The Ryzen "ThreadRipper"... 16 cores of awesome

[krumme]

Its pretty good to see its possible to discuss TR and skl x in the same thread without anyone going into defence mode saying this is only a xx thread. The cpu solves the same task and they target some of the same segment. So the ability to discuss power efficiency and perf next to each other is important.

 

[itsmydamnation]

Two cores running at 4.5ghz cannot produce more heat than 10 cores running at 4.3ghz.

That depends on where you are on the voltage curve on the two clock speeds, it can 100% happen because that curve can/is non linear.........

 

[Osjur]

Its partly because of the TIM used there but also because these chips are getting denser which means more hotspots and that is just harder to cool. Everybody who has done even a little bit of extreme oc knows that there is a point when it doesn't matter if you have custom loop with water at close to freezing point and direct die cooling, you still can't transfer the heat away fast enough and your temps skyrocket.

Two cores running at 4.5ghz cannot produce more heat than 10 cores running at 4.3ghz.

They are not producing more heat overall but they are producing heat in a hotspots and when you have poor chip to ihs contact, meaning poor thermal conductivity, we get a situation where you will see a lot higher core temps even if your cooler is capable of handling more.

 

[Wyrm]

It's not the power consumption that directly induces the temps on 7900X, it's the TIM. If it weren't for the thermal interface problem, even beefy air coolers could have handled SKL-X at stock and maybe a bit beyond.

IMHO, it's the other way around: power induces rising temperature. You can only lower the thermal plateau with faster heat transfer but you can't magically waive off 160W that the die has dissipated. One way or another those 160W need to come out. So, let me rephrase my statement: what thermal plateau are you expecting the 16 core TR will achieve if it's running 160W with solder on Prime95 with AVX2? My claim is that you will see the same high temperatures in this situation. I'm concerned that people will throw the TR chip under the bus when they see these thermals.

 

[Zucker2k]

That depends on where you are on the voltage curve on the two clock speeds, it can 100% happen because that curve can/is non linear.........

True, but remember that both of these scenarios are under turbo conditions. These chips demonstrably have a sweet oc spot at 4.5ghz. It's just impossible, given these characteristics for four threads running a bit higher voltage to outdo 20 sustained threads in heat generation, imo.
With heat spots, I think the favored core approach would more than likely factor in cores with the best thermal characteristics, among other things since these (2) cores are going to be the most active cores. This only makes sense.

 

[coercitiv]

IMHO, it's the other way around: power induces rising temperature. You can only lower the thermal plateau with faster heat transfer but you can't magically waive off 160W that the die has dissipated. One way or another those 160W need to come out.

Who said anything about magically waiving 160W?! You claimed TR will face the same issues as 7900X, and I pointed out 7900X is not a good comparison point since the TIM greatly affects cooling performance, significantly lowering the cooling efficiency even for systems capable of a lot more than 140W.

A good rhetorical question for you, straight from Tom's Hardware:

Why can't those liquid coolers keep up with a CPU like the -7900X? Back in the day, a normal all-in-one was good enough to keep the Core i7-5960X running cool, even overclocked to 4.8 GHz. We measured power consumption numbers of up to 250W back then. So, why did we have to force a constant 20°C in the loop to even start experimenting?

So, let me rephrase my statement: what thermal plateau are you expecting the 16 core TR will achieve if it's running 160W with solder on Prime95 with AVX2?

That's not a statement, it's a question, and you're the one who made the original claim, so be my guest and post your estimate.

 

[beginner99]

IMHO, it's the other way around: power induces rising temperature. You can only lower the thermal plateau with faster heat transfer but you can't magically waive off 160W that the die has dissipated. One way or another those 160W need to come out. So, let me rephrase my statement: what thermal plateau are you expecting the 16 core TR will achieve if it's running 160W with solder on Prime95 with AVX2? My claim is that you will see the same high temperatures in this situation. I'm concerned that people will throw the TR chip under the bus when they see these thermals.

But TR has a huge IHS and therefore will need new heatsink with an accordingly huge baseplate. more surface = more /faster heat transfer.

 

[tamz_msc]

I believe that it is a combination of ~1.3x increase in die size but in the same package, the 14nm+ process which runs hot when pushed beyond a certain point, and the use of thermal paste which causes the temperature issues in the 7900x.

Only time will tell how well AMD can cool two dies in the behemoth package that is TR.

 

[Kenmitch]

I believe that it is a combination of ~1.3x increase in die size but in the same package, the 14nm+ process which runs hot when pushed beyond a certain point, and the use of thermal paste which causes the temperature issues in the 7900x.

Only time will tell how well AMD can cool two dies in the behemoth package that is TR.

Seeing how EPYC has 4 dies I wouldn't think 2 would be a issue.

 

[Topweasel]

Keep in mind that the thermals on Skylake are going to be much more spotty and harder to manage because it just plain runs hotter. This has one of the highest thermal cut off's of any CPU we have seen. A TR or EPYC or OC'd Ryzen still need good cooling to stay operable. But they cut off earlier and generally need to run cooler. This works in AMD's favor as even though you have to make sure you are clearing 100-120-140-160w from the cooler the ambient temperatures in the case won't rise as much and contribute to hitting the thermal limits of the fans/cooler. Sure both are going to heat up a room at the same speed. But temps inside the case should be better, it's also one of the reasons I generally use my AIO's as an exhaust and not inlet.

So the big issue with Skylake X isn't power levels. It's what temp it will be running at with X level of cooling and how razer edge those margins between out of control and ok.

 

[maddie]

IMHO, it's the other way around: power induces rising temperature. You can only lower the thermal plateau with faster heat transfer but you can't magically waive off 160W that the die has dissipated. One way or another those 160W need to come out. So, let me rephrase my statement: what thermal plateau are you expecting the 16 core TR will achieve if it's running 160W with solder on Prime95 with AVX2? My claim is that you will see the same high temperatures in this situation. I'm concerned that people will throw the TR chip under the bus when they see these thermals.

T delta is lower with solder that TIM as you know, due to higher conductivity.

We should have some idea of the drop in temps for the same power dissipated. What are the temp drops from deliding and using liquid metal instead. That should point us to the difference between TR and the competition.

Hotspots are the wildcard..

 

[estarkey7]

Encoding is a tiny part of my workflow. That could be because I'm a non-professional and spend a lot of time mucking about attempting color correction, re-stabilizing footage that several stabilization packages couldn't manage, etc, but I'm going to assume I'm not alone in having encoding be a small percentage of workflow. As for "golden core", my assumption has been that the top turbo is roughly where I can expect an all-core OC to max out at. I find it useful to talk about those numbers for that reason, as opposed to believing that I'll somehow luck out at getting a single-core running. Decode is a several thread process in the software under discussion, so even if I'm doing stabilization or motion-tracking, or something else that's essentially single-threaded (some news-like overlays I use appear to be single-threaded which causes stuttering on my 4k footage -- there's a surprising amount of random stuff that isn't properly parallelized), I'm still going to be using a crap ton of cores. That doesn't mean I don't care about throughput of a single core, though. It's all about the long-pole. If it fits within a core's performance, no problem and more cores is better; but if it doesn't, all is not well in the garden....

Are you using premiere pro cc? If so, playback and color correction through Lumetri is gpu accelerated. So drop in a 1080 Ti or a VEGA and call it a day. That's my plan for my Threadripper build.

Sent from my rk3188 using Tapatalk

 

[dnavas]

Are you using premiere pro cc? If so, playback and color correction through Lumetri is gpu accelerated. So drop in a 1080 Ti or a VEGA and call it a day. That's my plan for my Threadripper build.

No, Edius. Edius only supports Intel's IGP for decode, which might not have been the worst decision at some point in the past given the state of art in nvdec and uvd, but it's getting long past time that they got support. They do support gpus for color correction, however I'd need to upgrade to Win10 (something I intend to do, just haven't).

But I think what I was trying to get at is the dozens of little gotchas. I add a feature, maybe some animated title, and suddenly I'm not RT anymore, because single-threaded titles worked fine in 1080, but in 4k not so much. There's an ecosystem that has to move to the next level, and it's a steep step-function to recode everything well-parallelized. If I can minimize my surface area of intersection by buying the chip that's a little faster, and clocks a little higher, the pain avoided might be worth the $200 spent (or whatever).

 

[ub4ty]

As it is possible to disable ryzen cores using an AMD utility and in Bios :

Wouldn't it be possible to simply just disable 4-8 cores on Threadripper when you're not doing super intensive multi-tasking workloads and drop the power utilization/heat output significantly? I am considering this as a straight-forward option to cut down on power/heat when I don't need all 16 cores. Has anyone done a detailed writeup on this w.r.t Ryzen 1700/1700x/1800x (power consumption/performance) with various quantities of cores disabled? This would give a good idea of what to expect.

Are people who are commenting about power/heat overlooking this simple option?

 

[The Stilt]

As it is possible to disable ryzen cores using an AMD utility and in Bios :

Wouldn't it be possible to simply just disable 4-8 cores on Threadripper when you're not doing super intensive multi-tasking workloads and drop the power utilization/heat output significantly?

Sure it is, as long as you don't mind performing a cold reset every time you want to change the core configuration.
The unused cores will be power gated anyway and the difference having them turned off completely is pretty much non-existent.

 

[coercitiv]

Wouldn't it be possible to simply just disable 4-8 cores on Threadripper when you're not doing super intensive multi-tasking workloads and drop the power utilization/heat output significantly?

Are people who are commenting about power/heat overlooking this simple option?

Your question has already been answered by CPU engineers years ago. CPU power management puts unused cores to sleep and reduces their power usage to values close to zero. This process has reached excellent granularity and response times, and disabling cores will almost always be a worse option than simply allowing the CPU to work as intended.

On top of that, well multi threaded workloads may actually generate less total heat when all cores all active.

 

[ub4ty]

Well, this is what I wanted to bring out...

The unused cores will be power gated anyway and the difference having them turned off completely is pretty much non-existent.

Obviously it's good to have a lower TDP. However, for the most part, people aren't going to running flat out 16 core full throttle. As such, power draw/heat output is generally going to be subdued on these mammoth high core count CPUs. In the case that you are running flat-out, you'd be most concerned with how well that heat-spreader is connected to the die so as to allow your heat-sink to dissipate the heat efficiently. This is where I have problems w/ Intel. 180W TDP is fine for 16 cores (AMD) vs. 140W TDP for 10 cores (Intel). At least I know with AMD that I can efficiently dissipate the heat when I begin pushing the CPU. Otherwise, in general operation, it really wont be using much.

 

[ub4ty]

Your question has already been answered by CPU engineers years ago. CPU power management puts unused cores to sleep and reduces their power usage to values close to zero. This process has reached excellent granularity and response times, and disabling cores will almost always be a worse option than simply allowing the CPU to work as intended.

On top of that, well multi threaded workloads may actually generate less total heat when all cores all active.

I was hoping to have this stated as it seemed the focus of attention w.r.t to performance/power/heat was glancing over this.

When pushing the CPU, TDP comes into full focus. Otherwise, you're using far less power and outputting far less heat scaled with your load. How well the CPU's heat spreader is attached to the die seems to be the starting point w.r.t to discussions about high load/power draw/heat output scenarios.

The 10 core intel is running at 140W TDP.
The 16 core amd is expected to be around 155W TDP w/ far more efficient heat spreader connectivity.

I'm not too concerned about power draw on Threadripper, heat, or heat dissipation.
I would be with intel... For 2 more cores over Ryzen 8core, I'm dealing with a 50% increase in power usage if I go w/ Intel plus a bad heat spreader connection. Instead, I can more than double the core count from Ryzen 8core to threadripper 16core only for 15W TDP above the intel chip + way better heat spreader connectivity. It's not even a question as to which option I go with. At low CPU util, I won't be using much above a Ryzen 8 core processor.
- Power on demand
- Double the core count for only 50% increase in power (at the high end of utilization)
- Well connected heat-spreader

 

[Markfw]

Well, this is what I wanted to bring out...


Obviously it's good to have a lower TDP. However, for the most part, people aren't going to running flat out 16 core full throttle. As such, power draw/heat output is generally going to be subdued on these mammoth high core count CPUs. In the case that you are running flat-out, you'd be most concerned with how well that heat-spreader is connected to the die so as to allow your heat-sink to dissipate the heat efficiently. This is where I have problems w/ Intel. 180W TDP is fine for 16 cores (AMD) vs. 140W TDP for 10 cores (Intel). At least I know with AMD that I can efficiently dissipate the heat when I begin pushing the CPU. Otherwise, in general operation, it really wont be using much.

Well, I have to differ with you on the point about not running all 16 cores flat out.

Think of usage situations, you are a gamer and a multitasker. 8 cores and 16 threads (Ryzen as an example) should be plenty, so I doubt many of them will buy TR.

Then you have people that run DR or mining. They want all the cores they can get, and run everything@100% all the time. Or small servers/data centers, they may also be inerested in TR, and might also run@100%.

Then you have bigger servers or data centers, and they most likely will go for EPYC CPUs and the like.

So, I may have missed some situations, but the prime candidates for TR I think WILL be running them@100%.

Thoughts ?

 

[tamz_msc]

Or in my case I'll be running a Linux VM all the time on at least 4 cores and Mathematica on Windows, or in other words, I'd gobbling up RAM actually putting quad-channel memory to use. Basically it'll be doing the job of the i7 3770 and E5 2640v3 combo I currently use but in one CPU.

 

[Abwx]

The 10 core intel is running at 140W TDP.

140/320 = 0.4375 W/mm2

The 16 core amd is expected to be around 155W TDP w/ far more efficient heat spreader connectivity.

155/378 = 0.41 W/mm2

 

[ub4ty]

Well, I have to differ with you on the point about not running all 16 cores flat out.

Think of usage situations, you are a gamer and a multitasker. 8 cores and 16 threads (Ryzen as an example) should be plenty, so I doubt many of them will buy TR.

Then you have people that run DR or mining. They want all the cores they can get, and run everything@100% all the time. Or small servers/data centers, they may also be inerested in TR, and might also run@100%.

Then you have bigger servers or data centers, and they most likely will go for EPYC CPUs and the like.

So, I may have missed some situations, but the prime candidates for TR I think WILL be running them@100%.

Thoughts ?

I generally concur with your use case scenarios but will leave room for (unique) and increasingly likely consumer use case scenarios for such horsepower.

In such scenarios in which the chips are being utilized to a notable degree, TDP is important as well as thermal dissipation/efficiency.

2 more cores than Ryzen 8 core and you're using 50% more power with Skylake-X. That a Yuge problem.
Whereas, with thread ripper, Running flat out 100% (~155W TDP), you get 6 more cores and 12 more threads for 50% increase in power over Ryzen 8 core and ~15W more in TDP vs. intel (10core). Plus you're assured you will have efficient thermal connectivity from heat spreader to the die ensuring that you will be able to dissipate the resultant heat. Between the extortionist pricing of intel's CPU, x1.5-2x the price w.r.t motherboard, the high TDP vs. core count, and the crappy heat spreader to die connectivity, I'm having a hard problem going with intel.

Earlier I mentioned : but will leave room for (unique) and increasingly likely consumer use case scenarios for such horsepower.
Gamers already enjoy pushing the envelope w.r.t to their setups and there are an increasingly larger number of multi-tasking operations being done while gaming. We are also entering the age of compute accelerators hanging off PCI-E. As such, from both a hardware and software standpoint, there is an increasing case to be made for bumping above 8 cores/16 threads (if priced right). AMD is pricing things right. So, if thread-ripper is priced right, there will be notable demand from the consumer market. If the CPU is not being pushed, power draws/heat aren't an issue. Power on Demand. So, why not have some extra in waiting? So, for some extra $$$, it will be worth it to step up into such a platform (if priced right).

The (unique) mention leaves room for new software developments on the horizon. I think you'll be surprised by how much highly multi-threaded applications take off in the near future (to a high degree).

 

[scannall]

x1.5-2x the price w.r.t motherboard,

The only place that I will disagree with you is on the Motherboard price. X399 will be expensive to build (4094 pins after all), and probably in the same ballpark, or slightly higher than X299 boards. Other than that, pretty spot on though.

 

[ub4ty]

The only place that I will disagree with you is on the Motherboard price. X399 will be expensive to build (4094 pins after all), and probably in the same ballpark, or slightly higher than X299 boards. Other than that, pretty spot on though.

The issue being the ridiculous price variability and sloppy configurations of X299 motherboards. Prices range anywhere from $258-$489. At the low range :
https://www.newegg.com/Product/Product.aspx?Item=N82E16813144058
If Threadripper boards come in under $258 and have a upper lower ceiling, that is what would be of note.
I do imagine a $200 or even lower threadripper motherboard (double the lower avg. price of a Ryzen 8) board.
Costs add up and AMD is doing well in K.I.S.S.

Skylake is all over the place.

 

[estarkey7]

Thanks for your reply. No worries. I am very frustrated as my 4790k is not cutting it for multi-camera 1080p60 jobs but finding something that does cut it is proving to be difficult. I am pretty sure TR will be lacking in clock speed to be a fit. I want to support AMD but a delidded 7900x from Silicon Lottery appears to be my best hope of improved ST with very nice MT. If I can get 4.5 GHz on all cores then I think there will be a performance difference from the 4790k.

My software is less threaded than Adobe Premiere, Sony Vegas etc... One would think monster cores would be the answer but it is not the case. Anyway, good luck with TR!

Hardware is not your issue. Maybe you need to switch editing software. Maybe move to Blackmagic Design Davinci Resolve, but Adobe Premiere Pro can't be beat.