TR Forums

So I've got a 7940x running at stock in an x299 Gigabyte Aorus 7 motherboard with 32 gigs of quad-channel DDR4-2666. On a lark I decided to run Cinebench R20 to see how it'd do... and I'm disappointed. I'd expect this Intel 14 core chip to at least pull even with a 1950X Threadripper, which pulled a score of 6670 cb, but even after disabling AVX and AVX-512 offsets, I'm only pulling 5282 cb. While I understand that it's not wise to tie too much to a theoretical benchmark, I'm more concerned that this is also affecting my BOINC and sustained video encoding performance.

Admittedly I've been lazy and have allowed the motherboard to handle voltage management. Is it possible that Gigabyte's "Auto" voltage management is using too much power by default, causing the chip to run hotter and throttle more readily? I'm using a be quiet! Dark Rock 4, so that shouldn't be a source of problems. Thanks for reading.

Have you check and see if the CPU is throttling or turbo boost is falling back under the stress test?

Krogoth wrote:

Have you check and see if the CPU is throttling or turbo boost is falling back under the stress test?

It's got to be - Task Manager indicates the clocks are drooping pretty appreciably. The benchmark was slightly slower with the AVX offset set to defaults, and the single core bench absolutely cratered before; with the offset the score was 239 cb, and with offset disabled that leaped to 438 cb.

Stock speeds? Doesn't seem particularly out of the ballpark given the clock deficit and core deficit to a 1950x. A 7940X at 4.5 GHz all-core can crank out about 8k points, so if you derate linearly to the base clocks, you're pretty spot-on.

Waco wrote:

Stock speeds? Doesn't seem particularly out of the ballpark given the clock deficit and core deficit to a 1950x. A 7940X at 4.5 GHz all-core can crank out about 8k points, so if you derate linearly to the base clocks, you're pretty spot-on.

I wondered if I was overreacting - under sustained x264 encodes with the AVX offset it'd sometimes dip down to 2.9 GHz across all cores, but given a guaranteed 3.1 GHz floor that makes sense. And it's not like the chip's been delidded. Still, this might be a good time to start playing with undervolting. Thank you.

Delid that bad boy. My 7960x runs at 4GHz on all cores with no AVX offset and around 1.03V. This chip scores 8039 in Cinebench R20 at those settings.

Yeah, delid if you're willing to take that risk.

The Skylake X chips are stupidly power-hungry and barely manage their base clock if you provide only the "official" 165W of cooling.

The 7960X at stock speeds is 75W hungrier than the Threadripper 1950X when measured using power draw at the wall, and the threadripper is rated at 180W to the Intel 165W. Now you can see what a joke that 165W TDP is, but that's only half the problem. Intel took such a ridiculously expensive and power-hungry chip and then used TIM instead of solder. Note that the 7940X is lower-binned silicon, it's likely to be leakier and more defective than the 7960X that is commonly reviewed. The complete lack of any reputable 7940X reviews makes me suspicious that Intel know that it's a power-draw and cooling nightmare, and therefore never sent any samples out to reviewers. I'm simply extrapolating based on the fact that lower-end models end up with the lower-binned dies.

You'll get 3.1GHz all-core speeds if you use a 240mm radiator, and it's unlikely you'll get a whole lot more, silicon lottery permitting.
You need to delid, overclock, and get a good quality triple radiator on that thing, otherwise it's a huge waste of potential.

Meanwhile, Ryzen 1950X will run using a 92mm air cooler in an mATX case at 3.7GHz+ all-core without any problems. I know because I have 16 of them in a farm used for VRay ray tracing purposes.

Chrispy_ wrote:

Yeah, delid if you're willing to take that risk.

The Skylake X chips are stupidly power-hungry and barely manage their base clock if you provide only the "official" 165W of cooling.

They 7960X at stock speeds is 75W hungrier than the Threadripper 1950X when measured using power draw at the wall, and the threadripper is rated at 180W to the Intel 165W. Now you can see what a joke that 165W TDP is, but that's only half the problem. Intel took such a ridiculously expensive and power-hungry chip and then used TIM instead of solder.

You'll get 3.1GHz all-core speeds if you use a 240mm radiator, and now a whole lot more. You need to delid, overclock, and get a 360 ro 280mm radiator on that thing, otherwise it's a huge waste of potential.

Meanwhile, Ryzen 1950X will run using a 92mm air cooler in an mATX case at 3.7GHz+ all-core without any problems. I know because I have 16 of them in a farm used for VRay ray tracing purposes.

It is because Intel never intended to operate their HCC Skylake-X SKUs at such high clockspeeds.

Chrispy_ wrote:

Yeah, delid if you're willing to take that risk.

The Skylake X chips are stupidly power-hungry and barely manage their base clock if you provide only the "official" 165W of cooling.

The 7960X at stock speeds is 75W hungrier than the Threadripper 1950X when measured using power draw at the wall, and the threadripper is rated at 180W to the Intel 165W. Now you can see what a joke that 165W TDP is, but that's only half the problem. Intel took such a ridiculously expensive and power-hungry chip and then used TIM instead of solder. Note that the 7940X is lower-binned silicon, it's likely to be leakier and more defective than the 7960X that is commonly reviewed. The complete lack of any reputable 7940X reviews makes me suspicious that Intel know that it's a power-draw and cooling nightmare, and therefore never sent any samples out to reviewers. I'm simply extrapolating based on the fact that lower-end models end up with the lower-binned dies.

You'll get 3.1GHz all-core speeds if you use a 240mm radiator, and it's unlikely you'll get a whole lot more, silicon lottery permitting.
You need to delid, overclock, and get a good quality triple radiator on that thing, otherwise it's a huge waste of potential.

Meanwhile, Ryzen 1950X will run using a 92mm air cooler in an mATX case at 3.7GHz+ all-core without any problems. I know because I have 16 of them in a farm used for VRay ray tracing purposes.

It's one thing to be power hungry and not offer top tier performance, it's an entirely different matter when your products consume more power whilst providing more performance, as is the case with Intel's HCC chips in the vast majority of workloads.

My 4GHz 7960x system with 32GB DDR4 3000, ~7x 120mm fans, D5 pump, 15x 7200RPM 4-8TB NAS Hard Drives, 1x 2.5" SATA SSD, 2x m.2 NVME SSDs, and a GTX 1660 pulls down no more than 435W from the wall when running Handbrake batch encodes with Plex Media Server running in the background actively streaming with hardware transcoding enabled. That's on an 850W Seasonic Focus Plus Gold which is 91.76% efficient in Hardocp's testing at 50% load: https://www.hardocp.com/article/2017/07/11/seasonic_focus_plus_gold_850w_power_supply_review/4. That works out to almost exactly 400W. Handbrake uses AVX also so that's about as demanding a scenario as you're going to find. If you like I could give you a sample file and my Handbrake settings so we could see just how fast Threadripper really is.

There's a reason AMD is doubling FPU width in Zen 2 afterall.

techguy wrote:

https://www.hardocp.com/article/2017/07/11/seasonic_focus_plus_gold_850w_power_supply_review/4. That works out to almost exactly 400W. Handbrake uses AVX also so that's about as demanding a scenario as you're going to find. If you like I could give you a sample file and my Handbrake settings so we could see just how fast Threadripper really is.

There's a reason AMD is doubling FPU width in Zen 2 afterall.

https://techgage.com/article/battle-of- ... st-cpus/4/

I don't know, it doesn't seem like there's a massive delta there.

Waco wrote:

techguy wrote:

https://www.hardocp.com/article/2017/07/11/seasonic_focus_plus_gold_850w_power_supply_review/4. That works out to almost exactly 400W. Handbrake uses AVX also so that's about as demanding a scenario as you're going to find. If you like I could give you a sample file and my Handbrake settings so we could see just how fast Threadripper really is.

There's a reason AMD is doubling FPU width in Zen 2 afterall.

https://techgage.com/article/battle-of- ... st-cpus/4/

I don't know, it doesn't seem like there's a massive delta there.

General tech reviewers don't know how to extract the most performance from Handbrake.

See this (and subsequent posts of mine in this thread): viewtopic.php?f=30&t=121911&p=1403534#p1403480

Feel free to copy my settings and try it for yourself on your 1950x system. When I first built this machine I went with a 7900x after reading every single review on the net comparing Skylake-X with Threadripper, specifically in Handbrake. Almost none of those reviews showed much of a difference, virtually no scaling beyond 8-10 cores. I found one at the time, I will try to find it again. I recall the review using drastically different settings than just about every other review out there though, which made me realize that most of the workloads being tested were not pushing these CPUs in the right way.

techguy wrote:

Feel free to copy my settings and try it for yourself on your 1950x system. When I first built this machine I went with a 7900x after reading every single review on the net comparing Skylake-X with Threadripper, specifically in Handbrake. Almost none of those reviews showed much of a difference, virtually no scaling beyond 8-10 cores. I found one at the time, I will try to find it again. I recall the review using drastically different settings than just about every other review out there though, which made me realize that most of the workloads being tested were not pushing these CPUs in the right way.

I'd have to virtualize on my NAS to do so, but it definitely handles many streams of 1080p transcoding without a hiccup via Plex. 4K transcoding on the fly is actually functional as well.

I'd be interesting in seeing what the tuning can do on each, though!

Waco wrote:

techguy wrote:

Feel free to copy my settings and try it for yourself on your 1950x system. When I first built this machine I went with a 7900x after reading every single review on the net comparing Skylake-X with Threadripper, specifically in Handbrake. Almost none of those reviews showed much of a difference, virtually no scaling beyond 8-10 cores. I found one at the time, I will try to find it again. I recall the review using drastically different settings than just about every other review out there though, which made me realize that most of the workloads being tested were not pushing these CPUs in the right way.

I'd have to virtualize on my NAS to do so, but it definitely handles many streams of 1080p transcoding without a hiccup via Plex. 4K transcoding on the fly is actually functional as well.

I'd be interesting in seeing what the tuning can do on each, though!

It really comes down to what you're asking the CPU to do. For me, I only use Handbrake on DVDs and other "low quality" (non-HD) source material, in an effort to get the most quality out of my media. Blu-rays and 4k Blu-rays I just mux to MKV with MKVToolNix. There's no transcoding in this process, the datastream is ripped from the M2TS files and dropped in an MKV container. The result is indistinguishable from dropping the disc in a player (I've done frame-by-frame comparisons).

Anyway, with lower quality source material the CPUs can really stretch their legs. It's the same concept as testing games at absurdly low resolutions to see which CPU is truly the fastest. Not practical in the real world, but meaningful to demonstrate the absolute capability of a given architecture.

Ah, I guess that's not how I use it. Maximal streams at a time and/or maximal speed for transcoding 4K. The 1950x is pretty capable for that.

Waco wrote:

Ah, I guess that's not how I use it. Maximal streams at a time and/or maximal speed for transcoding 4K. The 1950x is pretty capable for that.

CPU transcoding 4k needs an incredible amount of resources. I picked up a GTX 1660 recently to fill in that role so I can leave the CPU entirely to Handbrake, background tasks, and audio transcoding for Plex clients that have native support for the video codec, but not the audio (which happens pretty frequently in my experience).

How many clients have you been able to stream to while transcoding?

techguy wrote:

It's one thing to be power hungry and not offer top tier performance, it's an entirely different matter when your products consume more power whilst providing more performance, as is the case with Intel's HCC chips in the vast majority of workloads.
<anecdotes>

Well, yes. I agree with you but I have no idea why you quoted me; I didn't mention performance even once in my post. I'm simply saying that the 165W TDP of the 7940X is a joke of a lie. If provided with a '165W' cooler it will throttle and fail to meet base clocks.

We all know it's a 250W chip by simply looking at multiple stock-speed power draw results on the countless reviews around the web. If you want more than the rated all-core base speeds you are going to need to provide maximum cooling, no expense spared - as your 400W example proves.

In fairness to common sense, if you're going to buy a $1300 chip and expensive supporting board, you really do need to spend a proportional amount on cooling. Intel's crazy 28 core consumer part needed a 1770W Tetrafluoroethane refrigerant system, and Anandtech's 7980XE pulls down a kilowatt when all the taps are opened. Don't get me wrong, their 7980XE is ludicriously fast, but from a performance-per-Watt argument, you could genuinely build four 2990WX boxes in that power envelope - and Intel aren't going to hold a candle to 128C/256T of Threadrippers.

Concupiscence wrote:

I'm using a be quiet! Dark Rock 4, so that shouldn't be a source of problems.

See, the Dark Rock 4 is one of the better air coolers for your typical Ryzen/CoffeeLake processors but it is only rated to 200W by BeQuiet! themselves. If you want more performance, grab something like an Alphacool Eisbaer LT, EK Predator, or H150i.

techguy wrote:

How many clients have you been able to stream to while transcoding?

To be honest I haven't kept track since switching over from my Haswell Xeon quad build. I started by streaming to 5 different devices at once, 1080p reasonable bitrate sources, and when that didn't choke I considered it a success (I rarely have more than 5 streams active at a time). It can also transcode moderate bitrate 4K sources while also streaming a couple 1080p sources without much buffering.

Not very scientific, I know, but I didn't feel the need to justify its existence given that we won the CPU for free from the Intel 8086K / AMD 1950X swap deal.

Chrispy_ wrote:

techguy wrote:

It's one thing to be power hungry and not offer top tier performance, it's an entirely different matter when your products consume more power whilst providing more performance, as is the case with Intel's HCC chips in the vast majority of workloads.
<anecdotes>

Well, yes. I agree with you but I have no idea why you quoted me; I didn't mention performance even once in my post. I'm simply saying that the 165W TDP of the 7940X is a joke of a lie. If provided with a '165W' cooler it will throttle and fail to meet base clocks.

Can you provide evidence of this? I ran my 7900x at stock clocks and stock TIM under the IHS and it never got all the way down to base clock, or exceeded the 140W TDP as best my calculations can tell (usage under full load a full 100W less than the 7960x @ 4GHz). The lowest all-core speed I ever saw was 3.6GHz (base clock is 3.3) and 95% of the time the chip sat at 4.0GHz.

Chrispy_ wrote:

We all know it's a 250W chip by simply looking at multiple stock-speed power draw results on the countless reviews around the web. If you want more than the rated all-core base speeds you are going to need to provide maximum cooling, no expense spared - as your 400W example proves.

The purpose of including the rest of my system's specs is to demonstrate that not all of the power consumption can be attributed to the CPU. Unless you think 15 hard drives, 3 SSDs, 7 120mm fans, a GTX 1660, and a D5 water pump run on fairy dust. Also, I forgot to mention my Intel PCI-e RAID HBA. Keep in mind that this CPU running at 4GHz is easily consuming 50% more power than stock also. All this to say, Skylake-X is certainly not the most power-efficient architecture in history, but it does deliver the most performance in nearly all workloads. It's up to customers to decide it the extra performance justifies the additional power consumption.

Chrispy_ wrote:

In fairness to common sense, if you're going to buy a $1300 chip and expensive supporting board, you really do need to spend a proportional amount on cooling. Intel's crazy 28 core consumer part needed a 1770W Tetrafluoroethane refrigerant system, and Anandtech's 7980XE pulls down a kilowatt when all the taps are opened. Don't get me wrong, their 7980XE is ludicriously fast, but from a performance-per-Watt argument, you could genuinely build four 2990WX boxes in that power envelope - and Intel aren't going to hold a candle to 128C/256T of Threadrippers.

4x 2990wx systems *may* pull down the same power from the wall as a massively-overclocked i9 system, but they're also going to cost a whole lot more. Skylake-X and Zen+ Threadripper/Epyc serve different markets. Skylake-X excels at lightly-threaded and medium-thread-count workloads, Threadripper/Epyc excel at (most) highly-threaded workloads. There's no single solution that is best for all workloads. For me, Skylake-X wins more often than it loses, usually due to the use of AVX or the need for more clock speed.

I would be happy to consider a 64-core Threadripper 3 for a mere $2000 though

techguy wrote:

Can you provide evidence of this? I ran my 7900x at stock clocks and stock TIM under the IHS and it never got all the way down to base clock, or exceeded the 165W TDP as best my calculations can tell (usage under full load a full 100W less than the 7960x @ 4GHz).

Nope, I absolutely cannot; I do not have a 7940X and neither do you.

Your experience of a 7900X is irrelevant to a 7940X. Although the architecture is the same generation, the 7900X is an LCC silicon design and the best-binned variant of it in a (4x3)-2MC array Three months later, Intel rolled out their HCC silicon design with a (5x4)-2MC array for the 7980XE. It has a different transistor density to your 7900X, different power draw, different support 'uncore' and unlike the 7900X and 7980XE, the 7940X is not the best bin of flawless silicon - it is the second-worst bin of the Skylake X HCC - either partially defective, too leaky to make the cut, or possibly both.

Without owning or testing one myself, I'm simply drawing on the numerous tests of the various chips published by reputable review sites like PCPer, Anand, TechPowerup who ran these chips through comprehensive power-use testing at stock and overclocked speeds using multiple different power-draw testing methods at launch.

techguy wrote:

The purpose of including the rest of my system's specs is to demonstrate that not all of the power consumption can be attributed to the CPU

Which is why I'm quoting specific sites and power draws. Anand's 1KW overclock was a minimum build of just the CPU, RAM, board and a test drive. In your system it's too hard to know what percentage of your power draw is the processor because there are too many other variables. The best you can do is measure the difference between idle and peak when running a CPU-only benchmark or encode.

techguy wrote:

Can you provide evidence of this? I ran my 7900x at stock clocks and stock TIM under the IHS and it never got all the way down to base clock, or exceeded the 140W TDP as best my calculations can tell (usage under full load a full 100W less than the 7960x @ 4GHz). The lowest all-core speed I ever saw was 3.6GHz (base clock is 3.3) and 95% of the time the chip sat at 4.0GHz.

Intel itself says that the best way to know what a processor is consuming is to test it, which is exactly what Chrispy_ is pointing out that numerous reliable sources have done.

https://www.intel.com/content/dam/doc/w ... -paper.pdf

Intel wrote:

...there is a very accurate way to measure server power under a users “real world” conditions. And again, Intel and AMD both agree on this point: “The best way to measure a server’s power consumption is the power meter, an inexpensive tool that is plugged into the wall, and then your device, like a server, can be plugged into the power meter...

I've loved reading the comments and appreciate all the help. If you can oblige, I do have one more request.

Finding a straightforward guide to undervolting LGA 2066 chips hasn't been easy. This shows that it can be done and the results are impressive, but doesn't go into the methodology in a meaningful way. My eventual hope is to run the box at stock on air cooling with minimal throttling, and given the difference I saw in shunting an FX-8320 from 1.35v to 1.2125v, it'd be good to see my 7940x behave a bit better. Thank you again.

Alright, I may as well conclude this thread and put a bow on it. I went out and acquired a Noctua NF-F12 fan and strapped that onto the be quiet! Dark Rock 4, moving the default SilentWings 3 fan to the other side to push air to the case's vent fan. That definitely helped - Cinebench performance jumped from 5282 cb with AVX offsets disabled to 5566 with the motherboard running defaults. I may play with undervolting later, but this feels like it's managing thermal load much better... The real test is going to be subjecting it to Blu-ray rips with Handbrake and seeing how it fares. Thanks for all of your help!

Concupiscence wrote:

Alright, I may as well conclude this thread and put a bow on it. I went out and acquired a Noctua NF-F12 fan and strapped that onto the be quiet! Dark Rock 4, moving the default SilentWings 3 fan to the other side to push air to the case's vent fan. That definitely helped - Cinebench performance jumped from 5282 cb with AVX offsets disabled to 5566 with the motherboard running defaults. I may play with undervolting later, but this feels like it's managing thermal load much better... The real test is going to be subjecting it to Blu-ray rips with Handbrake and seeing how it fares. Thanks for all of your help!

So I actually own this processor. During cinebench r20 it maintains 3.8 all core (on a hyper 212 plus, lol) a bit above 6700. It seems pretty close to it's tdp except in one instance, prime 95 small fft all core load which pushes it to 3.5 (I think). I'll double check power usage etc. It's summer time and my ps4/monitor are on the same feed.

I used to own an MSI x299 sli plus (good board, bad for overclocking, but good stock). I currently use an MSI x299 carbon pro ac. I know my friend got similar scores at stock.

thecoldanddarkone wrote:

Concupiscence wrote:

Alright, I may as well conclude this thread and put a bow on it. I went out and acquired a Noctua NF-F12 fan and strapped that onto the be quiet! Dark Rock 4, moving the default SilentWings 3 fan to the other side to push air to the case's vent fan. That definitely helped - Cinebench performance jumped from 5282 cb with AVX offsets disabled to 5566 with the motherboard running defaults. I may play with undervolting later, but this feels like it's managing thermal load much better... The real test is going to be subjecting it to Blu-ray rips with Handbrake and seeing how it fares. Thanks for all of your help!

So I actually own this processor. During cinebench r20 it maintains 3.8 all core (on a hyper 212 plus, lol) a bit above 6700. It seems pretty close to it's tdp except in one instance, prime 95 small fft all core load which pushes it to 3.5 (I think). I'll double check power usage etc. It's summer time and my ps4/monitor are on the same feed.

I used to own an MSI x299 sli plus (good board, bad for overclocking, but good stock). I currently use an MSI x299 carbon pro ac. I know my friend got similar scores at stock.

PM me if you like. I'd love to bounce things off of you and see if I can't nudge its performance upward.

I know it's not kosher to necrobump, but I'm gonna swallow my pride and say that I found out why the 7940x wasn't performing well. Apparently sleep deprivation and system building don't go together... When it was time to install the Dark Rock 4, I failed to notice the "remove this before installing" adhesive sticker from the base of the heatsink. I had a HEDT CPU running with demanding workloads for nine months with a sticker impeding efficient thermal transfer.

Upon correction, and with a little time for the thermal paste to cure, my Cinebench score jumped to 6847. For a more productive and longer-term workload, my performance with a Blu-ray rip via Handbrake jumped from 47 fps to 58. Thank you to everyone who dropped in with suggestions; I'm gonna run this thing at stock until it keels over.

Ha. You'll never forget again, that's for sure. I did the same with my 6700K...the waterblock was sitting on a choke near the socket (just barely) and was touching the heatspreader on only one corner. I couldn't for the life of me figure out why it was so hot and clocked so poorly.

I did the waterblock install 3 times before I caught it purely by accident, I was on the road to convincing myself I was going to delid it at the time.

That's a mistake you only make once. I suspect most of us have done it.

I did it with my original Q6600 installation. I immediately pondered why idle temperature was way above ambient (~50C) with the first start-up. I figured it was poor mounting but when I was re-seating the heatsink. I'd immediately noticed that the plastic cover was still on the heatspreader.

The one and only good thing you can say about this kind of mistake is, it's the quickest and easiest temps problem to fix (after you identify it...)