I imagine you only use the high performance connect for things that need it. One for cost, but two for reliability. Ideally you don’t rip out all the compute engines and high performance memory because the usb controller died. The more chips you pack onto the device, the more chances there are for one of them to go bad and cause you to throw the whole thing away. Perhaps you build in redundancy and just ignore failures, but that increases complexity, leading to a higher rate of errors, etc.

In particular, bypassing the CoWoS heartburn-indigestion, looks to me like the healthier way to go about HPC/AI/ML gastronomy (if it works right … seeing how the 2D MCM image does look visually like 2.5D EMIB). This could be a winning tech from Eliyan.

I’m not sure whether to be amazed or frightened. I remember the day hawking 3rd party modem cards and convincing folks to go with brands that had the Lucent 56k chip as opposed to the Rockwell chip. Then suddenly….poof….supply dried up as everything started going into the motherboard. It’s like I’m the silicon version of a black hole creation . You first see the parent star swell and swell until it explodes, collapses into itself, only to be reborn into something even larger.

I think so.

So the socket will TRULY be the motherboard. I’m actually starting to believe this to be true . I mean, extrapolate this out. Does a desktop AI PC in 2035 ( whatever that means in 2035 much less today ) have a motherboard with nothing but one huge socket and traces on the motherboard for various physical connections like USB, HDMI, and Ethernet? Could you see a NUC sized mini PC where the motherboard is only as large as the socket itself as there is no need for any external connectors as everything is wireless in and out? Or another extreme. What happens to the Raspberry Pi when the socket becomes bigger than today’s motherboard? Will that become the ultimate expression of an SBC when quite literally the socket is the single board and the computer simultaneously?

They have done their PHY in 5 nanometer and 3 nanometer processes as far as I know. And they have one large scale customer, someone not quite a hyperscaler but more than a typical service provider.

But made of chiplets with interconnects.