Question Why did older CPUs not use more power / clock higher?

[beginner99]

Reading the Interview with Jim Keller I noticed this sentence:

which at the time we thought were huge. These were 300 square millimeters at 50 watts, which blew everybody's mind.

Yeah, I think 486s and below didn't even have heat sinks if I recall correctly.

But my question is why? Why weren't they designed to user more power and hence increase their computing capabilities? Was it process limitations? tool limitations in CPU design?

 

[IntelUser2000]

Intel could have reached 10 GHz in the P4 without TOO much trouble - they already had the integer pipeline half clocked and running at around what 3.8 GHz when they gave up on P4 I believe. I'm sure the plan was to expose the half clocked pipeline in future iterations, so that would make it a 7.6 GHz CPU. If they'd stuck with the P4 architecture a little longer they could have reached 10 GHz before long.

No CPU ever reached 9GHz. And only when you have a single core enabled, and with the long pipeline stage CPUs like Bulldozer and Netburst. And you need exotic liquid nitrogen or something of the sort to cool it with ridiculous voltages.

It was about the 5GHz level that needed water cooling and something even better. Nothing changed. I've seen a TH review that showed that it can do 5.4-5.5GHz on water cooling. All that happened was that coolers became several times larger and water cooling is much more common.

This is true regardless of ISA, process generation, and company and designs.

5.3GHz "Base" but 5.5GHz with water cooling! YIPPEE! No they just got better taking away overclocking headroom.

They probably abandoned the double-pumped ALU in Prescott because it was hindering clocks elsewhere. Also the ALU couldn't handle all operations.

 

[jamescox]

I remember a store tech asking me if I was over clocking because I had an aluminum heatsink that was a few inches high on an AMD k6-3. That is more like a chipset cooler by today’s standards (maybe not quite that tall). For a while, there was a drive to use massive amounts of copper; like pounds of copper hanging off your board. Then there was water cooling for the extreme. I always wanted to do a water cooling setup, but then heat pipes came along and water cooling became a lot less necessary. You often don’t see that much difference between a high end heat pipe based air cooler and water cooling. So there is a lot less of a reason to mess around with water cooling. Even the AIO seem to have limited life span compared to a good air cooler. Heat pipe based coolers seem like the reason that we can use such high power devices. I am not sure what you can go up to without heat pipes, at least without ridiculous noise levels. Servers often just use little heat sinks with fins, like my old K6, but they move ridiculous amounts of air over the heat sink and sound like a jet engine. My K6 just had a tiny little fan on top.

 

[jamescox]

Last I heard the current record for transistor switching speed is around 600 GHz, so they aren't THAT far away from terahertz. EUV was a couple process generations away for probably 15 years before it finally happened (and laughably, it wasn't Intel who did it as they still haven't sold a single chip made using EUV) 450 mm wafers were also about two process generations away at that time, but that will never happen.

Just because we have 5 GHz CPUs doesn't mean that's the speed transistors are switching at. Modern CPUs have something like 20 FO4 delays per cycle which implies we've had transistors switching at at least 100 GHz in our modern CPUs. Likely well in excess of that as you have to account for wire delay as well as FO4 delay in your worst case clock network.

Intel could have reached 10 GHz in the P4 without TOO much trouble - they already had the integer pipeline half clocked and running at around what 3.8 GHz when they gave up on P4 I believe. I'm sure the plan was to expose the half clocked pipeline in future iterations, so that would make it a 7.6 GHz CPU. If they'd stuck with the P4 architecture a little longer they could have reached 10 GHz before long.

The problem is that each cycle would not have been accomplishing all that much so the pipeline would be REALLY long, which is bad for branch delay and filling load/store delay slots so IPC would be terrible. It was also consuming a lot of power, at the time when consumers cared more and more about power usage in laptops and minimizing fan noise in desktops.

Reaching 10 GHz in a CPU that was power hungry and inappropriate for laptops, and wouldn't perform any better than 2.5 GHz PPro/PIII/"Core" architecture - and AMD being competitive at that time as well - would make continued sale of the P4 based on marketing MHz a difficult proposition. That's why they decided to hide clock rates behind model numbers going forward, so they wouldn't have to explain to consumers why the CPU in the latest PCs was clocked a lot slower than the one in last year's PCs after spending 20 years teaching consumers that more MHz = more performance.

I don’t think they could have reached 10 GHz. They thought they would be able to at the beginning, but that faded away. The power consumption with clock speed is non-linear. They would have been pulling hundreds of watts pretty quickly. Even with exotic cooling and the, large by modern standards, 217 mm2 die size, they would have hit power limitations. The last versions on a much smaller process were up to something like 120 watts or so. It wouldn’t have performed much better anyway, since it would just spend most of its time waiting on memory. Caches were small back then. The initial version with 256K L2 was not sufficient. It often did not outperform the P3 and only scaled to about 2 GHz. The following version with 512K was a lot better, but it was still horribly inefficient. When the power scaling stopped due to leakage (about 2006?), that pretty much put an end to inefficient architectures.