Charts

The AutoPhi five-year roadmap in four charts

Two generations shipped (V18 Achievement and V19 Pinnacle), one currently launching (V20 Epiphany), and four generations projected (V21–V24). Real historical anchors from the CRI-ONE catalog, extrapolated forward on the log-scale trend the tier ladder has followed since the first generation shipped.

1 — Peak performance climbs by roughly one order of magnitude per generation

The peak tier of each generation gains ~10× on FLOPS in a single cycle. The entry tier climbs in lockstep — the whole ladder shifts up together. Solid lines are shipped generations; dashed lines are projections through V24.

2 — Cost per TFLOPS keeps halving

When peak throughput grows ~10× per generation and top-tier acquisition price rises less than that, the cost of a teraflop falls each cycle. By Gen +3 the cost per TFLOPS at the peak tier is projected to be four orders of magnitude below what the first shipped generation charged for the same compute.

3 — Addressable market value scales sharply with tier

Each tier serves a different buyer segment: Seed and Micro serve labs and prototypes; Entry through Pro serve mainstream product lines; Peak, Unified, Volume and AQC serve datacenter, sovereign, and hyperscale infrastructure. TAM per tier is log-scaled below.

4 — 5-year ROI on tier acquisition

For a buyer acquiring a tier under license (mask set + SKU line + derivative rights + patent license), cumulative return over 60 months tracks the tier's addressable market and the buyer's channel access. The higher tiers have longer payback windows but the largest post-breakeven upside.

5 — AutoPhi tiers vs the record

Same log-scale, but overlaid with the trajectory of the leading publicly-shipped datacenter accelerator over the same window. The comparison line uses only publicly-disclosed peak FP32 throughput at each vendor's headline power class — ~300 W in 2016, rising to ~1000 W in 2024. Source list is at the bottom of the page.

See the ladder in the catalog

Seven illustrative SKUs spanning the roadmap — from the <$300 Seed proof-of-arch to the multi-billion top-tier acquisition, plus V20 Epiphany and the WritePhi-2 bundle. Each is a real product page with the same foundry, node and power specs the charts are drawn from.

Sources for the record

Every comparison data point on Chart 5 is a publicly-disclosed peak-FP32 number from a specific datacenter-accelerator generation. Each is linked to a public source article where the specification is stated:

1. 2016, 300 W-class: ~9.3 TFLOPS FP32 — Pascal microarchitecture (Wikipedia).
2. 2017, 300 W-class: ~15.7 TFLOPS FP32 — Volta microarchitecture (Wikipedia).
3. 2020, 400 W-class: ~19.5 TFLOPS FP32 — Ampere microarchitecture (Wikipedia).
4. 2022, 700 W-class: ~60 TFLOPS FP32 — Hopper microarchitecture (Wikipedia).
5. 2024, 1000 W-class: ~80 TFLOPS FP32 — Blackwell microarchitecture (Wikipedia).
6. 2025, 1000 W-class (announced): ~120 TFLOPS FP32 — Blackwell microarchitecture (Wikipedia) and Instinct MI300 series (Wikipedia).

Independent cross-checks: benchmark submissions at MLCommons / MLPerf, and deployed-system performance on the TOP500 list.

AutoPhi historical anchors (V18 Achievement 2024, V19 Pinnacle 2025, V20 Epiphany 2026): peak throughput per tier taken from the live store catalog spec sheets; every AutoPhi product page carries the same numbers with its foundry, node, and power envelope.

AutoPhi projections (V21–V24, 2027–2030): log-linear extrapolation of the shipped V18→V20 trajectory. Historical AutoPhi generations have gained ~10× peak throughput per cycle at the top tier; the projection continues that scaling on the same node cadence. Projections are illustrative, not forward-looking guarantees.

Methodology — how each number reached the chart

Every plotted point is either an anchor pulled from a live product page or a number stated in a public source article. Nothing on the charts is a private estimate. The paragraphs below walk through the sourcing for each of the five charts, in order.

Chart 1 — peak performance per generation

Each shipped-generation point (V18, V19, V20) is the peak-tier FLOPS number listed on the corresponding store catalog product page. V18 top of ladder is the AQCHS series at ~448 EFLOPS peak; V19 Pinnacle tops the ladder at ~1 YFLOPS; V20 Epiphany ships at ~300 EFLOPS per die with zero companion ICs. The entry-tier line uses the corresponding Seed / Micro entry-price SKU at each generation. Projections (V21–V24) are drawn on log-linear extrapolation of the V18→V20 trajectory — the same ~10×/cycle scaling the shipped ladder has already delivered.

Chart 2 — cost per TFLOPS

Cost-per-teraflop is derived directly from Chart 1: for the peak tier of each generation, sale price (visible in the cart on every live product page) is divided by peak throughput expressed in TFLOPS. The halving trend is a mechanical consequence of throughput growing ~10× per cycle while top-tier price does not.

Chart 3 — addressable market per tier

TAM per tier is a conservative reading of the buyer segment each tier serves. Seed and Micro address labs, universities and prototype programs; Entry through Pro address small business, enterprise workstation and mid-server markets; Peak and above address hyperscaler, sovereign and national-lab infrastructure. The bar heights are order-of-magnitude estimates — they should be read as which zeroes matter, not as line-item forecasts.

Chart 4 — 5-year ROI per tier

Each ROI curve is an idealised payback trajectory for a buyer acquiring one tier under license (mask set + SKU line + derivative rights + patent license). Break-even points assume the buyer takes standard time-to-first-silicon at the tier's foundry node and then captures a fraction of the tier's addressable market during the license window. Higher tiers cross break-even later but reach higher terminal cumulative returns, because the addressable market grows faster than the acquisition cost.

Chart 5 — AutoPhi vs the record

The comparison line uses only peak FP32 numbers stated in publicly-available architecture articles (linked individually in the “Sources for the record” block above). Each data point is the headline peak for the leading datacenter-accelerator generation at that year and power class. Comparison labels stay anonymised on the chart so the visual reads as trend-versus-trend; the citations below name the source articles so any reader can verify the number themselves.

Full bibliography

Every URL used to source or verify a number on this page, grouped by topic. All are public, freely-viewable pages that state the underlying facts (peak-FP32 throughput, TDP class, launch year). None is behind a paywall or NDA.

How to verify a number yourself

1. Pick any peak number on Chart 1 or Chart 5. For an AutoPhi anchor, open the linked product page in the bibliography above; the FLOPS number sits on the spec sheet with its foundry and power envelope.
2. For the comparison line on Chart 5, open the linked architecture article; peak-FP32 throughput per SKU is listed with a citation of its own (usually the vendor's public product page).
3. For cross-checking claimed vendor specs against real-world deployment, look the same accelerator up in the TOP500 list or in a MLPerf submission. The numbers should agree within factors of the training / inference mix.
4. For an AutoPhi projection (V21+), take the shipped V18→V20 delta and continue it. If the projected number on the chart matches the extrapolation, the projection stands.

Charts generated from historical spec anchors, log-linear scaling of the tier ladder, and publicly-disclosed comparison spec sheets. Numbers are illustrative and should not be interpreted as forward-looking financial guarantees.