On times with CPUs with so many cores, and with amazing single thread performance even at 3 GHz, overclocking is not what is used to be. E.g. in times when from a 33 MHz CPU you could get 21% speed-up running at 40MHz, or from 300 to 450 MHz on some budged CPUs (+50%, Celeron 300A with 128KB L2 cache), it was way more interesting (even knowing that the overclock could reduce the life of the components, because of electromigration, and some components running above their design limits).
I was under the impression that electromigration is a much more serious problem with modern processors due to the higher current densities—was it also a problem with older processors?
Not really. As long as you stayed within the thermal envelope, you were fine. I built hundreds of machines from 1995-2005, overclocking most, and I never had a single CPU failure that wasn't just a DOA chip.
Older chips had worse precision and control over overall process (polishing, chemical purity etc). This is a rather niche level knowledge, you would have to ask some ISTFA regular about it.
In my experience at least, the olde fears about shortening the life of the chip are non-issue in practice.
I've run overclocked PC's exclusively for about 17 years and never had a single failure, even though I've been running folding@home for most of those. By the time the chip is 4 yrs old it's time for a replacement for practical reasons.
My current system is a cheap Biostar board running a gen-1 i5 750 that I got second hand. The board dates back to 2009 I think. It's overclocked to around 3.6 ghz with turbo still enabled. About time for a replacement but until a year ago it benchmarked at the same level as $200 chips.
Overclocking is definitely still worth it if you want a fast pc for cheap
Yeh - the impact from overclocking isn't quite what it once was. It's nice that nowadays we're not left wanting for CPU speed even for most entry level processors. But I fondly remember feeling pretty excited by the "AXIA" AMD Athlons and pushing them to some pretty amazing results with fairly standard cooling (1.2/1.33GHz -> 1.66GHz and above, iirc)
I got a hand-me-down Athlon XP 2500+ around 2004-ish. I had read what you could do with it, so I did: bumped it up to 3200+ specs. It ran like that for years, until I saw the motherboard capacitors leaking and moved on.
Oh yeah. When I was a kid, my parents bought me an Athlon XP 2500+… overclocked to 3200+ and ran like that for many, many years. Eventually I gave that PC away to someone. I kinda regret that decision now :D
Its not the case of Overclocking no longer giving you expected speedup, its INTEL directly FIGHTING with overclockers and blocking every avenue that would enable you to gain performance for free.
Nowadays Overclocking is just another premium market segment being milked by Intel with K/Z series chips.
Every single Intel chipset is 'in theory' capable of overclocking (multiplier change, Turbo). Its all blocked on the Bios/OEM agreement[1]/microcode(!)[2] level. Sometimes OEMs grow bold and ignore it (Asrock pioneering H81 overclocking), sometimes they get spanked by Intel and quickly release 'updated' bios removing OC.
OC in itself works great. You an buy $72 3.2GHz G3258 haswell, OC to 4.3GHz (stock cooler, or ~4.5Ghz on custom) and enjoy >30% speed improvement. Especially great for single threaded games(written in python!) like World of Tanks where your oced $72 cpu performs like $300 4790K.
These firmwares disable power management completely, so that the microcode cannot even read the BCLK frequency and detect overclocking. As a side effect, you cannot use integrated graphics, and AVX instructions become slow — even slightly slower than at stock frequency (need power management to ramp up the AVX parts of the chip to full speed). But if you're not encoding video all day, you can save some money by buying a 6400 instead of a 6600K.
Not just overclockers, Intel has been gouging everyone ever since AMD stopped producing a competitive CPU (so basically since Core 2 or 2006.. man it's been that long?)
Here's hoping for Ryzen. Not only has Intel locked up the consumer segment but the server one as well. $200 for an E5-2407v2 is ludicrous, I'm really hoping AMD pulls a rabbit out of their hat because we need some real competition in the x86 space again.
The first time I built an Abit BP6 with dual Celeron 300's (overclocked to 450, naturally) I felt like I was getting a supercomputer for cheap. The performance increase going to dual-core was massive, and getting to use the cheap Celerons instead of the crazy-expensive Pentium II's paid for the rest of the system all on its own. The late 90's/early 2000's were exciting times for system builders.
Now? Meh. I still build my own, because you can spec out exactly what you want. But there's no longer any compelling performance reason to do anything beyond stock cooling and stock clocks.
I've got a 25% overclock on my i5 2600k going right now using the stock air cooler that's pretty rad. And I'm not an outlier, it's normal for these chips. The Haswell "devil's canyon" model was reportedly also a great overclocking chip.
These ~30% overclocks are still out there if you're willing to look for them. (But I believe that 50% celeron was some kind of binning issue caused by marketing).
The bigger issue these days is that GPU performance matters so much more for gaming. I've had the same CPU/motherboard for 5 years but I'm on my 3rd GPU.
50% Celeron OC lasted almost all the way to Coppermines. I was building PCs at the time specializing in cheap performance. Intel BX + 300A@450 2v, later 366@550 up to 700@1000 after modding slot1 adapters for 1.8-1.9V. Switched to selling AMD when disaster called P4 came out.
Yeah, I've had a 2500k running at 4.6 stable on all cores on stock voltages since it came out. I'm only now feeling a bit of an upgrade itch. If Ryzen CPUs are under $500 for the top models I'll build my next system on their platform and get a 2nd GTX 1080 or upgrade when the Vega 11 / Navi stuff comes out.
But what's interesting to note is that the number of pipeline stages hasn't increased all that much over the years, so I wonder if the upper limit on clock speed (disregarding power dissipation) has increased mainly due to smaller process sizes.
My understanding is that a long pipeline is only a benefit if the CPU gets fast enough to consume from that pipeline. CPUs around 2005-2010 had long pipelines in anticipation of higher CPU speeds which never happened. More recently, Intel have gone back to simpler pipeline design and focused on other areas.
I hope they decide to sell it for around $300. I'm certainly not entertaining any prices north of that. I'm not interested in funding their mobile chip ambitions when I buy a desktop processor. Which is why I haven't purchased a chip from them since the first generation i7. Their prices and products have been ridiculous for too long.
Kaby Lake is not great for overclocking.
You literally have to cut this CPU in half, to make it usable for OC. Better thermal paste between silicon and heatsink, will reduce temperatures by 10°C.
I was not very precise. I actually meant heat-spreader instead of heat-sink.
Kaby Lake has bad thermal conductivity between silicone and heat-spreader (metal top of CPU). If you crack-open the CPU and replace thermal paste, it will reduce temperature under load significantly.
It is common problem for Intel CPUs since Ivy Bridge. i7 are affected. Extreme version is soldered and does not have this issue.
60 degrees is not much, not really reaching full potential of that chip.
I had i7 2600K at 5.0 GHzx. Water cooled, 90 degrees under full load. It worked for 5 years 24/7, until motherboard died last year, CPU still works.
It is simple to set higher frequency, but that does not mean CPU is actually faster. Power management is very tricky. Sometimes it will put idle instructions, or turbo boost would not work.
You have to run benchmark for several hours to verify CPU is actually faster.
He means the thermal paste between the chip itself and the metal top (aka heatspreader). The metal top is the part the user-supplied thermal paste goes onto, and the cooler block sticks onto.
There's also another thermal paste sandwich UNDER that that you'd have to do some very delicate surgery to modify.
TL;DR everyone expected an IPC increase with Kaby Lake and Intel brought more performance only via higher clocks. They're either having issues innovating, betting that Ryzen will not be all AMD is talking it up to be, they're a bit caught with their pants down, or some combination. IMO worst-case for them is they push Coffee Lake and Skylake-X out sooner and drop prices across the board (especially on HEDT >4 core stuff) to compete with AMD.
Then once enthusiasts have access to <$2k PCs with 8 cores and the power of 2x1080GTX we can realistically power 4K VR @120FPS. Having games to actually make people want VR would also really help, so far not many AAA titles to use with it.
So the single-thread, integer-only, no-branch performance (admittedly a niche, but an interesting one) is still a fair way behind what was possible with the Pentium 4 (which could clock up to 11GHz)? I get that these days performance/watt is what matters and losing 100-odd cycles when you mispredict a branch isn't great, but it's still sad to think we might never match the best-case performance of the older line.
Not for the single-threaded integer-arithmetic case, AIUI - ADD is 1 IPC (at least when there are data dependencies) and pretty much always will be. In cases where branching happens, modern processes can maybe do more speculative execution, and obviously going from non-SIMD to SIMD is a huge improvement in the cases where a new processor lets you do that, so there are plenty of real-world improvements, but I don't think to the core case. I'm not an expert by any means though.
"but it's still sad to think we might never match the best-case performance of the older line."
More a positive note to those of us sticking with older hardware in various use-cases for simplicity and fewer backdoors. Good to know the integer performance still rocks. ;)
47 comments
[ 2.0 ms ] story [ 106 ms ] threadI've run overclocked PC's exclusively for about 17 years and never had a single failure, even though I've been running folding@home for most of those. By the time the chip is 4 yrs old it's time for a replacement for practical reasons.
My current system is a cheap Biostar board running a gen-1 i5 750 that I got second hand. The board dates back to 2009 I think. It's overclocked to around 3.6 ghz with turbo still enabled. About time for a replacement but until a year ago it benchmarked at the same level as $200 chips.
Overclocking is definitely still worth it if you want a fast pc for cheap
Nowadays Overclocking is just another premium market segment being milked by Intel with K/Z series chips.
Every single Intel chipset is 'in theory' capable of overclocking (multiplier change, Turbo). Its all blocked on the Bios/OEM agreement[1]/microcode(!)[2] level. Sometimes OEMs grow bold and ignore it (Asrock pioneering H81 overclocking), sometimes they get spanked by Intel and quickly release 'updated' bios removing OC.
OC in itself works great. You an buy $72 3.2GHz G3258 haswell, OC to 4.3GHz (stock cooler, or ~4.5Ghz on custom) and enjoy >30% speed improvement. Especially great for single threaded games(written in python!) like World of Tanks where your oced $72 cpu performs like $300 4790K.
[1] http://www.digitaltrends.com/computing/asrock-killed-overclo...
[1] https://www.reddit.com/r/overclocking/comments/3ji7im/nonz_o...
[2] http://www.sevenforums.com/windows-updates-activation/373250...
These firmwares disable power management completely, so that the microcode cannot even read the BCLK frequency and detect overclocking. As a side effect, you cannot use integrated graphics, and AVX instructions become slow — even slightly slower than at stock frequency (need power management to ramp up the AVX parts of the chip to full speed). But if you're not encoding video all day, you can save some money by buying a 6400 instead of a 6600K.
Now? Meh. I still build my own, because you can spec out exactly what you want. But there's no longer any compelling performance reason to do anything beyond stock cooling and stock clocks.
I'm not sure about this but I think any overclocking people do these days are for the fun of the journey, not for reaching the destination.
These ~30% overclocks are still out there if you're willing to look for them. (But I believe that 50% celeron was some kind of binning issue caused by marketing).
The bigger issue these days is that GPU performance matters so much more for gaming. I've had the same CPU/motherboard for 5 years but I'm on my 3rd GPU.
http://www.tomshardware.com/news/amd-fx-8150-overclock-9ghz-...
But what's interesting to note is that the number of pipeline stages hasn't increased all that much over the years, so I wonder if the upper limit on clock speed (disregarding power dissipation) has increased mainly due to smaller process sizes.
But overclockable i3 Kaby Lake is interesting.
But interestingly, the absolute first place winner had all 8 cores enabled http://hwbot.org/submission/2615355_the_stilt_cpu_frequency_...
You obviously don't have to cut the cores in half for daily usable overclocking. My quad Skylake runs 4.5GHz with all four cores enabled.
Kaby Lake has bad thermal conductivity between silicone and heat-spreader (metal top of CPU). If you crack-open the CPU and replace thermal paste, it will reduce temperature under load significantly.
It is common problem for Intel CPUs since Ivy Bridge. i7 are affected. Extreme version is soldered and does not have this issue.
I thought this:
> You literally have to cut this CPU in half, to make it usable for OC.
referred to turning off the cores, not delidding :D
But you don't need to do either to "make it usable for OC". Just get a good cooler.
I had i7 2600K at 5.0 GHzx. Water cooled, 90 degrees under full load. It worked for 5 years 24/7, until motherboard died last year, CPU still works.
It is simple to set higher frequency, but that does not mean CPU is actually faster. Power management is very tricky. Sometimes it will put idle instructions, or turbo boost would not work.
You have to run benchmark for several hours to verify CPU is actually faster.
There's also another thermal paste sandwich UNDER that that you'd have to do some very delicate surgery to modify.
Their CPUs are good because there isn't anything better, but their commercial practices have been worst than cancer.
Then once enthusiasts have access to <$2k PCs with 8 cores and the power of 2x1080GTX we can realistically power 4K VR @120FPS. Having games to actually make people want VR would also really help, so far not many AAA titles to use with it.
More a positive note to those of us sticking with older hardware in various use-cases for simplicity and fewer backdoors. Good to know the integer performance still rocks. ;)
Only thing I found with Google, was Intel predicting 10Ghz CPU by 2011.
https://encrypted.google.com/search?hl=en&q=pentium%204%2011...
http://www.geek.com/chips/intel-predicts-10ghz-chips-by-2011...