If you are getting a wafer fabbed for 5k, you are either on very old technology node, or you are talking about bulk pricing. You cannot get single wafers made for 5k on a modern node.
There’s a reason why modern nodes are used. If you manufacture in a node old enough to get cost in that ballpark, you will miss out on power efficiency and performance that just the technology gets you, and it won’t be a salable product. Gpus are high performance products. Older nodes are usually relegated to support chips.
You are also not likely to be able to reuse the work you did, on literally anything physical. You will perhaps be able to reuse logic in vhdl or verilog. Nothing synthesized, though.
This isn’t software. In software, you can write architecture agnostic code and compile and run it on processors that span decades. That ability is intentionally baked into the languages and processor architecture.
None of that is true in semiconductor manufacturing. There is no guarantee that you can close a design with the same logic just by rerunning synthesis and routing in a new technology. As a matter of fact, ask a physical design engineer. They’ll laugh at you. You’re acting as if they don’t do anything for a living.
Overall, I dont believe you can show a viable option on a node that old, and attempting to do so would be additional work that needs redone to be able to sell a product in a more recent node.
Incumbents have other advantages too. They have long term relationships with fabs, that helps with pricing and scheduling. There are also reciprocal licenses for patents between incumbents. You’d potentially accidentally run afoul of patent laws if you use covered techniques. You will want to use these techniques if you want to be competitive.
I understand the desire for a free and open option, but there are different obstacles in hardware than there is in software. Hardware requires heaps of money to get off the ground.
What is China making that is competitive with NA design houses in the gpu space? My understanding is that China is buying Nvidia parts nerfed by US law. Do you think there is not a big pile of funding that these Chinese efforts have been provided? The European union and Chinese government have very different economic levers they can pull
Rarely are full designs respun. That’s a huge waste of NRE cost. Lithography masks are millions of dollars.
Devices are characterized and techniques are explored on mpw designs. Multiple Pattern Wafer. Many customers design a small test die, and they are stitched together to make a single wafer and then you’ll get a few dies for evaluation. This way the cost is shared between many customers and experiments.
If even one mask (out of dozens) needs rework after a full design is getting fabbed, its considered a big error.
deleted by creator
If you are getting a wafer fabbed for 5k, you are either on very old technology node, or you are talking about bulk pricing. You cannot get single wafers made for 5k on a modern node.
deleted by creator
There’s a reason why modern nodes are used. If you manufacture in a node old enough to get cost in that ballpark, you will miss out on power efficiency and performance that just the technology gets you, and it won’t be a salable product. Gpus are high performance products. Older nodes are usually relegated to support chips.
You are also not likely to be able to reuse the work you did, on literally anything physical. You will perhaps be able to reuse logic in vhdl or verilog. Nothing synthesized, though.
This isn’t software. In software, you can write architecture agnostic code and compile and run it on processors that span decades. That ability is intentionally baked into the languages and processor architecture.
None of that is true in semiconductor manufacturing. There is no guarantee that you can close a design with the same logic just by rerunning synthesis and routing in a new technology. As a matter of fact, ask a physical design engineer. They’ll laugh at you. You’re acting as if they don’t do anything for a living.
Overall, I dont believe you can show a viable option on a node that old, and attempting to do so would be additional work that needs redone to be able to sell a product in a more recent node.
Incumbents have other advantages too. They have long term relationships with fabs, that helps with pricing and scheduling. There are also reciprocal licenses for patents between incumbents. You’d potentially accidentally run afoul of patent laws if you use covered techniques. You will want to use these techniques if you want to be competitive.
I understand the desire for a free and open option, but there are different obstacles in hardware than there is in software. Hardware requires heaps of money to get off the ground.
deleted by creator
What is China making that is competitive with NA design houses in the gpu space? My understanding is that China is buying Nvidia parts nerfed by US law. Do you think there is not a big pile of funding that these Chinese efforts have been provided? The European union and Chinese government have very different economic levers they can pull
Wdym by spin?
deleted by creator
Rarely are full designs respun. That’s a huge waste of NRE cost. Lithography masks are millions of dollars.
Devices are characterized and techniques are explored on mpw designs. Multiple Pattern Wafer. Many customers design a small test die, and they are stitched together to make a single wafer and then you’ll get a few dies for evaluation. This way the cost is shared between many customers and experiments.
If even one mask (out of dozens) needs rework after a full design is getting fabbed, its considered a big error.
deleted by creator
If that would be great, unfortunately it’s more like $18k (https://www.tomshardware.com/tech-industry/tsmcs-wafer-pricing-now-usd18-000-for-a-3nm-wafer-increased-by-over-3x-in-10-years-analyst). And that’s not the biggest cost issue, it’s actually the mask, which is more like in the millions of USD (https://semianalysis.com/2022/07/24/the-dark-side-of-the-semiconductor/)
deleted by creator