Why do think all these big IT companies buys H100 like crazy？They cannot do simple math？I don't think they are that stupid
Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
Why do think all these big IT companies buys H100 like crazy？They cannot do simple math？I don't think they are that stupid
I don’t know, maybe it has something to do with CUDA and its vast ecosystem?
CUDA matters a lot; much more than processing speed. Even at large AI projects; developer productivity and being able to sit on existing work matters more than cost and power of hardware.
Wrt. A100 to H100 as far as I can read. The H100 has higher power consumption and price than the A100, making the total cost difference much less.
Wrt. A100 to H100 as far as I can read. The H100 has higher power consumption and price than the A100, making the total cost difference much less.
Shorter training time using H100 v A100 is worth a lot more to commercial enterpreeises.
The problem with contact lithography has always been that it is easier to accumulate dirt and worsen yield. Notice how Canon has been trying to sell this for NAND manufacture where the structures are highly regular and areas with defects can just be disabled.
Just for reference, chip yields used to be like 20% in the early days of single chip processors when the Motorola 6800 came out. A CPU would have cost over 100 USD. When they switched from contact lithography to projection lithography, like with the MOS 6502, chip yield went up to like 70%. And the chip cost like 20 USD.
Just for reference, chip yields used to be like 20% in the early days of single chip processors when the Motorola 6800 came out. A CPU would have cost over 100 USD. When they switched from contact lithography to projection lithography, like with the MOS 6502, chip yield went up to like 70%. And the chip cost like 20 USD.
Last edited:
Is it exclusively for memory chips? Looking at their own video on YouTube, they did mention the point of damage occuring during the stamping process being one of the problems, but they claim to have overcome it.
It might be useable for any chip with highly regular structures. A DRAM or NAND chip. Or a GPU even. But I wouldn't use it on complex logic chips like CPUs that's for sure. It probably has horrible yields. On a memory chip or a GPU you just disable the affected blocks. You test the chip and then the affected areas can have their traces cut with a laser or whatever.
This is another reason why when people like Dylan Patel claim that SMIC cannot manufacture GPUs because they are so big compared with CPUs it just shows their lack of knowledge.
Nanoimprint tools capable of 45nm or less are already included in the latest BIS restrictions, although not for de-americanized tools (0% US origin content), which currently applies to DUV tools. But if Canon tries to export it to China surely US would reduce the percentage to 0% for it too.
- Status