Enflame S60 (SCORPIO-AO) AI accelerator

THE HIERARCHY POINT: this is the thinnest record in the slice and it is thin for a defensible reason: Enflame is one of the least transparent accelerator vendors in existence, its own website is behind a WAF that blocks automated access entirely, and the ONE authoritative independent source - a TechInsights teardown (cn-enflame-techinsights) - establishes the part's existence and process node but paywalls everything quantitative. WHAT IS ESTABLISHED: the S60 is Enflame's THIRD-generation AI accelerator card; its compute chiplet carries 'ENFLAME SCORPIO-AO' die markings; it was manufactured by TSMC on N6NTO-HPC. That process_nm: 6 field is the only hard number in this record and it comes from an independent teardown rather than a press release, which makes it the most trustworthy Enflame datum available. WHAT IS NOT ESTABLISHED: memory type, capacity, bandwidth, compute throughput at any precision, TDP, and form factor are all None because no source states them. TechInsights' analysis does include die utilisation broken out for logic, I/O memory and analog, and manufacturing cost, but that content sits behind a paid platform subscription and could not be read. THE OLDER ENFLAME PARTS ARE WORSE AND ARE DELIBERATELY NOT RECORDED HERE: T20 (SuiSi 1.5, PCIe, ~300 W, 2021), T21 (SuiSi 2.0, OAM, 400 W, 2022, claimed 160 TFLOPS TF32 and a cluster of 8000+ cards at 1.3 EFLOPS peak single-precision), i10/i20 inference line, and S60 itself. An unofficial Chinese AI-accelerator wiki (cn-cn-ai-accelerator-wiki) does publish T21 numbers and - to its credit - explicitly labels its FP16, INT8, memory capacity and memory bandwidth entries as unpublished-and-speculated. Those figures are therefore NOT recorded as an accelerator record: recording a part whose every number the source itself marks as a guess would violate this repo's core rule, and a record with four None fields and a note saying 'not recorded because the only source marks its own figures as speculation' would be documentation of nothing. VENDOR CONTEXT worth keeping: Enflame was founded March 2018 by two ex-AMD engineers, is Tencent-backed, licensed Arteris FlexNoC in 2018 for its on-chip backbone, sells predominantly to state-owned enterprises and has cooperated with local governments on AI computing hubs (Wuxi), was reported in June 2024 as submitting downgraded designs to TSMC to comply with US restrictions, and IPO'd on the SSE STAR Market in September 2026 (ticker 688801). The S60's third-generation status plus a TSMC N6 node plus a 2026 IPO is genuinely notable given the sanctions constraints - it implies Enflame secured N6-class foundry access, which is a materially harder procurement problem than the 7 nm N+1/N+2 parts the 910 line uses. MEMORY_BUS_WIDTH: STILL NULL, AND HERE IS PRECISELY WHAT WAS LOOKED FOR (2026-10-04). The vendor publishes NO memory-bus width for this part, and none was back-computed from the published bandwidth. NO FIGURE WAS BACK-COMPUTED FROM BANDWIDTH.[REPAIRED 2026-10-08 by w8-reverify. This sentence previously read '(2026-10-04). %s NO FIGURE WAS BACK-COMPUTED FROM BANDWIDTH' — a literal printf placeholder left in the JSON by the wave-3 pass (commit 9968769) across 34 accelerator records, and invisible to every check in this repo because validate.py never reads notes. The substitution text is the same boilerplate the 7 uncorrupted records in this family already carry, and it is only valid here because memory_bus_bit is null on every one of the 34, so "the vendor does not publish it" is the true state. A record with a non-null memory_bus_bit must NOT receive this filler.] That is the standing rule for this field in this repo and it was applied to every null here rather than to a subset: a width derived by dividing a published aggregate bandwidth by an assumed or half-assumed per-pin rate is two inferences stacked on one number, and if either inference is wrong the record looks populated while being wrong. The correct state of this field is 'the vendor does not publish it', and the note above names the specific document that was read and found not to carry it, so a future pass does not have to re-establish that from scratch. MEMORY_BUS_WIDTH: STILL NULL, AND HERE IS PRECISELY WHAT WAS LOOKED FOR (2026-10-04). No source located publishes a memory bus width for the S60. Enflame's own site is behind a WAF that blocks automated access, so even the vendor page could not be read. NO FIGURE WAS BACK-COMPUTED FROM BANDWIDTH. That is the standing rule for this field in this repo and it was applied to every null here rather than to a subset: a width derived by dividing a published aggregate bandwidth by an assumed or half-assumed per-pin rate is two inferences stacked on one number, and if either is wrong the record looks populated while being wrong. The correct state of this field is 'the vendor does not publish it', and the text above names the specific document read and found not to carry it, so a future pass does not have to re-establish that from scratch. flops IS EMPTY AND THIS PASS ESTABLISHED WHY, WHICH IS A STRONGER STATEMENT THAN 'NONE FOUND'. Enflame's own first-party S60 material is unreadable from here by FOUR independent client types, all returning HTTP 403: the scrape service ('document_antibot'), curl with a full desktop User-Agent (403, zero bytes), a REAL headless-browser session which got the WAF's own interstitial ('您的请求已中断 / 您提交的请求可能对网站造成威胁，无法访问 / Web应用防护服务检测您当前访问存在Web安全风险或访问不合规', with a request UUID), and - decisively for the fallback route - the Wayback Machine, whose availability API returns an EMPTY archived_snapshots object for support.enflame-tech.com/onlinedoc_hw/5-s6x/S60/index.html, for www.enflame-tech.com/product-technologies/s60 and for enflame-tech.com/en/product/s60.html, with CDX index queries returning zero rows. So there is no archived copy of the vendor pages either, which closes the mirror route. Note the contrast with MetaX and Cambricon: there the vendor is reachable but silent; here the vendor may well publish a figure and no reader in this environment could tell, which is a different and less comfortable fact. THE THIRD-PARTY FIGURES THAT CIRCULATE are consistent with each other on what they share - 48 GB, 672 GB/s, PCIe Gen5 x16, 300 W maximum, FP32/FP16/BF16/INT8 supported - and a secondary compilation states it examined Enflame's product manual and found it does NOT publish peak per-precision compute, correcting its own earlier 'FP16 100 TFLOPS (estimated)' as an estimate rather than an Enflame figure. That is consistent with the empty array but remains second-hand: the manual itself is behind the WAF. DO NOT populate flops from an estimate. The TechInsights floorplan-analysis source this record already cites is the one first-party-adjacent document that IS readable, and it gives floorplan and capacity facts rather than a vendor compute claim. See w5a-enflame-s60-vendor-site-unreachable-waf-confirmed.