Iluvatar BI-V150 / BI-V250 fabric - VERIFIED NEGATIVE: the vendor's own BI-V150 page loads and has no specification section at all, so there is no scale-up bandwidth to record

NULL, AND THE NULL IS THE ENTIRE FINDING - WHICH IS WHY THIS RECORD EXISTS INSTEAD OF A BI-V150 RECORD WITH A NUMBER IN IT. The vendor's own product page for the BI-V150, https://www.iluvatar.com/productDetails?fullCode=cpjs-yj-xlxl-tg150, was fetched on 2026-10-05 and RETURNED HTTP 200 WITH 3,970 CHARACTERS OF CONTENT. The page is not blocked, not paywalled, not an error page and not empty of text. It simply contains NO hardware specification section at all. SUBSTRING COUNTS OVER THE ENTIRE RETRIEVED BODY, which are the evidence rather than a claim of absence: 产品规格 / 规格 (specification-table heading) 0 occurrences, 接口 (interfaces) 0, PCIe 0, 'GB/s' 0, 'TB/s' 0, 互联 (interconnect) 0, and even the part number itself ('BI-V150' / 'V150' / 'BI') 0 occurrences. Its whole substantive body is 产品特点 bullets about SOFTWARE - '支持包括DeepSpeed、Megatron-DeepSpeed、Megatron-LM、Colossal-AI等主流框架，提供自研的推理框架无缝兼容vLLM和TGI热门框架', '生态友好，支持异构方案混训', '快速支持国内外新兴大模型，包括LLaMa，Qwen，Yi等最新版本和最大参数量'. Page assets are dated 20251021. NO SPECIFICATION TABLE EXISTS ON THIS PAGE. NOTHING WAS BACK-COMPUTED, ESTIMATED OR CARRIED OVER: the previous generation's 64 GB/s (interconnect/iluvatar-bi-v100-chip-to-chip) is deliberately NOT reused here, and the third-party aggregator's estimate is NOT substituted for the missing figure. Silently inheriting a number from the prior part is exactly the failure mode this repo has been burned by, and a bandwidth that is null because it was inherited is worse than one that is null because nobody published it.

NONE PUBLISHED. No topology, no port count, no link, no direction, no switch. There is nothing to describe, and this field is null rather than an empty string precisely so that 'the vendor told us there is no topology' stays distinguishable from 'nobody wrote the topology down'. The 互联 (interconnect) term does not occur once anywhere on the page.

NOT PUBLISHED, AND THE SAME NON-CLAIM APPLIES: this does not establish that the BI-V150 has no scale-up fabric, only that Iluvatar publishes none for it. A part marketed for the largest parameter counts of Llama and Qwen plainly connects to something; the vendor simply does not say what.

A VERIFIED NEGATIVE WITH THE URL NAMED, WHICH IS THE POINT. Tested URL: https://www.iluvatar.com/productDetails?fullCode=cpjs-yj-xlxl-tg150 - HTTP 200, readable, 3,970 characters, zero specification content. That distinction is the whole reason this record is filed instead of a BI-V150 interconnect record with an empty bandwidth field: an empty bandwidth on an otherwise ordinary-looking record reads as 'not applicable', which is how a reader concludes the part has no fabric or that nobody checked. Here the negative is stated, dated, sourced to the exact URL that was read, and set at verified with confidence 0.8 - the confidence is in the NEGATIVE being correctly established, not in any figure. BI-V250 IS DELIBERATELY NOT COVERED BY THIS NEGATIVE, and that restraint matters. The sibling URL https://www.iluvatar.com/productDetails?fullCode=cpjs-yj-xlxl-tg200 was also tested on 2026-10-05 and returned a Nuxt application-error page - 'An error occurred in the application and your page could not be served', 177 characters, no product content. A PAGE THAT ERRORS IS NOT EVIDENCE THAT CONTENT IS ABSENT: it may have been populated and is temporarily broken, and in this repo a wrong confident record is worse than an open question. So the BI-V250 appears only in the version string here, with no negative asserted for it, and the failed fetch is recorded in the source notes so a future pass does not repeat it. The name discrepancy in that URL matters too - the aggregator lists 'Tiangai 200 (BI-V200)' while the corpus record says BI-V250, and a renaming or a genuinely different part would produce exactly this signature. THE FINDING IS THE DISCLOSURE TREND, NOT THE MISSING NUMBER. Read alongside interconnect/iluvatar-bi-v100-chip-to-chip, which holds 64 GB/s bidirectional for the previous generation, the picture is that Iluvatar's disclosure went BACKWARDS. The BI-V100 (2021) publishes memory type, capacity, TDP and two interconnect figures, however briefly. The BI-V150 (2025), a part Iluvatar markets as supporting the latest and largest Llama and Qwen versions, publishes a shorter page than the one before it and contains no hardware specification section at all. For anyone evaluating Iluvatar as a scale-up vendor, the actionable fact is not that 64 GB/s is unimpressive - it is that the vendor's newest training part has no published interconnect whatsoever, so the fabric cannot be evaluated from primary material at all. WHAT WOULD OVERTURN THIS: any Iluvatar datasheet, product-selection guide, hardware manual or archived capture of the BI-V150 series stating an interconnect figure. The third-party aggregator mirrorfrog.com had already warned 'Detailed specifications of TianGai 150 are not fully disclosed by official sources', previously readable as third-party caution; that comment is now confirmed to be an accurate description of the vendor rather than hedging, which is the sort of aggregator statement worth re-checking against the primary source once and recording. KIND: filed as 'other' because the schema's enum has no entry for a vendor interconnect with no specified mechanism. 2026-10-10: set null_bandwidth=verified-negative. The record's own basis/notes already establish the reason; the field only names the case the schema now distinguishes. No number changed.