Quick Take
Atlas 960E SuperPoD is a Huawei NPO SuperPoD launched 17 September 2026: 4,096 NPUs, 8 EFLOPS FP8, 16 EFLOPS FP4, UnifiedBus plus Hi-ONE, fully liquid-cooled. It is announced, not a 2026 shelf SKU. Atlas 950 SuperPoD (1,024 NPUs shown) is the 2026 RFQ default. Do not invent SuperPoD-total kW; 1 PB HBM is media only.
What is Atlas 960E SuperPoD?
Definition: Atlas 960E SuperPoD is a Huawei NPO SuperPoD announced at HUAWEI CONNECT 2026 in Shanghai on 17 September 2026, in David Wang’s keynote Advancing the Agentic World, Building a Solid Silicon Foundation. Official interconnect is UnifiedBus plus Hi-ONE. Architecture is orthogonal and fully liquid-cooled. Memory addressing is unified via UnifiedBus. Search aliases 昇腾960超节点 / Ascend 960 SuperNode appear in channel copy — mention them once, then RFQ the English product name Atlas 960E SuperPoD.
A 960E RFQ is a SuperPoD hardware column for 2027 planning. It is not a 2026 ex-stock line, not an Ascend 960DT or 960PR chip PO, not TaiShan 950 SuperPoD (general-purpose / Kunpeng, up to 4,096 nodes and a 256 TB memory pool), and not Atlas 900 A3 SuperPoD (384 × Ascend 910 plus LingQu 630 V1). Keep those names on separate columns.
| Name | What it is | What it is not |
|---|---|---|
| Atlas 960E SuperPoD | 4,096-NPU NPO SuperPoD (announced 17 Sep 2026) | A 2026 shelf SKU; not 960DT/960PR chip POs |
| Atlas 950 SuperPoD | 1,024-NPU (shown) UnifiedBus SuperPoD | Atlas 960E; not TaiShan 950 |
| Atlas 900 A3 SuperPoD | 384 × Ascend 910 + LingQu 630 V1 map | 960E |
| TaiShan 950 SuperPoD | General-purpose SuperPoD, up to 4,096 nodes, 256 TB pool | Atlas AI SuperPoD |
| Ascend 960DT / 960PR | Next-gen NPUs (roadmap GA Q1 2027 / Q3 2027, pending vs OEM datasheet) | SuperPoD hardware SKU |
| KunLun G8680 V3 | Locked 8-NPU node 02315QNK | A SuperPoD |
| NVIDIA GB300 NVL72 | Peer scale-up class (~72-GPU rack) | A Huawei SuperPoD equivalent SKU |
- Fit: teams planning a 2027 NPO SuperPoD; comparing Atlas 950 (orderable path) versus Atlas 960E (architecture / NPO); facility teams sizing liquid cooling plus UnifiedBus versus 800G-module halls; CANN / Ascend ecosystem buyers.
- Unfit: treating 960E as 2026 ex-stock; mixing 960DT chip lines onto a SuperPoD cabinet PO; copying 384 LingQu 630 V1 56+48 onto 960E; inventing SuperPoD-total kW; treating 960E as drop-in NVIDIA NVL72; using 1 PB HBM as a datasheet lock.
Official Atlas 960E SuperPoD numbers (HUAWEI CONNECT 2026)
Huawei’s English news page for the 17 September 2026 launch is the Track 1 lock. Scale-up inside one SuperPoD is 4,096 NPUs. Compute is 8 EFLOPS FP8 and 16 EFLOPS FP4. Optics class is 5,500 Hi-ONE engines instead of 48,000 × 800G optical modules. Interconnect power is cut by over 550 kW. Availability language: doubles fault-free operating time; 99.8% system availability. SuperCluster can use UnifiedBus or RoCE; two-tier four-plane Clos up to 512,000 NPUs; multi-rail topology up to 1,000,000 NPUs. One petabyte of HBM is not on this official page.
| Item | Official value | RFQ note |
|---|---|---|
| Event | HUAWEI CONNECT 2026, Shanghai, 17 Sep 2026 | Keynote product, not a dated ATP |
| Product | Atlas 960E SuperPoD — world’s first NPO-based SuperPoD | English product name on the PO |
| Interconnect | UnifiedBus + Hi-ONE | Roles, not invented MPNs |
| Cooling | Fully liquid-cooled | Do not copy channel air-Q2 / liquid-Q3 splits |
| Scale-up | 4,096 NPUs in one SuperPoD | Not SuperCluster Clos |
| Compute | 8 EFLOPS FP8 / 16 EFLOPS FP4 | Official news page |
| Optics | 5,500 Hi-ONE vs 48,000 × 800G modules | NPO class, not a pluggable count to invent |
| Interconnect power | Cuts consumption by over 550 kW | Interconnect save, not hall kW |
| Availability | 99.8% system availability | Official wording |
| SuperCluster | Clos to 512,000 NPUs; multi-rail to 1,000,000 | Separate RFQ from one SuperPoD |
| 1 PB HBM | Not on the official news page | Media / pending confirmation |
Atlas 950 SuperPoD, shown as a true machine at WAIC on 17 July 2026, is 1,024 NPUs, 1 EFLOPS FP8 / 2 EFLOPS FP4, 256 TB unified global memory, and RTT class 3 μs. HiAscend cabinet class is 64 × 950DT NPU, about 100 kW, liquid plus air, 44OU ≈ 47U — cite as cabinet class, not 960E hall power. MWC language scales 64 NPU/cabinet up to 8,192 NPUs; do not collapse 1,024 versus 8,192 into one PO line.
What Hi-ONE NPO changes versus 800G modules and CPO
Hi-ONE is Huawei’s High-density Optical-interconnect-Node Engine. The same news-page FAQ locks NPO, 7.2 Tbit/s per engine, first NPO ready for mass production, and currently the only NPO with a built-in light source. Atlas 960E SuperPoD uses 5,500 Hi-ONE engines instead of 48,000 × 800G optical modules. That is the optics class freeze: NPO engines, not a paste of conventional SuperPoD pluggable counts.
CPO (co-packaged optics) cost-delta percentages that appear in channel blogs stay industry estimate / pending confirmation. Do not write an NPO-versus-CPO dollar table on this RFQ. Do not invent optic MPNs. UnifiedBus extras from Yang Chaobin’s 17 September 2026 interconnect news: protocol plus memory semantics; RTT class 7 μs → 2 μs; interconnect 100 GB/s → TB/s. LinkBlade (in-cabinet) is cable-free and reduces copper in a 4,096-NPU SuperPoD by about 196 km. LinkDevice (between cabinets) is 176 ports, 1.6 Tbit/s per port, 280 Tbit/s all-optical, RTT as low as 2 μs. UBG switch radix is up to 1,024 for million-NPU SuperCluster. Name those as roles; quantity stays pending a dated quote. Do not turn LinkBlade, LinkDevice, UBG, OceanStor M900, or Xinghe into invented MPNs.
- 960E optics: Hi-ONE NPO, 5,500 engines vs 48k × 800G.
- 950 optics: conventional SuperPoD optics class — freeze against a dated 950 quote, not against 960E Hi-ONE counts.
- CPO %: industry estimate / pending confirmation — not a PO line.
Atlas 960E SuperPoD vs Atlas 950 SuperPoD — which should I RFQ now?
RFQ Atlas 950 SuperPoD now if the hall must purchase in 2026. Put Atlas 960E SuperPoD on a 2027 planning appendix. Traditional 8-NPU servers stay a different envelope (PCIe plus NIC plus Ethernet/RoCE). Liao Heng’s public Huawei wording for an arithmetic demo only: system performance ≈ SuperPoD scale × chip spec. 4× cards × ~2× chip class ≈ 8× node FP8 (1 → 8 EFLOPS). Label that as Huawei public formula, not a quote.
| Dimension | Atlas 950 SuperPoD | Atlas 960E SuperPoD | Traditional 8-NPU servers (class) |
|---|---|---|---|
| Single-domain NPUs | 1,024 (shown) | 4,096 | 8 per box |
| FP8 / FP4 | 1 / 2 EFLOPS | 8 / 16 EFLOPS | N/A on this page |
| Interconnect | UnifiedBus (optical SuperPoD) | UnifiedBus + Hi-ONE NPO | PCIe + NIC + Ethernet/RoCE |
| Optics class | Conventional SuperPoD optics | 5,500 Hi-ONE vs 48k × 800G | Many pluggable 800G |
| Cooling | Liquid + air (cabinet page) | Fully liquid-cooled | Often air / mixed |
| Status | Nearer commercial / shown 2026 | Announced; not 2026 shelf | Shipping |
| RFQ now? | Yes — default 2026 SuperPoD RFQ | Planning / 2027 path | Different envelope |
Limitation of this page: it does not publish channel net, SuperPoD-total kW, or a dollar TCO versus NVIDIA GB300 NVL72 / Vera Rubin NVL72. Compare architecture class only. Atlas 900 A3 384 remains the nearer LingQu 630 V1 map — back-link that RFQ instead of copying 56+48 onto 960E.
Nested BSP, UnifiedBus, and the Liao Heng formula
Heng Liao’s paper Nested Parallel von Neumann Architecture and Nested BSP, arXiv:2609.16787, submitted 15 September 2026, is Track 1 architecture, not a SKU. Nested BSP (software) plus Nested Parallel von Neumann (hardware) plus Unified Bus; slogan Physically Sparse, Logically Tight. Huawei also uses Peerium Computing Architecture in architecture PR — keep it as an architecture name, not a purchase-order part.
Chip roadmap is a separate column from the SuperPoD PO. Public coverage (Tom’s Hardware / Straits Times, pending confirmation versus OEM datasheet) places Ascend 960DT ready Q1 2027 and Ascend 960PR Q3 2027, with 970 (2028) / 980 (2029) as later marks. Do not treat chip GA as Atlas 960E SuperPoD ship date. CANN is now community open-source as of the official 17 September wording; migration cost still pending confirmation on the RFQ.
Real-world deployment scenarios
Four enterprise pictures decide which column gets a dated RFQ. Do not collapse them into one mashup PO.
- 2026 train/infer hall that must PO this year. Default the SuperPoD hardware column to Atlas 950 SuperPoD (or Atlas 900 A3 384 if the LingQu map still fits). Atlas 960E SuperPoD stays on the planning appendix until OEM prints a ship window.
- 2027 10-trillion-parameter / agentic SuperCluster plan. RFQ 960E architecture plus SuperCluster Clos / multi-rail as separate lines. Ship remains pending OEM. Do not copy 2026 950 cabinet ~100 kW as 960E hall power.
- Optics and power team. Freeze Hi-ONE NPO versus 48k × 800G and the official interconnect save of over 550 kW. Still no invented hall kW. LinkBlade / LinkDevice / UBG stay roles.
- SME / small SuperPoD. Atlas 650E / 8–16 NPU UnifiedBus appliances from the Yang keynote are not this 4,096 page. Send those buyers to a different envelope.
How to RFQ Atlas 960E SuperPoD versus Atlas 950 SuperPoD
Six steps freeze product, chip generation, scale-up versus SuperCluster, optics and cooling, software, and a dated RFQ without invented MPNs or SuperPoD-total kW.
- Step 1 — Lock the product column. Atlas 950 SuperPoD vs Atlas 960E SuperPoD vs Atlas 900 A3 384 — one PO, not a mashup.
- Step 2 — Lock NPU generation. 950DT vs 960DT/960PR vs 910 — chips are not SuperPoD SKUs.
- Step 3 — Lock scale-up vs scale-out. One SuperPoD (1024 or 4096) vs SuperCluster (512k Clos / 1M multi-rail) vs TaiShan 950 CPU SuperPoD.
- Step 4 — Freeze optics and cooling. 960E = Hi-ONE NPO + full liquid; do not paste 800G-module counts or air-cooled SuperPoD GA from channel blogs.
- Step 5 — Freeze software. CANN now community open-source (official 17 Sep) — migration cost still pending confirmation on the RFQ.
- Step 6 — Send a dated RFQ. Cabinets / UnifiedBus roles (LinkBlade, LinkDevice, UBG) as roles not invented MPNs; support years; no public net; no SuperPoD-total kW.
Shanghai TaoHan Industrial reviews the split BOM against hall drawings for China enterprise procurement. Soft lead times stay industry estimate / pending confirmation until the channel prints a dated ATP.
Common RFQ pitfalls
Name mix, media HBM, invented kilowatts, and chip-versus-SuperPoD mashups are the four expensive mistakes after HUAWEI CONNECT 2026.
- 960E as 2026 shelf. The official news page announces; it does not print a commercial ship date.
- 960E = 960DT. SuperPoD hardware is not a chip PO. Keep Q1/Q3 2027 chip GA off the cabinet line.
- 1 PB HBM as datasheet. Not on the official Atlas 960E page — media / pending confirmation.
- Invented SuperPoD-total kW. Atlas 950 cabinet ~100 kW is cabinet class. Interconnect save over 550 kW is not hall power.
- 56+48 LingQu on 960E. That pair belongs to the Atlas 900 A3 / 800T A3 map, not this NPO SuperPoD.
- TaiShan 950 on an Atlas PO. CPU SuperPoD, 4,096 nodes, not Atlas 960E.
- Drop-in NVL72. Compare architecture; do not invent dollar TCO.
FAQ
Atlas 960E SuperPoD is the 4,096-NPU NPO SuperPoD announced 17 Sep 2026, not a 2026 shelf SKU and not Ascend 960DT. RFQ Atlas 950 SuperPoD for 2026 POs. Hi-ONE is NPO at 7.2 Tbit/s. 1 PB HBM is media only. NVIDIA NVL72 is architecture class, not a drop-in SKU.
What is Atlas 960E SuperPoD?
Is Atlas 960E the same as Ascend 960DT?
Can I buy Atlas 960E SuperPoD in 2026?
Atlas 960E vs Atlas 950 SuperPoD — which should I RFQ now?
What is Hi-ONE, and how is NPO different from CPO?
Did Huawei publish 1 PB HBM for Atlas 960E?
How does Atlas 960E compare to NVIDIA GB300 NVL72?
References
Validate Atlas 960E SuperPoD numbers on Huawei’s 17 September 2026 news page. Validate Atlas 950 SuperPoD on the July WAIC news page and HiAscend cluster tag 950. Architecture extras sit on the LingQu / UnifiedBus news page and arXiv:2609.16787. This page publishes no list dollars.
- Huawei — Atlas 960E SuperPoD news (HUAWEI CONNECT 2026)
- Huawei — UnifiedBus / LingQu interconnect news (17 Sep 2026)
- Huawei — Atlas 950 SuperPoD WAIC news
- HiAscend — Atlas 950 cluster hardware
- arXiv:2609.16787 — Nested Parallel von Neumann / Nested BSP
- Huawei — HUAWEI CONNECT event page
Key Takeaway
Key Takeaway: Freeze 2026 SuperPoD CapEx on Atlas 950 SuperPoD (or Atlas 900 A3 384). Keep Atlas 960E SuperPoD on a 2027 NPO appendix: 4,096 NPUs, Hi-ONE, full liquid, no invented kW, 1 PB HBM media-only. Shanghai TaoHan Industrial helps China enterprise buyers split those columns on a dated RFQ — without publishing list prices.
Related reading
Need Atlas 950 on the 2026 PO and 960E on a 2027 appendix?
If product column, Hi-ONE NPO optics, or SuperCluster Clos lines are still mixed, send the draft BOM to Shanghai TaoHan Industrial before CapEx freeze. We help split Atlas 950 SuperPoD, Atlas 960E SuperPoD, and Atlas 900 A3 384 — without publishing list prices.
Contact TAOHAN · SuperPoD BOM splitEmail: info@021taohan.com