$0 mag for all filament clumps. Equivalent young-stellar masses inferred from total IR luminosities cluster around a few $d_L = 16.2$1 per clump, with a total of $d_L = 16.2$2 in all measured filament clumps (Elmegreen et al., 2018).

The paper interprets these structures through classical filament instability. For an isothermal, self-gravitating cylinder,

$d_L = 16.2$3

and the fastest-growing mode has $d_L = 16.2$4 in the simplest near-critical case, or more precisely $d_L = 16.2$5. In M100 the measured ratio of adjacent clump separation to clump diameter peaks at $d_L = 16.2$6, directly matching the predicted fastest-growing mode. A later PHANGS-ALMA CO(2–1) study extends the same picture kinematically. Using FILFINDER to trace the brightest spiral-arm ridges and a dendrogram decomposition to identify “leaves” and “branches,” Zhou et al. describe nested hub–filament systems from galaxy-cloud to cloud-clump scales. In M100, galaxy-cloud scale hubs correspond to branches with $d_L = 16.2$7–2 kpc and $d_L = 16.2$8–$d_L = 16.2$9, while cloud-clump scale hubs correspond to $\mu \simeq 31.05$0–500 pc and $\mu \simeq 31.05$1–$\mu \simeq 31.05$2. After subtraction of the large-scale rotation field, the local velocity gradient obeys

$\mu \simeq 31.05$3

over $\mu \simeq 31.05$4–3000 pc. Because pure free-fall would imply $\mu \simeq 31.05$5, the observed trend supports gravitational collapse, but a collapse slower than a pure free-fall gravitational collapse (Zhou et al., 2024).

4. Molecular cloud populations, star formation, and dust heating

ALMA 12CO(1–0) feathered observations resolve 165 giant molecular cloud associations in M100. Using CPROPS, these were classified by environment into 11 circumnuclear ring (CNR) GMAs, 21 bar GMAs, 62 spiral-arm GMAs, and 71 inter-arm GMAs. The environmental contrasts are pronounced: the CNR GMAs are massive and compact, with all $\mu \simeq 31.05$6 and $\mu \simeq 31.05$7 up to $\mu \simeq 31.05$8; bar GMAs have elevated velocity dispersion; inter-arm GMAs are diffuse, with low surface density and radii up to $\mu \simeq 31.05$9 pc. The mass–size relation is not universal: $K_s$00 in the CNR, $K_s$01 in the bar, $K_s$02 in the spiral arms, $K_s$03 in the inter-arm region, and $K_s$04 for all GMAs combined. The virial parameter spans $K_s$05–$K_s$06, with median values $K_s$07 in the CNR, $K_s$08 in the bar, $K_s$09 in the spiral arms, and $K_s$10 in the inter-arm region; only the spiral GMAs are in general self-gravitating. Star formation activity decreases in order over the CNR, spiral, bar, and the inter-arm GMAs, and the local Kennicutt–Schmidt slopes differ accordingly: $K_s$11 in the CNR, $K_s$12 in the bar, $K_s$13 in the spiral arms, $K_s$14 in the inter-arm region, and $K_s$15 for all GMAs (Pan et al., 2017).

Age dating from the dust-corrected H$K_s$16/FUV ratio provides a second view of star formation. In M100, the circumnuclear ring at $K_s$17 kpc is dominated by the youngest 0–4 Myr bin, with an azimuthal age gradient of $K_s$18–4 Myr around the ring. Along the two-armed spiral outside the ring ($K_s$19–10 kpc), most pixels fall in the 4–6 Myr bin, with older populations toward the outer edges of the arms. In the short S–SW arm at $K_s$20 kpc, the age changes from $K_s$21 Myr on the inner edge to $K_s$22 Myr on the outer edge over a projected width of $K_s$23 kpc, corresponding to $K_s$24. The interpretation advanced in that work is sequential star formation associated with bar-driven gas inflow near resonances in the central ring and density-wave triggering across the spiral arms (Sánchez-Gil et al., 2011).

Radiative-transfer modelling with SKIRT adds an energy-balance constraint. The M100 model contains an old stellar bulge, an old stellar disc, a young non-ionising stellar disc, a young ionising stellar disc, and a dust disc. For this model, 33% of the bolometric stellar light is absorbed by dust. The effective attenuation curve rises steeply into the UV, shows a clear $K_s$25m bump, has a UV slope roughly $K_s$26, and gives $K_s$27 and $K_s$28. The bolometric dust-heating fraction by young stars,

$K_s$29

has a global value $K_s$30, dropping to $K_s$31 in the bulge region and $K_s$32 in the bar region. Radially, $K_s$33 reaches $K_s$34 at $K_s$35 kpc, dips to $K_s$36–40% over $K_s$37 kpc, and then approaches $K_s$38 in the outer disc. A tight empirical relation is reported between $K_s$39 and local sSFR, with a log–log fit $K_s$40 giving $K_s$41, $K_s$42, and a Spearman $K_s$43 (Nersesian et al., 2020).

5. Transients, anomalous H I, and halo structure

M100 is the host galaxy of the peculiar Ca-rich SN 2019ehk. Discovery images place the SN on a spiral arm at a projected distance of order $K_s$44–$K_s$45, corresponding to $K_s$46–3.0 kpc at 16.2 Mpc. The line of sight has small Galactic foreground reddening, $K_s$47 mag, but substantial host reddening: late analyses bracket $K_s$48 in the range 0.5–1.0 mag. Early flash spectroscopy shows narrow H$K_s$49 and He II emission lines from dense, hydrogen-rich circumstellar material within $K_s$50–$K_s$51 cm. At late times, the inferred [O I] luminosity is $K_s$52, the [Ca II] luminosity is $K_s$53, and the synthesized oxygen mass is $K_s$54–0.069 $K_s$55. These measurements are argued to favour a Type IIb core-collapse supernova from a stripped low mass progenitor of $K_s$56–9.5 $K_s$57, rather than a thermonuclear helium detonation event. In the same galaxy, X-ray mosaics identify SN 1979C as a bright source with $K_s$58–$K_s$59 in the 0.3–2 keV band decades after explosion (De et al., 2020, Marasca et al., 12 Nov 2025).

The outer H I disc contains two kinematically anomalous clouds discovered in the DIISC survey. These clouds lie at projected galactocentric radii $K_s$60 and 15.5 kpc and are offset by $K_s$61 from the rotating disk at their positions. Their measured properties are $K_s$62 and $K_s$63, peak column densities $K_s$64 and $K_s$65, and projected sizes $K_s$66 and $K_s$67 kpc. One of them is directly associated with a compact star-forming region seen in GALEX FUV and H$K_s$68, with $K_s$69 and $K_s$70. The proposed origin is not unique: star-formation feedback-driven outflows, ram-pressure stripping, and tidal interactions with satellite galaxies are all considered plausible. At larger radius, an HST-COS sightline at 38.8 kpc yields a $K_s$71 upper limit $K_s$72, indicating that the inner CGM is predominantly hot or highly ionized (Gim et al., 2021).

Rotation-curve decomposition has also been used to constrain the halo. One study fitted nine dark-matter profiles—Pseudoisothermal, Burkert, NFW, Moore, Einasto, core-modified, DC14, coreNFW and Lucky13—to VLA H I data. Four models, DC14, Lucky13, Burkert and Moore, were rejected as not suitable for this galaxy. The remaining accepted profiles gave reduced $K_s$73 values of 1.25 for Pseudo-Iso, 0.49 for NFW, 0.52 for Einasto, 3.29 for core-modified, and 1.64 for coreNFW. The Pseudoisothermal profile was identified as the best fitting because its inner linear rise and outer flatness match the observed curve, while all successful fits imply a total dark matter mass within 10 kpc of $K_s$74. A cautious implication is that the central $K_s$75 kpc are better described by a cored or partly cored halo than by a steep cusp (Shen et al., 2021).

6. M100 as a computing-architecture designation

In a wholly separate context, M100 is the name of Li Auto’s “orchestrated dataflow” architecture for general AI computing. The system targets AI inference in Autonomous Driving, LLMs, and intelligent human interactions, and is organized around compiler-architecture co-design and explicit management of data movement rather than conventional cache hierarchies. Its top-level structure includes a Central Control Block, described as a 4-core RISC-V + Vector Engine cluster, and 14 clusters with 4 Tensor Processing Blocks each. Each TPB contains 2 MB High-Bandwidth Shared Memory, a Tensor Computing Unit with an $K_s$76 MAC array, a Configurable Vector Unit, Data-Transform DMA, Gather/Scatter DMA, a Scalar/CPU Starter Unit, and a local Synchronization Unit. The architecture largely eliminates caching: tensor computations are driven by compiler- and runtime-managed data streams between computing elements and on/off-chip memories, with producer–consumer counters implementing dataflow synchronization (Xie et al., 20 Apr 2026).

The compiler maps a dataflow graph $K_s$77 into a space–time schedule

$K_s$78

subject to

$K_s$79

Tensor-level granularity is the fundamental design choice. Large tensors are tiled, runtime firmware performs just-in-time assembly of long TPB instructions, and execution cost is modelled as

$K_s$80

Reported application benchmarks compare M100 with Thor-U under an identical power budget. On UniAD, M100 reaches 30 FPS versus 7.9 FPS overall, with module-level speedups of $K_s$81 for RegNet, $K_s$82 for BEVFormer, $K_s$83 for TempFusion, $K_s$84 for TrackFormer, and $K_s$85 for MapFormer. On LLaMA2-7B, decode is near parity at 21.34 ms versus 20.00 ms for W4A16, whereas prefill is 79.00 ms versus 154.00 ms for W8A8. On the MindVLA LLM component, decode is 0.10 ms versus 0.30 ms and prefill is 0.84 ms versus 1.74 ms. The same report attributes to cache elimination an NPU L2+ cache reduction of $K_s$86, a die-area saving of $K_s$87, and an RTL-size reduction of $K_s$88, while stating that under iso-power $K_s$89 active NPU) the design delivers $K_s$90 higher application-level throughput in AD tasks (Xie et al., 20 Apr 2026).