Hacker News Reader

Comment by DeathArrow | Hacker News Reader

I don't know if overclocking pays off much these days. The chips are already taken close to limits with factory settings, so there's not much overclocking potential unless you use extreme cooling.

I fondly remember the Core 2 Duo days when I overclocked my CPU with more than 50% using the stock cooler. It's nice to 50% performance for free, particularly when you are poor.

I even remember overclocking an old laptop with K6-II (those days laptop CPUs were socketed) by around 15%.

View on HN

One argument used to be that there's always manufacturing tolerances that you can exploit to your advantage. However, those tolerances were in many cases optimized out of the process by better binning -- if a particular CPU reliably worked at a higher clock speed, it would simply be sold at a higher speed.

But also, whether a chip is actually stable at settings X or Y has become a lot harder to determine due to complexity: multicore architectures, frequency scaling and power management in general make it hard to predict how a chip will behave under real conditions. It has become unrealistic for a user to test all cores under all possible combinations of workloads. Consequently, and from anecdotal experience at least, many overclocked systems seem a lot less stable than their owners claim.

>But also, whether a chip is actually stable at settings X or Y has become a lot harder to determine due to complexity

There are many good software for stress testing CPUs and GPUs. Just run AIda64 or OCCP stress tests for a long time and you are good to go if no problems arise. If there are problems, you redo the settings and run the tests again.

I think that's naive, or let's say oversimplified. There are too many cases where systems pass these tests and then fail in real-world scenarios, sometimes quickly. There's simply no way that a specific set of software routines can guarantee that a processor behaves in spec.

Undervolting (removing the safety margin put in place for binning) and removing power limits is where the gains are today. Ryzen 5000 could get some pretty nice multicore gains by undervolting using PBO2.

The most basic "overclocking" these days is actually undervolting. The clock speed you can leave to the chip.

The undervolting makes the chip run cooler, which in turn means it can stay at higher frequencies longer.

You need higher voltage to have a chip (CPU, GPU, RAM) stable at a higher frequency. Higher voltage means more heat, so you need better cooling.

This actually depends on silicon quality. The mainboard manufacturers generally configure the OOB CPU voltages higher than they need to be in anticipation of handling crappy silicon.

If you have a higher binned CPU then you can reduce that voltage and still be stable at a given clock, the chip will then be able to boost higher because there will be more thermal headroom.

The days of creating a resistor bridge with a pencil to enable extra cores and altering the multiplier are long gone. I still remember pushing my duron with a massive aluminium block on top of it getting 100+mhz more out of it.

I remember when you could enable cores in a GPU just by flashing a different firmware.

Wasn't there a time when you could flash a GeForce to a Quadro and you'd lose some performance in gaming but you gain floating point precision for rendering? Or was this a rumor?

Overclocking these modern CPUs is usually just removing the powerlimit. They can clock themselves pretty close to the thermal limit anyway and little is gained in raw performance above that. But the Performance/Watt drops significantly when overclocking.

I have a 10850k cpu, with a HUGE Be Quiet heatsink. Any heavy CPU workload will draw 300+watt, without any overclocking, the cpu will downclock, but not throttle around 100 Celsius. Overclocking has never been more dead than it is today. The days with old Celeron 300 or Core 2 Duo days with 50% + overclock speed are long gone. Today you will most like hit max temperature and cpu throttling.

> the cpu will downclock, but not throttle around 100 Celsius.

Downclocking is literally throttling. But there's like another 10% performance that can be unlocked if you'd go full custom watercooling.

>But there's like another 10% performance that can be unlocked if you'd go full custom watercooling.

It's actually less than 10%. I'd say closer to around 5% - as a custom watercooling enthusiast. 2-3% gains going from a great dual tower air cooler to an All-in-One closed loop (but extra noise from the pump) and another 2-3% gains for a great CPU waterblock and thicker, denser radiators of custom watercooling (and least noise of all, assuming you choose an appropriate pump speed).

5% performance gains, and 500% price increase <:o)

fine with that. for 99% of users it's great that factory settings allow the chip to run basically at max potential without having to be an expert in the complex web of acronyms/timings/voltages for overclocking

Frankly this is aimed more at XMP, where vendors routinely bump IMC voltage to get those swanky binned gaming kits onto the QVL. A ddr4-4800 kit obviously isn’t going to work on stock voltage with a memory controller rated at 3200, the gap is covered by increasing VCCIO and VCCSA a lot, and this results in quite a few chips with failed memory controllers over time.

Also, on AMD there is an extremely widespread culture of overclocking the fabric which again, can result in degradation over time (progressively lower clocks at a given voltage, or instability at a fixed clock). People very much still believe in the fairy-tale of “24/7 safe” overclocking - obviously the manufacturer has every incentive to certify the chip as fast as it’ll go, they set the limit at the 24/7 safe point already, but the gains are so potent that people tell themselves whatever they need to hear.

https://www.amd.com/en/legal/claims/gaming-details.html

AMD and intel have been very very clear for a very long time that this is not covered by warranty (see AMD GD-106/GD-112) but people have low key continued to do it and warranty the CPUs when they fail, because there wasn’t a way for AMD and intel to tell. And to be fair - AMD and intel have often muddled the waters by using these numbers in their benchmarks etc. But there remains a general sense of “if you didn’t turn on XMP you didn’t do the build right” and “XMP can’t hurt anything as long as you don’t increase the voltages” despite this already being done automatically by the board.

XMP absolutely does cause damage and I killed a 9900k with no other overclocking simply by enabling XMP on a fast kit. But consumers still don’t get the memo, they think it is only a problem if you instantly blow up the chip like the 7000 series at launch. XMP is widely viewed as safe and even something you're supposed to be doing.

Electromigration happens very quickly at 7nm and 5nm nodes, and people are not used to it, and it’s low key also a thing on 14nm and 10nm class too. But it is already something that CPUs have to be designed around and designed to detect and monitor and tolerate as part of the boost algorithm (that stability margin is also headroom the boost algorithm wants to exploit), they devote an enormous amount of effort to electromigration already. And the mismatch between that public understanding of the problem and the reality is causing a lot of warranty claims. You can see from enterprise fleets that CPUs don’t normally fail, yet every enthusiast knows someone who had a chip die “even without overclocking! I just enabled XMP!”. Of the failed CPUs getting returned for warranty, probably a plurality are memory overclocking failures. And this is only going to continue to get worse over time - 2.5D and 3D packaging are both progressively even more sensitive due to thermals, and have tighter requirements for voltage stability since they need to talk to the other dies at very low voltages (especially across heterogeneous dies with different metal stacks/totally different processes etc).

https://semiengineering.com/3d-ic-reliability-degrades-with-...

https://semiengineering.com/on-chip-power-distribution-model...

Reply on news.ycombinator.com