天池 1024 卡超节点

天池 1024 卡超节点 (Baidu Tianchi 1024-card supernode) was announced by Baidu group EVP Shen Dou at the opening of Baidu World 2026 on 12 November 2026. IT Home’s on-site report (2026-11-12 12:09:14) said the cluster will go to market next year. Sina Tech carried a matching account the same day.

Per that on-site account, the supernode uses an optical engine and raises inter-card interconnect bandwidth by 3×; versus the previous generation, training performance rises 2.5× and inference 3×. A companion G3.5 KV Cache cabinet is said to double long-context inference performance, cut TTFT by 90% when multi-turn cache hits, and raise overall throughput 7×. A single suite is said to support training a 5-trillion-parameter model and to scale horizontally toward 10,000-card clusters. IT Home also noted that Baidu had last year targeted a thousand-card Tianchi supernode for 2028, and that the schedule is now one year earlier; it additionally mentioned the Kunlunxin N series moving to a 2028 launch. All of these figures come from the report, not from independent tests.

The existing official product page cloud.baidu.com/product/tianchi.html, checked in this scan, still mainly describes the current Tianchi supernode (for example “32-card one-layer all-to-all” and “1U 4-card dense design”) and did not mention “1024” or the new performance claims. Treat the 1024-card specs as keynote reporting until the product page or docs are updated.

Key points:

• On-site report: https://www.ithome.com/1/026/026.htm ; existing product line page: https://cloud.baidu.com/product/tianchi.html ;
• This is a specific new SKU name under Tianchi supernodes. Do not describe it as the already listed 32-card system, or as a cloud instance you can order today.

Why it matters

Baidu named a 1024-card supernode at Baidu World on 12 November 2026 and gave a public go-to-market window (next year). That is a concrete infrastructure product name, distinct from the current Tianchi supernode already on the product page.

This page only restates on-site reporting and notes that the official product page has not yet been updated. Multipliers, the 5-trillion-parameter claim and TTFT figures are official statements via the press, not third-party benchmarks.

Limits

The SKU is not on sale yet. The official page still describes the existing 32-card design. Do not treat the 1024-card supernode as orderable, or the reported speedups as independent measurements. The companion G3.5 KV Cache cabinet should wait for fuller official documentation before being treated as a separate entry.

1. Read the IT Home on-site report https://www.ithome.com/1/026/026.htm and check whether https://cloud.baidu.com/product/tianchi.html has been updated with 1024-card specifications;
2. For planning or procurement, wait for Baidu AI Cloud’s later official docs, quotes and delivery terms; this page is not an ordering guide;
3. When quoting bandwidth, training/inference multipliers, 5-trillion-parameter capacity, TTFT or “available next year,” cite the 2026-11-12 on-site report and re-check upstream;
4. Do not describe the Tianchi 1024-card supernode as the already listed 32-card Tianchi supernode.