Huawei's not that bad 1152 GB HiZQ 2.0, *can be stacked* this isn't DGX Station
Dmitry Rybin@DmitryRybin1
Unreleased 8x Ascend 950 SuperPoD (codenamed 850E)
8:51 PM · Jul 1, 2026 · 4.5K Views
Huawei reveals Atlas 850E SuperPoD server with Ascend 950 NPUs, claiming 14.24 PFLOPS of FP4 compute
The system's per-chip specs slightly beat Nvidia's H200 GPU.
514752035.5K
Sentiment
Some users see the Huawei Atlas 850E SuperPoD as useful for MoE inference and small experiments due to its memory capacity, while others dismiss interest in the hardware with insults.
Pos
50.0%
Neg
50.0%
2 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS20.5KBOOKMARKS15LIKES96RETWEETS1REPLIES3
Zephyr@zephyr_z9
looks nice Specs are solid Total memory bandwidth is 4TB/s per chip A bit stronger than H100s
Dmitry Rybin@DmitryRybin1
Unreleased 8x Ascend 950 SuperPoD (codenamed 850E)
3hViews 20.5KLikes 96Bookmarks 15
Zephyr@zephyr_z9
looks nice Specs are solid Total memory bandwidth is 4TB/s per chip A bit stronger than H200s
Dmitry Rybin@DmitryRybin1
Unreleased 8x Ascend 950 SuperPoD (codenamed 850E)
3hViews 7.2KLikes 9Bookmarks 0
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
will be funny if they offer many such boxes
tphuang@tphuang
中昊芯英 unveils next gen TPU AI chip as part of 泰则 2.0 compute cluster that will have 896 TFLOPS compute per chip (1792 TOPS INT8). Each chip uses just 600W of power.
In a 8 TPU + 2 CPU setup, box will have 7.168 PFLOPS of compute. Natively supports PyTorch, vLLM, SGLang & other tools.
Completed integration w/ Qwen, DeepSeek, GLM & Minimax.
Company uses Chiplet + 2.5D packaging & connected in a cluster via optical modules & w/ OCS.
Founder/CEO 杨龚轶凡 was a core member of Google TPU design team for v2/3/4.
This does feel like the Chinese Google TPU.
3hViews 1.6KLikes 2Bookmarks 0
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
@zephyr_z9 It's weaker than H200s but has slightly more total memory I think this is pretty good for MoE inference and small experiments
Zephyr@zephyr_z9
looks nice Specs are solid Total memory bandwidth is 4TB/s per chip A bit stronger than H100s
2hViews 551Likes 4Bookmarks 0
IamPengu@iampengu911
@zephyr_z9 Do you think memory stocks have topped for the year?
3hViews 130
Peter Wigand@Peter_Wigand
@teortaxesTex You should counsel @yacineMTB on shopping-therapy issues, @teortaxesTex. Just which computer parts with lengthy names I can't be arsed to look up are the ones which define your innermost hopes and dreams?
3hViews 3