As we all know, the past couple of years have been tough for HW, but can things turn around for at least its Hisilicon division when Ascend deliveries installations ramp up? I think they could.
As an example, it looks like HW has 8 EFLOPS over a couple of years. Let's say it sells 5 EFLOPS of GPUs a year and consumes another 5 EFLOPS for personal use. For reference, I think it was Megvil that said you need 10000 A100 to have a competitive GPT platform. That works out to 3.2 EFLOPS. Given the craze over GPT, I don't think 5 EFLOPS for internal usage + 5 EFLOPS for external sales is that crazy.
Let's see how much 5 EFLOPS would cost. It would work out to be around 15000 Ascend 910 GPUs (since each is about 320 TFLOPS)
if we get a Ascend 920 that does 800 TFLOPS (which is where H100 is at), then that's would require 6000 Ascend 920
How much would that cost? well according to this
Each A100 retails for about 12k at least and H100 at least 33k.
Let's say HW GPUs are a little cheaper so 12k for Ascend 910 & 30k for Ascend 920. Also in both cases, they come with not just the GPUs, but the HBM memory and whatever else GPGPU needs and all the support.
12000 * 15k = 180 million
30k * 6k = 180 million
In both cases, I think HW can make $180 million from just the GPUs and this is not even counting all the software platform and such associated with setting up the data centers and also the CPUs, server racks and such.
If HW needs 10 EFLOPS between external & internal sales? That would be about 30k Ascend 910 or 12k Ascend 920.
How many GPUs can you fit on a 12-inch wafer? Let's say they stack 2 die of about 600mm2 (similar to Biren). Each die will be around 25x25. on a 150mm diameter, you can fit 5 across. Works out to be about 80 per wafer. If we assume 60% yield at this point (TSMC is at low 60s right now for 3nm based on report) for a very highly complex wafer, then you get about 50 good die per wafer -> 25 Ascend 920. To get 12000, you will need just 500 wafers. That's really not a lot.
Let's say China really hit the accelerators with 40 EFLOPS a year and 3/4 of that demand ends up with Ascend GPUs. Even with that, we are looking at just 1500 wafers a year and just a couple of billion $ in revenue for HW. Maybe I'm overcounting the # of GPU die you can fit per wafer. But even if you can only get 20 good die per wafer, 10 Ascend 920, you are still just looking at most 5000 wafers, which is not all that much.
In another year of two, if SMIC can do 5nm production, even if it's only 3k wpm, that's enough to handle all the most important HPC needs of Chinese data centers. It's the mobile SoC that tend to use up all of TSMC capacity. At some point, SMIC will need to get into that business too.
