Chinese semiconductor thread II

[tphuang]

If Huawei's Ascend 910C truly has comparable performance to H100's, then US strategy of offering crippled H100s (i.e., H20s) has objectively and completely backfired. US wants to keep a foot in the door in the second largest AI market, yet inadvertently ceded the entire market to Huawei monopoly with its outdated and draconian sanction regime. It's truly a gift that keeps on giving. The cherry on top is that generative AI and LLMs have remarkably little-to-zero discernible military implications. So economically and militarily, you really have to question how US got so arrogant to believe they can simply stop AI development in China.

Not exactly. Ascend-910C might be comparable to H100 in TFLOPS and such. But there are questions about chip-to-chip interconnect and HBM and such. We are going to have to wait and hear back on reports. They are getting tested now.

The question is just how many AI chips does China need in the next 18 months. That's kind of time frame we are looking at to get LLMs that consumed the entire universe of data. Let's say they devote 3k wpm to Ascend-910C with 30 usable die per wafer. That would be 90k AI chips per month and 1.1m per year. Or around 1.6m over next 18 months. That seems a little less than ideal. Chinese hyperscalers will be fight an uphill battle here vs American hyperscalers.

Now there are rumors that SMIC has to allocate some 7nm capacity to other chip designers. Maybe that's a sign SMIC is able to add more 7nm capacity. And that's certainly good news.

 

[staplez]

Not exactly. Ascend-910C might be comparable to H100 in TFLOPS and such. But there are questions about chip-to-chip interconnect and HBM and such. We are going to have to wait and hear back on reports. They are getting tested now.

The question is just how many AI chips does China need in the next 18 months. That's kind of time frame we are looking at to get LLMs that consumed the entire universe of data. Let's say they devote 3k wpm to Ascend-910C with 30 usable die per wafer. That would be 90k AI chips per month and 1.1m per year. Or around 1.6m over next 18 months. That seems a little less than ideal. Chinese hyperscalers will be fight an uphill battle here vs American hyperscalers.

Now there are rumors that SMIC has to allocate some 7nm capacity to other chip designers. Maybe that's a sign SMIC is able to add more 7nm capacity. And that's certainly good news.

Well there's also the simple problem of capacity. Just look at the ascend 910b. Initially everyone thought it would just take over. However, it turned out Huawei simply could not make enough and customers started buying the NVidia H20 cards. Will the new ascend 910c be able to sell enough to cover everyone's needs? Probably not to be honest.

And I'm not trying to diminish Huawei's success here. What has been accomplished in a mere 5 years is incredible. But the reality is even though in pure technology terms China is but a year or two behind, China simply doesn't have the capacity at the moment to cover all their needs.

 

[curiouscat]

Not exactly. Ascend-910C might be comparable to H100 in TFLOPS and such. But there are questions about chip-to-chip interconnect and HBM and such. We are going to have to wait and hear back on reports. They are getting tested now.

The question is just how many AI chips does China need in the next 18 months. That's kind of time frame we are looking at to get LLMs that consumed the entire universe of data. Let's say they devote 3k wpm to Ascend-910C with 30 usable die per wafer. That would be 90k AI chips per month and 1.1m per year. Or around 1.6m over next 18 months. That seems a little less than ideal. Chinese hyperscalers will be fight an uphill battle here vs American hyperscalers.

Now there are rumors that SMIC has to allocate some 7nm capacity to other chip designers. Maybe that's a sign SMIC is able to add more 7nm capacity. And that's certainly good news.

I think another concern is energy efficiency. The number one cost for most data centers is energy and HUAWEI most likely has substantially worse energy efficiency than Nvidia because of the node disadvantage from not having EUV.

 

[SanWenYu]

US wants to keep a foot in the door in the second largest AI market, yet inadvertently ceded the entire market to Huawei monopoly with its outdated and draconian sanction regime.

The US could have caused a much bigger trouble for China if they started the tech war with a total ban. But the greed and arrogance in them demand 即要 (market share) 又要 (to slow China down). Their moves so far are effectively 添油战术. It is no surprise that they are quickly losing both the market share and the tech war.

 

[latenlazy]

I think another concern is energy efficiency. The number one cost for most data centers is energy and HUAWEI most likely has substantially worse energy efficiency than Nvidia because of the node disadvantage from not having EUV.

Err no. Nvidia’s chips are extremely energy inefficient because they employ suboptimal architecture for AI acceleration tasks. Nvidia adapted them from GPUs. Huawei’s architecture isn’t.

 

[Moonscape]

I think another concern is energy efficiency. The number one cost for most data centers is energy and HUAWEI most likely has substantially worse energy efficiency than Nvidia because of the node disadvantage from not having EUV.

Energy is cheaper in China than in the US due to large scale investment in green energy.

Intelligent industrial policies lead to synergies not possible in a neoliberal regime.

 

[tphuang]

Well there's also the simple problem of capacity. Just look at the ascend 910b. Initially everyone thought it would just take over. However, it turned out Huawei simply could not make enough and customers started buying the NVidia H20 cards. Will the new ascend 910c be able to sell enough to cover everyone's needs? Probably not to be honest.

Doesn’t work like that. There are reasons why Nvidia chips still sell. Better chip interconnect speed & HBM3E availability among them.

and it will take some time for huawei to take over the market. Things don’t happen overnight.

And I'm not trying to diminish Huawei's success here. What has been accomplished in a mere 5 years is incredible. But the reality is even though in pure technology terms China is but a year or two behind, China simply doesn't have the capacity at the moment to cover all their needs.

That’s is certainly an important question. That’s why various Chinese chip firms all placed rush orders with TSMC recently.

On the other hand, if we agree that AI chip is more important than kirin chips, then why does huawei continue to use more wafers for the latter?

I think another concern is energy efficiency. The number one cost for most data centers is energy and HUAWEI most likely has substantially worse energy efficiency than Nvidia because of the node disadvantage from not having EUV.

Energy efficiency is not the issue here. China produces plenty of low cost energy. Think a little harder on why energy efficiency is important.

Err no. Nvidia’s chips are extremely energy inefficient because they employ suboptimal architecture for AI acceleration tasks. Nvidia adapted them from GPUs. Huawei’s architecture isn’t.

We don’t know this since we don’t have any data on the new chip.

 

[latenlazy]

We don’t know this since we don’t have any data on the new chip.

Very doubtful that Huawei would have built their chip around a GPU based architecture since they don’t have any GPU products.

 

[tphuang]

Very doubtful that Huawei would have built their chip around a GPU based architecture since they don’t have any GPU products.

They did build chip on their own NPU architecture, but I haven’t seen any evidence with ascend 910 for example that it would have notable energy efficiency advantage over Nvidia chips.

 

[pecopls]

I think another concern is energy efficiency. The number one cost for most data centers is energy and HUAWEI most likely has substantially worse energy efficiency than Nvidia because of the node disadvantage from not having EUV.

The cost of the chips themselves far outway the cost of the electricity. It's almost not a factor.

An H100 system draws maybe 1.4kW per hour. Assuming 75% up-time, it will draw about 9,000 kW per year. The cost of each kW is about US$ 0.1, so about US$ 900/year. By comparison, NVIDIA charges US$ ~25,000-30,000 per H100.

The reason why node shrink is important is to minimize the number of interconnects by maximizing per card compute.