Everybody, get your money!!!
cgracey
Posts: 14,152
in Propeller 2
At 5nm, it will cost $500 million or more to design a “reasonably complex SoC,” Johnson said. In comparison, it will cost $271 million to design a 7nm SoC, which is about 9 times the cost for a 28nm planar device, according to Gartner.
I just wanted to be sure everyone starts saving now for Prop3.
http://semiengineering.com/going-to-gate-all-around-fets/
Comments
If prices come down over time, whilst everyone is saving over time, then perhaps we'd meet in the middle.
So 5nm at "just" $250 million may not seem so bad :-)
It's amazing how rarefied small processes have become. Never have transistors been so cheap and so expensive at the same time. I bet there are only 5 customers for this 5nm process. There may be 3 for 3nm.
Potential yield increases a factor of 16...
- the 90nm P2 in production
- a year or two after that, 45nm P3
Extrapolating from current P2 design, that should allow for 64 cogs with 2MB of Hub and 256 smart pins. DROOL!
How about just a P2 with full 1MB Hub RAM and 1MB flash (plus the speed increase).
Forget all that...I say bring back the Commodore 64 and let's live like it's 1983 again. Demand goes down which means cost goes down.
:-D
P.S., I'm serious about the Commodore 64 part.
A better way would be to group them - say one hub ram per 8 cores - and have a "superhub" shared by all groups.
Yes, eye-watering numbers, however there is also movement at the 'trailing edge'.
This recent news, gave a test run of parts, for a much more modest $50k, in 65nm - effectively a mask FPGA.
http://forums.parallax.com/discussion/164828/more-shuttle-choices
"The Shuttle is for the FG65L-5 device, which is a 65nm technology up to 600,000 usable gates, up to 900kbit of memory, 4 PLLs/DCMs with maximum of 242 user IOs."
The challenge then is, how much Prop can actually fit into that FG65L-5 ?
A company is doing microcoded 8086 and 8051, in impressively small LE counts - low 300's ..
http://www.embedded.com/electronics-blogs/max-unleashed-and-unfettered/4442147/Micro-sequencer-based-8051-core-requires-only-312-FPGA-LUTs
http://www.eetimes.com/author.asp?section_id=216&doc_id=1328967
Even if you could sell 10 million units, the NRE cost, alone, would be $50 per unit. You'd better be selling it for $100. Those metrics could work in an iPhone, I suppose. It's just nuts how much money that is. It's surreal.
And they sell many many millions of units worldwide.
Imagine a chip similar to a P1 in a tiny QFN20 package with inbuilt xtal osc.
The die would be really tiny, so lots on the wafer!!!
Maybe the chip cost would be <$0.10, sell for $1.
Could you sell 500M chips ???
But the real problem is getting the ~$250 funding
Best be sure you have a winning chip!!!
I recently read that Apple is at 16nm for iPhone and moving to 10nm for iPad. They may move to 7nm in 2018 and 5nm is likely to be after 2020. Even selling 100+ million units per year does not make the cost and risk viable right now.
I suspect the electromigration issue is being veiled by continuous replacements, due to upgrading.
10nm is being finalized right now for first deliveries in 2017. 7nm will have first trial runs in 2017. No has even finalized designs for 5nm yet.
I'm used to the idea that if I build a thing with chips those chips are going to still be working long after I am not.
Now it seems acceptable that they can fizz out just after the warranty period.
That might be fine for phones and such that only last a few months anyway because the screens get busted. Have you ever seen an iPhone without a cracked screen?
But what about people building infrastructure that is supposed to work reliably for thirty years?
Is it realistic to expect that the whole world upgrades its entire infrastructure every two years or so?
Meanwhile, space probes and jet airplane flight controllers will not be using 7nm. Ever.
There are parts of the Prop2 that have power densities of 7kW/cm2, like in the resistor sections of the 75-ohm DACs, where 36mW is being dissipated across a small area. Even a little op-amp circuit consuming 250uA can have a very high power density. You just need to spread those sections out from each other. Meanwhile, they cause steep temperature gradients on the die, which concentric layout can compensate for.
It seems that 10-year-device-lifespan electro-migration guidelines are in use for advanced processes these days. The automated chip design tools push up against these guidelines. You could tell the tools to work more conservatively, but I get the feeling nobody does that, because they want the performance.
I often have trouble with my internet connection, being a line-of-site microwave system, and on a few occasions their tech guy asked me if my router was over 3 years old, as it might need replacing. Kind of like how tires wear out, to him. They do deal with a lot of people and their routers.
These days, it isn't even good on motor vehicles. The days of the cars expiring due to rust or motor/gearbox exiration in a few years are long gone.
I have a 28 year old Toyota Supra that I have had since new. Still runs fine.
Also a Toyota Prado 4WD, 19 years old, 400,000km on the clock, and still runs fine.
As for car batteries, you can buy with 2 year or 3 year warranty. The good old car battery technology is well understood. They just make one of the return bars the right thickness so that it lasts precisely 2 or 3 years + a tiny bit. Built in replacement requirement.
Why did it take so long to change from making the old filament light bulbs? Because they wear out. Then the US changes the rules, and voila, they can make a better filament light bulb. In Oz, we banned the filament bulbs so we switched to fluorescent bulbs. Great for the environment, not! They have a small amount of mercury in them so they should not be dumped in the rubbish - but most people don't even know this.
I don't see the problem, with the housing prices of lately, you have to work like 500 years to pay the mortgage !