A mix-and-match chiplet marketplace for processor makers is still a long way off

As Moore's so-called Law continues to slow, many chipmakers are turning to advanced packaging and chiplet techniques to drive greater efficiencies and performance than what's possible with process shrinks alone. AMD's Instinct MI300 family of accelerators, which it showed at its datacenter and AI event in June, is just the …

[steelpillow]

"In many respects, the chip package is becoming a complete system unto itself. Rather than spreading discrete components like memory controllers, CPUs, and GPUs across a motherboard, they can all be packaged together and communicate over a low-power, high-bandwidth fabric."

Indeed. Historically you got some lumps of stuff and made your own components. Then people started selling ready-made components in a range of sizes - a whole component in a blob of resin - and all you had to do was fix them down and wire them together. Next came the printed circuit board, all ready wired up and ready to stick them on. In due course you just bought your circuit board ready-built and plugged it in. Then came microprocessors - a whole miniature circuit board in a blob of plastic. Anybody remember the Intel Pentium II? CPU and core memory ICs mounted on a PCB as a single plug-in-and-watch-it-fall-out package? It has taken over thirty years of headscratching and failed products to take that forward to the point where we can reliably throw away that PCB and separate modules, and just stick those chips together before encasing in plastic.

I suppose the next thing will be to do away with the LAN and routers and stuff, and plug every system into a single 5G box under the BOFH's desk. We will all interact with it via Metaverse contact lenses welded to our eyeballs. What could possibly go wrong.

[Ken Hagan]

Nah. It'll be built into the contact lens and get power from the temperature difference between the two sides of the lens. It'll also have near-UV and near-IR cameras built in.

What could go wrong? Well, the medical regulators might need to be swatted away before anyone can sell it, but how hard could that be? (Elon's already working on it.)

[steelpillow]

Sometimes I get the feeling he's been field-testing too many prototypes at once.

But hey, this cage fight against the Borg, maybe he's planning a Matrix-style upload of Kung Fu mastery. That would be worth watching.

[Anonymous Coward]

Bow down before the Masters of Yield

stitch together 47 chiplets

My gob is smacked, I bow down before the mastery of yield and process control that even makes this thinkable.

[Sparkus]

Sounds like a job for....

Infiniband. At the substrate level anyway..........

[PRR]

> "It's the biggest inflection point in semiconductors since the dawn of RTL and synthesis way back in the late '80s," semi-con biz Synopsys VP John Koeter said

I sure as heck remember RTL- the Crown 800 audiotape deck was full of them. RTL was a less compact less efficient predecessor to TTL and then N/PMOS CMOS, the stuff that made mass logic practical.

But WTF does he mean by "synthesis"?

I was wondering if he meant "LSI", many-many gates on one chip. But that was not RTL era or even really TTL era. There was IIL which was compact/cheap but I have not seen that in decades. Does he mean we write code and an algorithm translates that to silicon? Gate array?

I've been here since early 1960s but the only "synthesis" comes to mind is a way to generate multiple frequencies without multiple dedicated crystals.