How much would it cost to design a custom ASIC that can compute large amount of numeric data on a hard drive? The algorithm is static and we were wondering if we can do this faster than GPUs.
Are there large setup fees involved to manufacture prototypes?
Unless you have like $100 million dollars in the bank abandon all hope. Also your idea sounds dumb as shit.
>>55444103
Have you tried a FPGA (field programmable gate array)?
For repeated algorythms, same task over and over again its much faster than a regular processor. More expensive and difficult to work with, but much cheaper than trying to start a production of ASICs.
>>55444103
why would you want to compute that?
and yes, it is expensive, at least 30m just to start and probably no you can't make it faster than gpus because gpus are glorified parallel calculators
>>55444103
disregard >>55444142 , he's a retard
ASIC is pricy. I've heard estimations that it's worth it when you need more than 10k units.
But do not despair, FPGAs are really cheap nowadays and sound just like the right tool for what you need.
Their speed is about the same as ASIC. Good luck
>>55444274
>you need more than 10k units
i'm pretty sure there are companies that make asics case by case, and it is cheaper than "conventional" supercomputer
>>55444329
you can make it case by case but you're gonna spend about the same as if you need 10k units, so you will save a lot of money if you use an equivalent FPGA
Anyone know which is the best hardware to buy to run this - https://github.com/JarrettR/FPGA-Cryptoparty
>>55444380
You can now buy unused wafer space from some fabs for low-run ASICs for five digits or even less if you're doing the design and layout yourself, the cores aren't too big and are self-contained, you satisfy all the requirements of the process, and you don't have any more metal layers than are already on the wafer. People have started pooling this - some universities are even getting students' ASICs fabbed as part of their course.
Don't expect to get to use the latest and greatest processes this way, but I got ten test ASICs, for five digits. It was wasted money: they didn't work above about 1MHz, but that was my fault for fucking up the design spectacularly - in validation the process was absolutely fine.
Prototyping is best done on FPGAs, or with simulators: taping out test units early on isn't something that's really advisable.
You can also get circuit boards done amazingly cheap in low quantity now, if you don't want complex ones and you're not super-picky about the quality.
China, man. Ask around in Shenzhen.
>>55444103
If you're asking how much it would cost then you shouldn't even begin thinking about an asic. FPGA is all of the ASIC without the hassle, I'm pretty sure most CS students learn to design their own GPU on whatever market flavor of FPGA.
Bitcoin miners were using FPGAs before spinning their own silicon.
Also "compute a large amount of numeric data" is incredibly vague, maybe you're going about whatever you're trying to do wrong.
