Alibaba T-Head Zhenwu 810E
Product Overview
Zhenwu 810E is Alibaba T-Head's first-generation training-inference integrated AI chip of the Zhenwu series, unveiled in January 2026 and positioned in T-Head's official roadmap as the baseline (first-generation) product of the series.
The 810E features a training-inference integrated design that serves both large model training and high-density inference. It is equipped with 96 GB HBM2e memory and an inter-chip interconnect bandwidth of 700 GB/s (composed of 7 proprietary ICN high-speed links), delivering single-card bandwidth comparable to the NVIDIA H20. According to the South China Morning Post, T-Head describes the 810E as a Parallel Processing Unit (PPU) for AI training and inference.
In terms of power, the 810E's board-level power is about 400 W, lower than the H20's 550 W, giving it better energy efficiency at equivalent bandwidth. The interface is PCIe 5.0 ×16.
⚠️ Timeline discrepancy (important): The roadmap chart of some English media (WCCFtech) marks the 810E as "2024 Q2", which is inconsistent with the "January 2026 unveiling" reported by TrendForce, Synced, and domestic media (TMTPost, etc.). This site adopts January 2026 (cross-validated by multiple sources); the 2024 Q2 entry in the roadmap chart is likely an internal project initiation or tape-out date.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | T-Head proprietary parallel computing architecture (PPU) |
| Process Node | Not disclosed |
| Memory Capacity | 96 GB HBM2e |
| Inter-Chip Interconnect Bandwidth | 700 GB/s (7× proprietary ICN links) |
| Interface | PCIe 5.0 ×16 |
| TDP | 400 W (board-level power, estimated) |
| Debut | January 2026 |
| Series Positioning | First-generation baseline product of the Zhenwu series |
| Comparison Baseline | Single-card bandwidth comparable to the NVIDIA H20 |
| Price | Not disclosed |
⚠️ Specification notes: Process node, absolute compute, and memory bandwidth are all undisclosed. The 400 W TDP is an industry-estimated value (based on the statement that it is below the H20's 550 W); refer to T-Head's subsequent official datasheet for confirmation.
Generational Comparison with the M890
| Parameter | Zhenwu 810E | Zhenwu M890 | Improvement |
|---|---|---|---|
| Memory Capacity | 96 GB HBM2e | 144 GB HBM3 | +50% |
| Inter-Chip Interconnect Bandwidth | 700 GB/s | 800 GB/s | +14% |
| Architecture | Proprietary parallel computing architecture (first generation) | Proprietary parallel computing architecture (full upgrade) | 3× overall performance |
| Positioning | Training-inference integrated (baseline) | Training-inference integrated (Agentic-optimized) | — |
| Unveiled | January 2026 | May 2026 | 4 months apart |
📌 Only 4 months separate the 810E from the M890, reflecting T-Head's significantly accelerated iteration cadence in 2026 and its established "one generation per year" roadmap commitment.
Relationship to the Site's Alibaba PPU Card
The site's "Alibaba PPU" card records the early project-stage information of the Zhenwu series (temporary codename "PPU", 96 GB HBM2e, 700 GB/s, 400 W, TSMC tape-out in 2023). As the Zhenwu brand entered commercial use, the temporary codename was superseded by official models such as the Zhenwu 810E / M890. Refer to this card and the "Zhenwu M890" card for the latest information.