Hacker News Reader

Comment by Vt71fcAqt7 | Hacker News Reader

How do they combine the two seperate chips into one? Is there a performance/size cost? Could we have 1,000 chiplets per CPU and acheive >99% yield?

View on HN

There are a variety of 2.5D (chip next to chip) and 3D (chip on top of chip) packaging techniques. Each comes with different tradeoffs.

Here are a couple of articles with more detailed info:

https://semiwiki.com/semiconductor-manufacturers/tsmc/306329...

https://www.anandtech.com/show/16051/3dfabric-the-home-for-t...

https://www.eetimes.com/amd-tsmc-imec-show-their-chiplet-pla...

https://www.techpowerup.com/292256/amd-details-its-3d-v-cach...

They use an internal interconnect, similar to how a server which took multiple distinct CPUs would use a motherboard-level interconnect.

In addition to splitting the CPU or GPU into distinct units, you can also take other functions and use different processes for them. For example, in package I/O or L2 cache don't really see the same advantages for newer processes, so you can make these using more established (and cheaper/more available) processes.

You get most of the economic value from the first few divisions. Going from 80% yield to 90% yield shaves 14% off your cost/unit. Going from 95% to 99% only saves you 4%.

I'm not an expert at how they combine chips. Like I said for AMD, they also wanted their units to be composable w/ small number of chips, so they basically have a die w/ a few cores and different die w/ memory, and I believe they have a proprietary communication mesh for connecting them. I think there is some considerable signal/energy overhead to communicating between chips. The cost of masks and interconnects is probably high enough to make a high number of diverse chiplets inviable, but I wouldn't say it's impossible in the future.

Thanks.

>I think there is some considerable signal/energy overhead to communicating between chips.

I tried looking for some info on this but couldn't find any. Do you have any source I could read on this?

This is a fundamental effect because of the Shannon-Hartley theorem[1], which says your communication channel's capacity (i.e. bitrate) has a log dependency on the channel's Signal to Noise Ratio. In a practical wireline communication system, you have a transmitter with noise, a lossy interconnect, and a receiver with noise. As the interconnect gets longer, which is what happens when you go from on-die communication to die-to-die communication, your loss increases. This means you have to reduce your transmitter and/or receiver noise which requires using additional power.

Another way of looking at it is bit error rate [2].

I think you'll be hard-pressed to find concrete numbers since these designs are closely-guarded trade secrets. You might find some examples by searching for wireline transceiver / PHY papers on IEEE xplore, especially at ISSCC (conference) or in JSSCC (journal).

[1] https://en.wikipedia.org/wiki/Shannon%E2%80%93Hartley_theore...

[2] https://en.wikipedia.org/wiki/Bit_error_rate

Reply on news.ycombinator.com