Giant wafer
System
Posts: 45
in Propeller 2
This discussion was created from comments split from: P2 Eval Board Owners, I need your help!.
Comments
Imagine the wafer yields on this thing ?! - no mention of how they cool it, but there are what look like mounting holes in the die !
https://www.edacafe.com/nbc/articles/1/1693023/Cerebras-Systems-Unveils-Industrys-First-Trillion-Transistor-Chip
"Optimized for AI work, the Cerebras Wafer Scale Engine (WSE) is a single chip that contains more than 1.2 trillion transistors and is 46,225 square millimeters. The WSE is 56.7 times larger than the largest graphics processing unit which measures 815 square millimeters and 21.1 billion transistors1. The WSE also contains 3,000 times more high speed, on-chip memory, and has 10,000 times more memory bandwidth."
That monster makes the 'large' P2 die, look tiny...
What wafer size? Seems to me that thing is a wafer, basically.
Yields may be expressed in terms of functioning processing units.
80 percent chips, 90, 95, 100.
Can we get that in a 40 pin DIP?
215mm x 215mm = 46224 sq mm.
So 21.5cm x 21.5cm or about 8.46in x 8.46in
(Sorry, I had to)
One wafer, one processing unit. Run what works on it, and make a lot of them. Sheesh.
So this is over half of the wafer!!!
Maybe they should just make a round chip
And what about power? Need a KW power supply? And it’s own air conditioner?
Wonder what the speed is? Maybe 4-5GHz?
Will it run Arduino
Or with 18GB it will run 4 copies of Windoze 10 at once
Begs the question of why not just make it a round package and and not trim the wafer at all. Then they could use the whole 70,000 mm2.
Probably get another 100,000-150,000 cores and 4-7GB
Any free engineering samples???
I'm guessing they decided to make it square, externally at least, for everyone's sanity.
Those large holes you see on the wafer are the I/O - you just hand solder wires to them
On the giant wafer chip I suspect the reason they squared the wafer is that there is an X-Y addressing scheme for the cores which doesn't deal well with the missing corners if they were to try to make use of the chords they discarded.
Done!
I often wondered what the maximum die size was. Now we know... the whole wafer which is currently 300mm dia AFAIK. Thats ~70,000mm2.
So, using OnSemi 200mm wafers at 180nm, that’s ~31,000mm2. This gives ~47B transistors in ~142mm square, or ~70B transistors on the whole wafer.
Or another way, the P2 is 8.5mm sq = ~72mm2, so on a whole wafer we could fit ~430 P2’s or 3,440 P2 cores with 215MB Hub Ram. Of course it doesn’t quite scale like this, but it puts it in some perspective.
The P2 fits 32M transistors in 72.3 mm2, for a density of 32M/72.3 = 443k transistors per mm2.
The density difference is 14.3M/443k = 32.3x. The HUGE chip process size must be about 180nm/sqrt(32.3) = 31.6nm. It's probably a 28nm process, which TSMC offers.
https://www.businesswire.com/news/home/20190819005148/en/
If that HUGE chip fits 1T transistors into 46,000 mm2, then it fits 1T/46k = 21.7M transistors per mm2.
The P2 fits 32M transistors in 72.3 mm2, for a density of 32M/72.3 = 443k transistors per mm2.
The density difference is 21.7M/443k = 49x. The HUGE chip process size must be about 180nm/sqrt(49) = 25.7nm.
We know it's a 16nm process, though.
I’m drooling to think we could have a P2 with say 256 cores, each with 64KB private cog/lut, and a giant central hub of banked 256 x 1MB. Even at 180nm and 350MHz it would be awesome 😎
I’d even settle for a list of failed cores and hub ram blocks.
What would i do with it ? Dunno 🤷♂️
I've been reading about a project that was wanting to do the very same thing (Wafer-scale) since 1999 as a Smalltalk processor.
The project is called SiliconSqueak. Merik Voswinkel and Jecel Assumpcao Jr. are the guys behind it.
A quick google search found a few documents including a PHD dissertation all by the same guys.
The project might prove an interesting read :-)
https://www.researchgate.net/project/Agora-Wafer-Scale-Integration
Dec 30th 2016
https://www.researchgate.net/publication/312029203_Plurion_an_Object_Oriented_Microprocessor
Jan 2005
http://www.merlintec.com/download/jecel_phd_deposited.pdf
May 2015
http://www.merlintec.com/
The Brazilian research company working on SiliconSqueak.
J
it's 1.2 Trillion... time to redo the math again.
the article says
as they probably need a lot of space for connectivity as well ...