Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

A CPU that large reminds me of the famous remark made by Grace Hopper about how light can move 30cm in one nano second, I guess theoretically meaning that CPU could have some kind of maximum size.

Of course since current CPU contains cores, it doesn't apply.



Clock distribution is already quite complex in big circuits. It should arrive at the same time everywhere but it does not. You need to add extra components to de-skew them.


Instead of placing the dies next to each other on a flat surface, the dies could be above each other, or the CPU made spherical, for optimally close distances. And RAM of the motherboard could also be in a circle or sphere around it.

Is there any other concern than cooling with this?


Cooling is pretty much the one and only concern, there's nothing stopping AMD from doing this with their zeppelin dies except for the simple fact the chips on the bottom of the stack would get roasted at these TDP's.


They could go back to the SECC days but then put heatsinks on both sides.


RAM of the motherboard could also be in a circle or sphere

Cray did this back in the day https://en.m.wikipedia.org/wiki/Cray-1


I'd imagine that a circle/sphere would add all sorts of manufacturing, engineering, and build (ie actual installation in the box) concerns.


IIRC, cooling is already the primary limiting factor on current designs. People have achieved some remarkable stable overclocks by using LN2 cooling.


Silicon microchannels can cool pretty much anything but they're expensive.


Well for one thing, light doesn't bend. There's something to be said for the flat rectangular shape.


There is no light used in a CPU (or any other kind of chip found on a PC motherboard). Grace Hopper was talking about the speed of light, which is roughly how fast a change in voltage will propagate. That can and does curve all around your CPU because the wires are not always routed in straight lines.


Really? I'm pretty sure Einstein would disagree.


If your cpu is dense enough to gravitationally bend light by any significant amount then you might have other problems.

But we bounce light off of electric fields all the time, so the a flat chip is not really required (even if you made an all-optical chip). As others have said, it's all about heat distribution.


[Brief network outage provoked this far-too-long answer ;) ]

I'm pretty sure anyone who's dealt with fibre optics would disagree.

> Einstein

This likely refers to the behaviour of a classical beam of light in General Relativity passing through vacuum near a massive object sourcing an exterior Schwarzschild metric (e.g. a non-rotating but otherwise typical star).

There are redundant degrees of freedom in such a system. Normally one would set a coordinate condition or fix a gauge such that the beam of light is moving across star-centred/star-fixed coordinates [1], and one would say that the path it takes is curved compared to a beam that started parallel with the "curved" beam at a great distance from the star, and which never gets very close to the star.[2] Thus, the star's mass curves the beam of light, or equivalently, the beam(s) of light pass through curved spacetime (more technically they each follow a null geodesic of the Schwarzschild solution), with the near-to-star region of spacetime being more curved than the far-from-star region of spacetime.

In Special Relativity, spacetime is flat, so in vacuum beams of light that start parallel will always remain parallel.

Of course, when one introduces non-vacuum media (including hollow optical fibres with vacuum cores) or even a mirror, then a classical beam of light can be made to follow very different paths than it would in free space.

I think in this context that to the extent that Einstein would focus more on the non-vacuum behaviour of light and -- if we are talking about modern optical computing -- not consider the behaviour of a classical beam of light but rather a set of photons interacting with matter in a way in which quantum mechanical behaviour is not just observable but outright relevant. This would be directly analogous to considering the behaviour of fundamental charges in semiconductors in modern electronic computing, as quantum mechanical behaviour is often un-ignorable.

- --

[1] One could choose any arbitrary coordinate condition. For instance (this is something we can do in GR that is very non-Special Relativity-like) one could fix a set of coordinates on the notional leading edge of a (non-eternal) beam of light and treat the curvature of spacetime as a force acting on the beam, increasing in magnitude near the star, and pointing towards the star. While this is wholly legitimate -- the covariant formulation of the physics is identical -- one would generally take the position that the gauge discussed above is preferable as the behaviour of the light is simpler to describe (there's no need to consider the restoration of the behaviour of the beam of light as it leaves the region near the star or to consider any (pseudo-) force fields).

[2] Strictly speaking this is the path through spacetime, however, one could certainly consider deviation in the spacelike non-g_tt directions, and quantify this with the geodesic deviation equation applied between the two originally-parallel beams. In any event, a curved path through spacetime is shorter than a non-curved path, which is markedly different from Euclidean geometry in which a curved path through space is longer than a non-curved path.[3]

[3] Compare a path of a pulse of light through an optical microchip in a lab here on Earth, propagating from one edge of the chip to another. In the lab one will measure the non-straight path through curved waveguides or reflecting off mirrors as longer -- and taking longer -- than a straight path. But that's because the spacetime in the lab is so very nearly flat that it can't presently be distinguished from flat on the length scale of a (parallel-to-the-floor) microchip.


With all the cores it would still apply to inter-core communication and memory sharing.


There more i read about computer internals, the more i am reminded of the hermetic saying "as above, so below".

The layout of a CPU, the layout of a computer internally, and the layout of a network of computers seems to be more and more the same.


It goes even deeper than that. They layout of the circuits at the transister/component level have to deal with this too. Take a look at a motherboard between the cpu and the memory slots, you'll see traces that make weird turns and then back onto them selves. All to make sure the signals end up arriving at the same time (well, as close as is manageable).


Yes, but you can buffer inter core communication, cores don't need to be in sync. I think. I'm not a CPU designer.




Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: