[simne]

Because very large share of market now are datacenters. Difference from desktop is dramatic - for desktop really acceptable very simple chips with bad energy efficiency, but DCs already deal with extremely high power consumption, as they typically "compress" so much consumption in one rack, that constantly working near to physical constraints.

[scotty79]

You can't make desktop computer 4 times larger but there's very little preventing you form putting 4 racks where you had 1 before. If the floor space is the expensive part of data center then probably some incentives are misaligned.

[simne]

For about price of land and connectivity - in large city land price begin on few millions dollars per square kilometer, and usage of cable channels could cost from 50$ per meter (easy could be 200$/m).

Plus, space arrange could last years.

Heat dissipation in range of megawatts could be just prohibited by local regulations.

So, space in large cities is very serious problem, and for business it is usually easier to "compress" as much computing power as possible in one rack.

[IAmBroom]

> in large city land price begin on few millions dollars per square kilometer,

There's little need to put large datacenters in downtown Chicago and Manhattan.

[sbarre]

Bigger chips = more distance to cover for your electrons = more power required = more generated heat = slower throughput for your data.

Surely you don't believe that the entire chip industry had not thought of "wait what if we just make the chips bigger".

[moffkalast]

AMD hiding Threadripper behind their back: Uh yeah what a terrible idea, we definitely didn't actually do that. Making a CPU that's twice the size, how ridiculous would that be right?!

[sbarre]

You found an exception to my two-line generalization. Congrats!

[fuzzfactor]

The main reason to reduce feature size since quite some time has been to make more money per wafer, and faster.

Same reason that so much work was put into increasing wafer diameter over the decades.

More chips per wafer means a lot.

Much more than for performance sake.

[simne]

You cannot place dc anywhere, in large cities space is extremely constrained, and land is extremely expensive.

Also big problem - connectivity - you cannot place DC where it cannot be connected to power grid and to very powerful network.

So yes, DC floor space is severely limited.

And the third issue - last decades, rack servers dissipate extremely large amounts of heat, I hear numbers up to tens Kilowatts per rack, which is just hard to dissipate with air cooling (as example, all IBM Power servers have option of liquid cooling, but this is totally different price range).

[moffkalast]

That's the AI hype narrative, but aren't server CPUs only like 25% of the total market? That's tiny compared to consumer volume, though revenue is likely on par given the higher cost per unit.

[simne]

> aren't server CPUs only like 25% of the total market?

Yes and no. If just formally calculate, yes, servers are small market volumes. But, they are much less constrained financially, than private person, so from same fab one could earn much more money if sell to server market, than if sell to consumer market.

[zozbot234]

I don't think that's correct, server chips aren't really "more expensive" than consumer chips when you correctly account for performance. Older-gen server chips have comparable performance to new top-of-the-line consumer chips and sell for a similar price. Newer-gen server chips in turn are priced at a premium over the current value of the older-gen, to account for their higher performance. The lower financial constraints don't enter into it all that much.

[adrian_b]

For many years until about a decade ago (more precisely until the launch of the Intel Skylake Server processors) the server CPUs had a performance per dollar comparable to desktop CPUs so the expensive server CPUs were expensive because of their higher performance.

But since then the prices of server CPUs have ballooned and now their performance per dollar is many times worse than for desktop CPUs. Server CPUs have very good performance per watt, but the same performance per watt is achieved with desktop CPUs by underclocking them.

The only advantage of server CPUs is that they aggregate in a single socket the equivalent of many desktop CPUs, including not only the aggregate number of cores, but also the aggregate number of memory channels and the aggregate number of PCIe lanes. Thus a server computer becomes equivalent with a cluster of desktop computers that would be interconnected by network interfaces much faster than the typically available Ethernet links.

While for embarrassingly parallel tasks a server computer will cost many times more than a cluster of desktop computers with the same performance, it will have a much less disadvantage or it might even have a better performance/cost ratio for tasks with a lot of interprocess/interthread communication, where the tight coupling between the many cores hosted by the same socket ensures a lower latency and a higher throughput for such communication.

The owners of datacenters are willing to pay the much higher prices of modern server CPUs because the consolidation into a single server of multiple old servers brings economies in other components, due to less coolers, less power supplies, less racks, simpler maintenance and administration, etc.

While the prices of server CPUs at retail are huge, the biggest costumers, like cloud owners, can get very large discounts, so for them the difference in comparison with desktop CPUs is not as great as for SMEs and individuals. The large discounts that Intel was forced to accept during the last few years, to avoid losing too much of the market to AMD, were the cause why Intel's server CPU division has lost many billions of $.