Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
These "new" chips look more like their sanctioned chips with a couple functional units disabled. They are likely recovering low binned GPUs with lots of broken units due to surface errors, disabling the broken units, and selling them to the Chinese market.
From Chinese internet
转：[NV新品解读]近期H20 L20相关信息在传播，目前H20的设计思路是阉割单卡算力，满配HBM存储，满配NVLINK速率。由于单卡算力阉割的厉害，集群算力规模较低，算力阉割至1/6(原有性能20%不到)，按照传输与计算均衡来预估，集群算力规模在1～2万P，再往上计算能力较差，难以跟H100/H800的规模化10万P以上算力抗衡，因此难以胜任GPT-4等训练任务，但是在大模型推理场景下有一定适用度。
Redirect: [NV New Product Interpretation] Information related to H20 L20 has been spreading recently. The current design idea of H20 is to reduce the computing power of a single card, fully equipped with HBM storage, and fully equipped with NVLINK speed. Due to the severe emasculation of the computing power of a single card, the scale of the cluster computing power is relatively low, with the computing power reduced to 1/6 (less than 20% of the original performance). It is estimated based on the transmission and computing balance that the cluster computing power scale is 1 to 2 10,000 P, and the computing power above is poor, and it is difficult to compete with the large-scale computing power of H100/H800 of more than 100,000 P. Therefore, it is difficult to be competent for training tasks such as GPT-4, but it has certain applicability in large model inference scenarios.
转：[NV新品解读]近期H20 L20相关信息在传播，目前H20的设计思路是阉割单卡算力，满配HBM存储，满配NVLINK速率。由于单卡算力阉割的厉害，集群算力规模较低，算力阉割至1/6(原有性能20%不到)，按照传输与计算均衡来预估，集群算力规模在1～2万P，再往上计算能力较差，难以跟H100/H800的规模化10万P以上算力抗衡，因此难以胜任GPT-4等训练任务，但是在大模型推理场景下有一定适用度。
Redirect: [NV New Product Interpretation] Information related to H20 L20 has been spreading recently. The current design idea of H20 is to reduce the computing power of a single card, fully equipped with HBM storage, and fully equipped with NVLINK speed. Due to the severe emasculation of the computing power of a single card, the scale of the cluster computing power is relatively low, with the computing power reduced to 1/6 (less than 20% of the original performance). It is estimated based on the transmission and computing balance that the cluster computing power scale is 1 to 2 10,000 P, and the computing power above is poor, and it is difficult to compete with the large-scale computing power of H100/H800 of more than 100,000 P. Therefore, it is difficult to be competent for training tasks such as GPT-4, but it has certain applicability in large model inference scenarios.
Last edited:
Yeah, The EUV, immersion and even the nanoimprint market in China is pretty much open right now, this is a huge opportunity.
I posted this earlier in the thread already. It'd be nice if you can look back and acknowledge that they are linked, so I don't have to look through the entire article to find out that it's the same server machine
yes, you are right bout that
dylan always gets alarmed. It gets seriously old
Chinese chip maker SMIC’s sales fall for third straight quarter despite boost from Huawei’s 5G smartphone revival
- SMIC reported a 15 per cent fall in revenue in the September quarter to US$1.62 billion, missing analysts’ projection of US$1.64 billion
- Net income dropped 80 per cent to US$94 million, compared with estimates for US$178.1 million
- Status