Really interesting interview. I think the comments on "you wouldn't believe how many knobs there are for OC'ing" validates the idea Intel has always been aware of its heat problems.
It also lends credence that Intel (& AMD to a lesser extent) are quite dependent on high clocks (e.g., 4.8+ GHz) to achieve high performance. Unfortunately, high clocks are on a very flat part of the power vs perf curve. While this works on mobile workstations -> HPC, I'd hope after testing Apple's M1, Intel refocuses back on much wider designs and noticeably lower stock clocks. It's just whether Intel is able to do it well.
Intel, it wouldn't be a crime to reduce PL1 / PL2; we can make progress here, instead of assuming the standard ultra-thin laptop should always boost to over 50W (!). We used to think laptops were "15W", but the boost era has really increased the heat output of ultra-thin laptops to yester-decade's mobile workstations.
While Intel is more honest with Alder Lake, it's not lost on me that Intel CPUs are the absolute hottest on the market: far more than AMD & Apple.
Intel CPUs have Speedstep and Turbo boost technologies which allow the CPU to run at very low clocks and maintain performance at the tap. Every CPU does but Intel processors have specific set of C states which do the work amazingly on a laptop.
This whole focus of Laptop BGA trash should not be a part of OC at all. All those processors have poor silicon, locked BIOSes, garbage power limits and poorer cooling. They used to have individual heatsinks for CPU and GPU when this thin and light nonsense trend was not there. It helped longevity I have a 2014 machine which has both CPU and GPU socketed and it runs latest 2022 games because I changed my GPU (see MXM).
Apple processors also need higher boost clocks and same for others on ARM like Qcomm. Clockspeed is a requirement for all processors, ultra thin garbage does not have 50W PL2. You talking about U series which are ULV class 15W TDP class with 30W max and they have power limits too. Mainstream gaming processors have higher TDP 115 W is the max for 12700H. And it operates at 45W base. You want performance or portability ? Also longevity is best for socketed components because BGA balls get problems after a certain period of time due to heat cycling.
X86 has limitations in wider archs. Its very hard, limiting to impossible to make a functioning architechture as wide as Apples M1. ARM 64 is a much newer arch with heavy focus on multicore and wide decode.
It would be interesting, generally seeing his opinion about why Apple M1's are that much efficient on comparable power envelopes paralleled to other cpu/memory/chipset combinations (if meaningful on that level, because of variance of interfaces and thermal standards)?
Both company's cpus are quite performant even at lower frequencies. I mean that's where AMD gets its efficiency lead from compared to intel, since intel is dealing with fewer performance cores while AMD can stay at better frequency curves for all core work. The over the top boosting just prevents them from leaving single core performance on the table. If the silicon can handle it then you may as well use it. I agree that there should be more emphasis on lower PL1, though.
>Then, inside Intel, we're always tuning knobs, and adjusting things. We had control of the ratio - at one point, something called ‘re-marking’ was going on in the market, and it was not a good thing, as third parties would modify the label on the processor, and indicate that it was that 50 MHz instead of the 33 MHz. We needed a way to prevent that. So again, enabling a rock-solid, stable, reliable processor, people could count on us. So we'd lock the ratio, and at that point, before K SKU existed, the only form overclocking would have would be base clock overclocking.
IMHO, this example does kind of date this problem, too. No manufacturers are doing that today. AMD has unlocked all its CPUs without any such problems.
I like AMD's method: unlock every CPU and just use the motherboard (the far more relevant ingredient in OC quality) to lock / unlock overclocking.
Overclocking should be motherboard / chipset-dependent, not CPU SKU dependent. OEMs aren't buying B660 / Z690 chipsets for their high-volume PCs. Intel can leave H670 / H610 locked.
Good point about AMD not having problems with illegal branding of their processors. Today as you go up the stack you get more cores. You can "fake" that with software tweaks!
the moment overclocking became commercialized with special edition CPUs, any "value" reason for overclocking went out the door. Couple that with the fact that computers have not been bottlenecked by pure CPU performance for decades...
OC as a hobby to learn the technology is definitely an interesting path to fool. But it is a luxury at the top end of the scale.
Not really if you ask me. An Intel K CPU is a top bin, and that means it's the best of the silicon for the wafer, like a 10900K for instance. A 10850K is not a 10900K and never will be because I saw a 10850K which has very good VIDs but cannot break past 5.3GHz stable no matter what. While a 10900K can. Also a 10700K binning is not superior to 9900K either. Because CFL chip bin maxed at 9900K quality and CML at 10900K.
Even on mobile a Haswell 4700MQ is nothing like 4930MX, the latter is a top of the line i7 binned silicon which can run at pretty good Undervolt and Overvolt which achieves getting max out of the CPU.
I understand the old notion but that's way too old, at-least Intel respects Overclocking which is a major benefit to people who want to understand Hardware, which is how I learnt about CPUs and PC, even smartphones of old era, Qcomm and Samsung had bins too. Now it's useless because of complex scheduler EAS and other junk.
probably, also underclocking might get lower interest, if power cores could be hotswapped in/out and also efficiency cores could be sent to idling, if performance demand and load doesn't require all cores active (That interesting question is, where's the tradeoff between performance for all possible configurations (now) enabled and power/thermals/cooling power: bigLittle (on x64's), under/overclocking, core/cluster swap in/out, finer voltage level granularity, adjustable response times, on peripherals/chipset/direct cpu data transfer configuration (powersave vs. performance), <- (OS) platform independent client/user tool (bios, UEFI, hardware switches) ?)
Interesting interview! One question I missed and would have liked Dan Ragland's answer to is cooling, and what Intel does to make that as capable as they can. After all, cooling is absolutely vital for any attempt at overclocking. Specifically, how does Intel tries to make sure (or not) that their CPUs and sockets are optimized for good heat removal and transfer? The recent issues with LGA 1700 sockets and the processor clamps bending/warping the CPUs by as much as a full mm (!) aren't conducive to good heat transfer. The best water cooling solution and heat transfer paste or liquid can't do much with those gaps in place. I sometimes have the strong impression that the chip circuitry is designed by great and highly knowledgeable experts, but the cooling solutions and socket specifications by some first year Summer intern.
The problem is LGA1700 mounting ILM. Intel did not design it properly, the clamping force of the ILM and the reverse force of Backplate is not enough which is causing the CPU to bend at the center causing a concave shape making the TIM being major in the cycling of the CPU and overtime stress. Rectangular CPU and clamping force in horizontal center line.
This is why AMD's AM5 has better ILM mechanism which screws down the ILM properly with backplate with another set of screws.
Good interview, and an interesting comment. To be fair I think the socket spec is more cohesive than you're giving credit for here, but that doesn't mean it's perfect, and the biggest problem is the inconsistency of the overall socket-motherboard design.
I'm an embedded systems designer with an engineering background, so I inherently 'get' issues like this, and there are still some aspects to the LGA1700 mounting issues I've yet to see properly addressed by some of the enthusiast websites - or specialist OC forums.
One of them is distortion/flex in the motherboard PCB, which is far more pronounced on the Z690 series than I've seen on previous boards - my Z690 Hero was warped across three or four different areas when I got it out of the box, and is about half a millimeter thinner than the Z390 predecessor it replaced. I'm guessing the two-DIMM boards with even thinner PCBs might be even more negatively affected.
I've made solid aluminium cooling solution mounting backplates for years going back to Socket 478, if not earlier - and the one I made for my Z690 board was deliberately intended to contact the rear plate of the clamping mechanism, creating a fully structural stressed 'sandwich' effect around the CPU socket, taking the pressure off the mainboard's PCB (and solder) and preventing the board, socket and CPU bowing under pressure from the cooling solution. It also does some backside heatsinking as a bonus.
Coincidentally or otherwise, I've tried the CLM raising mod and it made no difference (well, it actually *raised* temps by a couple of degrees). Possibly because, with the solid backplate installed from day 1, my processor won't/can't bow and the cooling solution can generate good contact at lower overall clamp pressure.
The backplate also pre-emptively fixed some issues I'd have encountered had I used the OE mounting mechanism for my (Corsair) AIO too, which is another subject for another post.
If you haven't yet, Igor's Lab has taken a good look at Intel's LGA1700 mounting issues, and even what AMD is doing differently, really interesting reads.
I have had many low grade staffing houses from India spam my inbox with such thermal and cooling positions at Intel and I can confirm that “summer intern” is not far off the mark based on their job descriptions and wages.
I think it's very hard, more than all core OC. Which is actually the best part of Intel processors. Unlike AMD which is XFR PBO2 based dynamic clock behavior which you cannot control at all, I mean we can but it will lose that turbo XFR advantage.
if it comes to 'per core' overclocking, people should know where cores are located (what memory/gpu_vpu(_npu)/peripheral controllers are close) on a chiplet and having core-wise thermal sensors and datasheet-wise thermal conductivity numbers available (if not, that might be some statistics from OC labs for the customers?)
I used to overclock everything. Now I don't bother because there's virtually no headroom left for all-core OC these days, unless you're running sub-ambient. Additionally, there is often a regression in single-threaded performance when focusing on all-core OC since there are so few chips that can push all cores higher than what the out of box Turbo bins can achieve. I still watercool just to run the lowest temps and ensure there is no thermal throttling when Turbo kicks in, but it's admittedly not as fun as it used to be.
Well it was a good interview Ian. But I wish you asked for more on the Silicon Engineering and binning aspects.. Like how the degradation happens, how electromigration effects, plus how the K series is actually binned by Intel vs non K. How do they vary by the wafer positioning and etc because right now everyone believes that center of wafer is top bin and nobody knows how a batch works, a good VID CPU with golden silicon sample from a batch can have horrible CPUs, cut from the same wafer.
Also for the Halo Mobile, I do not know if you are aware. Intel had Xtreme Mobile series, a top of the line binned K series class SKUs, Intel Ivy Bridge had 3920XM and 3940XM they overclock to 4GHz without breaking sweat on a proper machine like M18x R2. Or the Haswell 4930MX and 4940MX, these are problematic due to FIVR heating up the CPU faster. Anyways the point is Intel had them. They had that on the BGA trash too like HQ is bare min locked bins suprisingly Intel had unlocked 2 bins on Haswell mobile like 4700MQ, 4800MQ, 4900MQ, but SKL 6700HQ and KBL 7700HQ are trash locked parts lol, truly a time when rPGA was in mobile and people upgraded from an i5 mobile to an i7 Extreme Mobile CPU. Shame how BGA trashed the whole market. Now they removed undervolting too because of SGX issue and even with 11th gen and 12th which do not have SGX they locked undervolting.
Plus I like how the man has Comet Lake, I bet it's a 10900K golden sample. It's the best CPU to have fun and OC. 11900K is horrible and it has poor IMC on top which is worse. 12900K is a first gen of big little so it's not worthy at all to own and run for long. X99 also a superb HEDT. Ring bus fast CPU with ton of PCIe lanes and fun to OC, I hope SPR HEDT arrives and reinvigorates the OC and DIY space again, AMD killed the SRTX40 prematurely.
On a final note ..
Intel publishes documentation for their CPUs which is really a treasure trove of Information I wish he touched on that aspect. You know why I didn't chose AMD ? They do not publish anything. An Intel 10th gen for eg has 2.0v maximum for VCore, that's a hard lock we know. For any of the VCCIO and VCCSA we know limits are better at 1.2 to 1.3v max for high IMC as it OCes the Uncore and Memory. All these are there in the documentation. For AMD's IOD issues many say up this voltage and that or do this Core Recycler etc and try to do that Curve Optimizer on reddit and etc to get over that DRAM and IF Instability lottery on even stock. But we do not know what's per spec and it's a headache.
AMD pushed their Silicon to absolute limits if you ask me with the OCN thread spanning 80 pages when Zen 3 launched with higher boost clocks and tons of RMA, they tamed it later. Intel started doing it with Rocket Lake the TVB of that CPU almost killed all OC. Alder Lake is also in similar boat, plus ADL is having very very high Heat density more than Intel RKL which makes them run very hot on past 250W, like a 10900K at 300W = Alder Lake at 250W. Makes sense because ADL die is very small vs CML and RKL is too massive but also a backport .. anyways that's all.
I've had to ban two people so far today for making crude attacks on Mr. Ragland.
Let it be known that any further comments along those lines will also earn swift bans. This behavior is unacceptable and has no place in the AT comments section.
Great interview, and an insightful look into a part of Intel that doesn't have much of a public facing presence. Really interesting that they're involved with all aspects of chip design, but it does make sense given how highly binned Intel's top end CPUs have become.
Looking forward to more coverage from your trip here to Hillsboro.
Data rate on DDR memory is arguably semantically accurate as MHz as the term is used to denote that an *insert event here* happens every 1 second, if your physical transistor memory clock rate is 100 Hz and you perform two "transfers" per clock (even semantically speaking) then you are still doing 200 transfers per second, which would mean the event (transfer) is happening 200 Hz. I think this is overblown honestly.
A clock cycle is a periodic function, measured in Hz - a Hz isn't simply 'per second', it's 'cycles per second'. It's all about complete, never-ending, repetitive cycles. Not only is clock-edge data transfer a reverse cycle (e.g. the second edge is opposite to the first), it's not a Hz measurement by virtue of data transfer not being cyclical. Sometimes data is transferred, sometimes it isn't. The data rate in this case gives the peak transfers per second, minus any additional encoding overhead that might occur in other schemes. Beyond that, we can simply look at the units:
Cycles per second (C/s) * transfers per cycle (T/C) = Transfers per second (T/s)
It's all covered in the video, including Linus' argument. Linus is now using MT/s.
Plus, if the company considered applying XMP profiles warranty-breaking act that's another reason no one should pay attention to Intel and overclocking — except, of course, for employees of companies that use overclocking to push product. They can play around as usual, for vacuous hype.
Locking the multiplier made sense when the difference between SKUs was the clock speed. Today we're looking at core counts, L3 size and clock speed is not as much of an issue. Not sure locking the multiplier is actually serving a purpose today aside from forcing the upsell to the K SKUs. It's not like someone can overclock an i5 to an i7 or an i7 to an i9, the active components in the chips are different, parts are physically disabled.
Clarification, we're looking at the core counts and L3 size, the clock speeds aren't as much of an issue as they were in the past, a Pentium III went from 450MHz to 1.3GHz across 3 cores, today we really only have a couple of SKUs, a cheap i5 and an overclockable i5, a cheap i7 and an overclockable i7, etc...
very interesting inverview (thx), maybe it could be extended with some overview, where on a chip/chiplet power consumption occurs at what amount (not the expert details, being internals, but some rough overview would be a really nice picture to have: cores (power/efficient), cache, controllers, peripherals, gpu/vpu/tpu&npu, memory (x64?) )?
Another item while reading: What's the definition (and profiling standard) for comparing cpu's performances (internal or not vendor related)? What benchmark tools are utilised for overview?
I really hope Intel and AMD create CPU SKUS specifically for buyers that intend to liquid cool their CPUs. It would make things so much simpler for a meaningful amount of high-end CPU users and boutique shops.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
41 Comments
Back to Article
linuxgeex - Monday, January 24, 2022 - link
Definitely prefer MT/s!ikjadoon - Monday, January 24, 2022 - link
Really interesting interview. I think the comments on "you wouldn't believe how many knobs there are for OC'ing" validates the idea Intel has always been aware of its heat problems.It also lends credence that Intel (& AMD to a lesser extent) are quite dependent on high clocks (e.g., 4.8+ GHz) to achieve high performance. Unfortunately, high clocks are on a very flat part of the power vs perf curve. While this works on mobile workstations -> HPC, I'd hope after testing Apple's M1, Intel refocuses back on much wider designs and noticeably lower stock clocks. It's just whether Intel is able to do it well.
Intel, it wouldn't be a crime to reduce PL1 / PL2; we can make progress here, instead of assuming the standard ultra-thin laptop should always boost to over 50W (!). We used to think laptops were "15W", but the boost era has really increased the heat output of ultra-thin laptops to yester-decade's mobile workstations.
While Intel is more honest with Alder Lake, it's not lost on me that Intel CPUs are the absolute hottest on the market: far more than AMD & Apple.
Silver5urfer - Monday, January 24, 2022 - link
Intel CPUs have Speedstep and Turbo boost technologies which allow the CPU to run at very low clocks and maintain performance at the tap. Every CPU does but Intel processors have specific set of C states which do the work amazingly on a laptop.This whole focus of Laptop BGA trash should not be a part of OC at all. All those processors have poor silicon, locked BIOSes, garbage power limits and poorer cooling. They used to have individual heatsinks for CPU and GPU when this thin and light nonsense trend was not there. It helped longevity I have a 2014 machine which has both CPU and GPU socketed and it runs latest 2022 games because I changed my GPU (see MXM).
Apple processors also need higher boost clocks and same for others on ARM like Qcomm. Clockspeed is a requirement for all processors, ultra thin garbage does not have 50W PL2. You talking about U series which are ULV class 15W TDP class with 30W max and they have power limits too. Mainstream gaming processors have higher TDP 115 W is the max for 12700H. And it operates at 45W base. You want performance or portability ? Also longevity is best for socketed components because BGA balls get problems after a certain period of time due to heat cycling.
Kurosaki - Tuesday, January 25, 2022 - link
X86 has limitations in wider archs. Its very hard, limiting to impossible to make a functioning architechture as wide as Apples M1. ARM 64 is a much newer arch with heavy focus on multicore and wide decode.back2future - Thursday, January 27, 2022 - link
It would be interesting, generally seeing his opinion about why Apple M1's are that much efficient on comparable power envelopes paralleled to other cpu/memory/chipset combinations (if meaningful on that level, because of variance of interfaces and thermal standards)?whatthe123 - Monday, January 24, 2022 - link
Both company's cpus are quite performant even at lower frequencies. I mean that's where AMD gets its efficiency lead from compared to intel, since intel is dealing with fewer performance cores while AMD can stay at better frequency curves for all core work. The over the top boosting just prevents them from leaving single core performance on the table. If the silicon can handle it then you may as well use it. I agree that there should be more emphasis on lower PL1, though.ikjadoon - Monday, January 24, 2022 - link
>Then, inside Intel, we're always tuning knobs, and adjusting things. We had control of the ratio - at one point, something called ‘re-marking’ was going on in the market, and it was not a good thing, as third parties would modify the label on the processor, and indicate that it was that 50 MHz instead of the 33 MHz. We needed a way to prevent that. So again, enabling a rock-solid, stable, reliable processor, people could count on us. So we'd lock the ratio, and at that point, before K SKU existed, the only form overclocking would have would be base clock overclocking.IMHO, this example does kind of date this problem, too. No manufacturers are doing that today. AMD has unlocked all its CPUs without any such problems.
I like AMD's method: unlock every CPU and just use the motherboard (the far more relevant ingredient in OC quality) to lock / unlock overclocking.
Overclocking should be motherboard / chipset-dependent, not CPU SKU dependent. OEMs aren't buying B660 / Z690 chipsets for their high-volume PCs. Intel can leave H670 / H610 locked.
Hulk - Monday, January 24, 2022 - link
Good point about AMD not having problems with illegal branding of their processors. Today as you go up the stack you get more cores. You can "fake" that with software tweaks!jmke - Monday, January 24, 2022 - link
the moment overclocking became commercialized with special edition CPUs, any "value" reason for overclocking went out the door. Couple that with the fact that computers have not been bottlenecked by pure CPU performance for decades...OC as a hobby to learn the technology is definitely an interesting path to fool. But it is a luxury at the top end of the scale.
jmke - Monday, January 24, 2022 - link
"path to *follow"Silver5urfer - Monday, January 24, 2022 - link
Not really if you ask me. An Intel K CPU is a top bin, and that means it's the best of the silicon for the wafer, like a 10900K for instance. A 10850K is not a 10900K and never will be because I saw a 10850K which has very good VIDs but cannot break past 5.3GHz stable no matter what. While a 10900K can. Also a 10700K binning is not superior to 9900K either. Because CFL chip bin maxed at 9900K quality and CML at 10900K.Even on mobile a Haswell 4700MQ is nothing like 4930MX, the latter is a top of the line i7 binned silicon which can run at pretty good Undervolt and Overvolt which achieves getting max out of the CPU.
I understand the old notion but that's way too old, at-least Intel respects Overclocking which is a major benefit to people who want to understand Hardware, which is how I learnt about CPUs and PC, even smartphones of old era, Qcomm and Samsung had bins too. Now it's useless because of complex scheduler EAS and other junk.
back2future - Thursday, January 27, 2022 - link
probably, also underclocking might get lower interest, if power cores could be hotswapped in/out and also efficiency cores could be sent to idling, if performance demand and load doesn't require all cores active (That interesting question is, where's the tradeoff between performance for all possible configurations (now) enabled and power/thermals/cooling power: bigLittle (on x64's), under/overclocking, core/cluster swap in/out, finer voltage level granularity, adjustable response times, on peripherals/chipset/direct cpu data transfer configuration (powersave vs. performance), <- (OS) platform independent client/user tool (bios, UEFI, hardware switches) ?)eastcoast_pete - Monday, January 24, 2022 - link
Interesting interview! One question I missed and would have liked Dan Ragland's answer to is cooling, and what Intel does to make that as capable as they can. After all, cooling is absolutely vital for any attempt at overclocking. Specifically, how does Intel tries to make sure (or not) that their CPUs and sockets are optimized for good heat removal and transfer? The recent issues with LGA 1700 sockets and the processor clamps bending/warping the CPUs by as much as a full mm (!) aren't conducive to good heat transfer. The best water cooling solution and heat transfer paste or liquid can't do much with those gaps in place. I sometimes have the strong impression that the chip circuitry is designed by great and highly knowledgeable experts, but the cooling solutions and socket specifications by some first year Summer intern.Silver5urfer - Monday, January 24, 2022 - link
The problem is LGA1700 mounting ILM. Intel did not design it properly, the clamping force of the ILM and the reverse force of Backplate is not enough which is causing the CPU to bend at the center causing a concave shape making the TIM being major in the cycling of the CPU and overtime stress. Rectangular CPU and clamping force in horizontal center line.This is why AMD's AM5 has better ILM mechanism which screws down the ILM properly with backplate with another set of screws.
Short_Circuit - Monday, January 24, 2022 - link
Good interview, and an interesting comment. To be fair I think the socket spec is more cohesive than you're giving credit for here, but that doesn't mean it's perfect, and the biggest problem is the inconsistency of the overall socket-motherboard design.I'm an embedded systems designer with an engineering background, so I inherently 'get' issues like this, and there are still some aspects to the LGA1700 mounting issues I've yet to see properly addressed by some of the enthusiast websites - or specialist OC forums.
One of them is distortion/flex in the motherboard PCB, which is far more pronounced on the Z690 series than I've seen on previous boards - my Z690 Hero was warped across three or four different areas when I got it out of the box, and is about half a millimeter thinner than the Z390 predecessor it replaced. I'm guessing the two-DIMM boards with even thinner PCBs might be even more negatively affected.
I've made solid aluminium cooling solution mounting backplates for years going back to Socket 478, if not earlier - and the one I made for my Z690 board was deliberately intended to contact the rear plate of the clamping mechanism, creating a fully structural stressed 'sandwich' effect around the CPU socket, taking the pressure off the mainboard's PCB (and solder) and preventing the board, socket and CPU bowing under pressure from the cooling solution. It also does some backside heatsinking as a bonus.
Coincidentally or otherwise, I've tried the CLM raising mod and it made no difference (well, it actually *raised* temps by a couple of degrees). Possibly because, with the solid backplate installed from day 1, my processor won't/can't bow and the cooling solution can generate good contact at lower overall clamp pressure.
The backplate also pre-emptively fixed some issues I'd have encountered had I used the OE mounting mechanism for my (Corsair) AIO too, which is another subject for another post.
hansmuff - Tuesday, January 25, 2022 - link
If you haven't yet, Igor's Lab has taken a good look at Intel's LGA1700 mounting issues, and even what AMD is doing differently, really interesting reads.willupowers - Thursday, January 27, 2022 - link
I have had many low grade staffing houses from India spam my inbox with such thermal and cooling positions at Intel and I can confirm that “summer intern” is not far off the mark based on their job descriptions and wages.Jorgp2 - Monday, January 24, 2022 - link
I do feel like most people don't appreciate per core overlcockingSilver5urfer - Monday, January 24, 2022 - link
I think it's very hard, more than all core OC. Which is actually the best part of Intel processors. Unlike AMD which is XFR PBO2 based dynamic clock behavior which you cannot control at all, I mean we can but it will lose that turbo XFR advantage.back2future - Thursday, January 27, 2022 - link
if it comes to 'per core' overclocking, people should know where cores are located (what memory/gpu_vpu(_npu)/peripheral controllers are close) on a chiplet and having core-wise thermal sensors and datasheet-wise thermal conductivity numbers available (if not, that might be some statistics from OC labs for the customers?)techguymaxc - Monday, January 24, 2022 - link
I used to overclock everything. Now I don't bother because there's virtually no headroom left for all-core OC these days, unless you're running sub-ambient. Additionally, there is often a regression in single-threaded performance when focusing on all-core OC since there are so few chips that can push all cores higher than what the out of box Turbo bins can achieve. I still watercool just to run the lowest temps and ensure there is no thermal throttling when Turbo kicks in, but it's admittedly not as fun as it used to be.Silver5urfer - Monday, January 24, 2022 - link
Well it was a good interview Ian. But I wish you asked for more on the Silicon Engineering and binning aspects.. Like how the degradation happens, how electromigration effects, plus how the K series is actually binned by Intel vs non K. How do they vary by the wafer positioning and etc because right now everyone believes that center of wafer is top bin and nobody knows how a batch works, a good VID CPU with golden silicon sample from a batch can have horrible CPUs, cut from the same wafer.Also for the Halo Mobile, I do not know if you are aware. Intel had Xtreme Mobile series, a top of the line binned K series class SKUs, Intel Ivy Bridge had 3920XM and 3940XM they overclock to 4GHz without breaking sweat on a proper machine like M18x R2. Or the Haswell 4930MX and 4940MX, these are problematic due to FIVR heating up the CPU faster. Anyways the point is Intel had them. They had that on the BGA trash too like HQ is bare min locked bins suprisingly Intel had unlocked 2 bins on Haswell mobile like 4700MQ, 4800MQ, 4900MQ, but SKL 6700HQ and KBL 7700HQ are trash locked parts lol, truly a time when rPGA was in mobile and people upgraded from an i5 mobile to an i7 Extreme Mobile CPU. Shame how BGA trashed the whole market. Now they removed undervolting too because of SGX issue and even with 11th gen and 12th which do not have SGX they locked undervolting.
Plus I like how the man has Comet Lake, I bet it's a 10900K golden sample. It's the best CPU to have fun and OC. 11900K is horrible and it has poor IMC on top which is worse. 12900K is a first gen of big little so it's not worthy at all to own and run for long. X99 also a superb HEDT. Ring bus fast CPU with ton of PCIe lanes and fun to OC, I hope SPR HEDT arrives and reinvigorates the OC and DIY space again, AMD killed the SRTX40 prematurely.
On a final note ..
Intel publishes documentation for their CPUs which is really a treasure trove of Information I wish he touched on that aspect. You know why I didn't chose AMD ? They do not publish anything. An Intel 10th gen for eg has 2.0v maximum for VCore, that's a hard lock we know. For any of the VCCIO and VCCSA we know limits are better at 1.2 to 1.3v max for high IMC as it OCes the Uncore and Memory. All these are there in the documentation. For AMD's IOD issues many say up this voltage and that or do this Core Recycler etc and try to do that Curve Optimizer on reddit and etc to get over that DRAM and IF Instability lottery on even stock. But we do not know what's per spec and it's a headache.
AMD pushed their Silicon to absolute limits if you ask me with the OCN thread spanning 80 pages when Zen 3 launched with higher boost clocks and tons of RMA, they tamed it later. Intel started doing it with Rocket Lake the TVB of that CPU almost killed all OC. Alder Lake is also in similar boat, plus ADL is having very very high Heat density more than Intel RKL which makes them run very hot on past 250W, like a 10900K at 300W = Alder Lake at 250W. Makes sense because ADL die is very small vs CML and RKL is too massive but also a backport .. anyways that's all.
Ryan Smith - Monday, January 24, 2022 - link
I've had to ban two people so far today for making crude attacks on Mr. Ragland.Let it be known that any further comments along those lines will also earn swift bans. This behavior is unacceptable and has no place in the AT comments section.
m53 - Tuesday, January 25, 2022 - link
Thank you for keeping it civil.GeoffreyA - Tuesday, January 25, 2022 - link
Thanks, Ryan.thestryker - Monday, January 24, 2022 - link
Great interview, and an insightful look into a part of Intel that doesn't have much of a public facing presence. Really interesting that they're involved with all aspects of chip design, but it does make sense given how highly binned Intel's top end CPUs have become.Looking forward to more coverage from your trip here to Hillsboro.
AshlayW - Tuesday, January 25, 2022 - link
About MHz vs MT/s:Data rate on DDR memory is arguably semantically accurate as MHz as the term is used to denote that an *insert event here* happens every 1 second, if your physical transistor memory clock rate is 100 Hz and you perform two "transfers" per clock (even semantically speaking) then you are still doing 200 transfers per second, which would mean the event (transfer) is happening 200 Hz. I think this is overblown honestly.
Ian Cutress - Tuesday, January 25, 2022 - link
Slightly incorrect. https://www.youtube.com/watch?v=5fZO77I-6CgA clock cycle is a periodic function, measured in Hz - a Hz isn't simply 'per second', it's 'cycles per second'. It's all about complete, never-ending, repetitive cycles. Not only is clock-edge data transfer a reverse cycle (e.g. the second edge is opposite to the first), it's not a Hz measurement by virtue of data transfer not being cyclical. Sometimes data is transferred, sometimes it isn't. The data rate in this case gives the peak transfers per second, minus any additional encoding overhead that might occur in other schemes. Beyond that, we can simply look at the units:
Cycles per second (C/s) * transfers per cycle (T/C) = Transfers per second (T/s)
It's all covered in the video, including Linus' argument. Linus is now using MT/s.
Oxford Guy - Tuesday, January 25, 2022 - link
Intel killed the overclocking warranty program and wants us to care about overclocking, in 2022?Droll.
Oxford Guy - Tuesday, January 25, 2022 - link
Plus, if the company considered applying XMP profiles warranty-breaking act that's another reason no one should pay attention to Intel and overclocking — except, of course, for employees of companies that use overclocking to push product. They can play around as usual, for vacuous hype.0ldman79 - Wednesday, January 26, 2022 - link
Excellent interview.Locking the multiplier made sense when the difference between SKUs was the clock speed. Today we're looking at core counts, L3 size and clock speed is not as much of an issue.
Not sure locking the multiplier is actually serving a purpose today aside from forcing the upsell to the K SKUs. It's not like someone can overclock an i5 to an i7 or an i7 to an i9, the active components in the chips are different, parts are physically disabled.
0ldman79 - Wednesday, January 26, 2022 - link
Clarification, we're looking at the core counts and L3 size, the clock speeds aren't as much of an issue as they were in the past, a Pentium III went from 450MHz to 1.3GHz across 3 cores, today we really only have a couple of SKUs, a cheap i5 and an overclockable i5, a cheap i7 and an overclockable i7, etc...back2future - Thursday, January 27, 2022 - link
very interesting inverview (thx), maybe it could be extended with some overview, where on a chip/chiplet power consumption occurs at what amount (not the expert details, being internals, but some rough overview would be a really nice picture to have: cores (power/efficient), cache, controllers, peripherals, gpu/vpu/tpu&npu, memory (x64?) )?back2future - Thursday, January 27, 2022 - link
Another item while reading: What's the definition (and profiling standard) for comparing cpu's performances (internal or not vendor related)? What benchmark tools are utilised for overview?Gracie C. Wilson - Saturday, January 29, 2022 - link
https://en.wikipedia.org/">AnchorsGracie C. Wilson - Saturday, January 29, 2022 - link
<a href="https://en.wikipedia.org/">Anchors</a&g...
Gracie C. Wilson - Saturday, January 29, 2022 - link
[Anchors](https://devdojo.com)
Azekk - Monday, January 31, 2022 - link
Where is the fun? My friend told me that this place is entertaining and interesting, I guess he lied to me. https://vengeio.ussuperdave643 - Wednesday, February 2, 2022 - link
Wonderful insights, thank you Dr. Ian!lilkwarrior - Tuesday, February 22, 2022 - link
I really hope Intel and AMD create CPU SKUS specifically for buyers that intend to liquid cool their CPUs. It would make things so much simpler for a meaningful amount of high-end CPU users and boutique shops.I'm glad this interview asked about it.