AMD ThreadRipper Reviews [Aug 30 - Tom's 1920X]

[Ajay]

@Markfw AT test setup: "Memory G.Skill Trident Z RGB DDR4-3200 C14 4x8GB"

Anyway, given the non-deterministic behavior of Windows 10 thread scheduling, any assessment of CPU core frequency vs load vs total thread occupancy would need to be done on Linux.

 

[csbin]

https://www.3dcenter.org/news/threa...tate-zur-anwendungs-performance-im-ueberblick

 

[Zucker2k]

hmmmm I missed that. But the question is, was that with all 8 slots filled ? And overclocked ? I will be at 4 ghz, and 3600 and try to run tests at that and 3200 and see what happens. Since Ryzen/EPYC/Threadripper are so sensitive to memory speed, I would guess that 3600 could make a huge difference in benchmarks.... We will see next Monday if I get it in time to test. Or Tuesday.

Edit: The other thig I disagree with is the comment that faster memory only helps in memory limited scenarios. From that I have read, since the zen core talks to all cores at memory speed, ALL things run faster with faster memory.

Since you're overclocking (which'll disable turbo), maybe this power management measure won't kick in. In stock mode, however, story is different with the 3200mhz+ rams.

 

[Bouowmx]

Some info about the memory power "siphon", from Tom's Hardware: http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1950x-cpu,5167-14.html

Both of AMD’s CPUs are designed for a maximum power consumption of 180W at their default settings. If the memory’s overclocked, then the CPU gets 15W less, which might affect performance in usage scenarios that employ all of the cores and, consequently, get too close to the limit.
<break>
The Asus X399 ROG Zenith Extreme motherboard limits power consumption to exactly 180W, just as it should, when using the default settings. Things look a whole lot different once the processor is manually overclocked to maximize its frequency, though. The 1950X needs 1.35V to achieve 3.9 GHz. At that point, AMD's new processors join Intel's overclocked Core i9-7900X well beyond 300W.

Which settings to change are unclear, and the ASUS ZENITH EXTREME manual does not go over all the BIOS options.

What if I want to run stock frequencies, and just remove power limit?

 

[Markfw]

To all previous replies to my memory question:

It seems very unclear how all of this works out in both stock and overclocked setting with 3600 ram, and I feel that much testing was not done in this area. Its known that the Zen core loves fast memory, yet no one really pushed the benchmarks in this area. With my own hardware, I will ATTEMPT to do some testing in this area, and I ask all that have good 3600 memory on their TR chips to help me next week when we all get our hardware.

 

[Edrick]

It is very funny watching the same people criticize who Sky-X for not being a super gaming CPU make excuses for TR being in the same exact boat.

Lets face it, HEDT is not only for gamers. TR and Sky-X shine in many other areas.

 

[mattiasnyc]

It is very funny watching the same people criticize who Sky-X for not being a super gaming CPU make excuses for TR being in the same exact boat.

Lets face it, HEDT is not only for gamers. TR and Sky-X shine in many other areas.

Find us the people who did that, quote them, and then we'll believe that.

My hunch is that whomever made that argument, if it happened at all, did that in response to something else, something specific. And ignoring the reason the point was made is, well, missing the point.

 

[mattiasnyc]

That PCWorld review guy seemed a bit high or drunk....

 

[IEC]

Find us the people who did that, quote them, and then we'll believe that.

My hunch is that whomever made that argument, if it happened at all, did that in response to something else, something specific. And ignoring the reason the point was made is, well, missing the point.

Most of the people ripping (haha, get it?) the new architectures don't see the value because their workloads don't scale to that many cores. The fact of the matter is that AMD beginning the "core wars" has brought more cores for cheaper to EVERYONE (insert Gary Oldman gif from The Professional here) - AMD *and* Intel. So if your workloads favor Intel processors - that's great, you're saving a bundle over the Broadwell-E series. I am legitimately happy for you - this kind of CPU power in an inexpensive desktop chip was unthinkable just a short year ago. If you prefer the X399 platform, you're happy AMD is actually competitive again. I am, and I threw more Benjamins at AMD in the past year than I did in the past decade (by far). If you're a 1337 hax0r gamer, why are you even looking at HEDT again?

It's about time we had some movement in the CPU arena, and regardless of brand preference people should be happy about the current state of desktop and HEDT CPUs. The best is yet to come and I can't wait for AMD and Intel to pick up the pace.

 

[krumme]

Threadripper downclocks a notch to compensate for the extra power draw from the memory controllers. So you actually lose performance if you go crazy on the ram overclocks. Ouch!

So now amd is the brand that enable their tdp limiter while intel have disabled it on top of the tim. So it regulary exceeds tdp out the box. Selling factory oc configurations out the door.
Way to ruin your presence on the b2b market in no time.
Look no further than amd rx570/580 to see where it goes and does for brand.
Shortsighted beancounter decisions of lowest professional standard.
To see Intel in such a desparate state is just sad.

 

[VirtualLarry]

It's about time we had some movement in the CPU arena, and regardless of brand preference people should be happy about the current state of desktop and HEDT CPUs. The best is yet to come and I can't wait for AMD and Intel to pick up the pace.

I'm seriously wondering, now that we've seen "engineering samples" of ThreadRipper de-lidded, with FOUR die, not just TWO... could they be failed / downbinned Epyc CPUs? Will the TR4 socket support more than two dies? I realize that the memory channels and PCI-E lanes of two of the dies will not be connected, supposedly, but couldn't the CPU dies remain active, and just use the Infinity Fabric to communicate with "non-local" memory / PCI-E space?

I was just thinking towards the future, if AMD could offer a 32-core / 64-thread ThreadRipper, without releasing a new socket?

 

[Abwx]

I'm seriously wondering, now that we've seen "engineering samples" of ThreadRipper de-lidded, with FOUR die, not just TWO... could they be failed / downbinned Epyc CPUs? Will the TR4 socket support more than two dies? I realize that the memory channels and PCI-E lanes of two of the dies will not be connected, supposedly, but couldn't the CPU dies remain active, and just use the Infinity Fabric to communicate with "non-local" memory / PCI-E space?

I was just thinking towards the future, if AMD could offer a 32-core / 64-thread ThreadRipper, without releasing a new socket?

Those two dies are here ony for mechanical reasons to make sure that the heatsink is homogeneously contacted....

They could release SKUs with higher core count but frequency would be severly reduced, by something like 30% for a 32C, not sure that it would have some traction.

Otherwise the dies used for TR are selected in the 2% best dies.


http://www.hardware.fr/articles/967-1/amd-threadripper-1950x-1920x-test-quelque-chose-epyc.html

 

[Zucker2k]

So now amd is the brand that enable their tdp limiter while intel have disabled it on top of the tim. So it regulary exceeds tdp out the box. Selling factory oc configurations out the door.
Way to ruin your presence on the b2b market in no time.
Look no further than amd rx570/580 to see where it goes and does for brand.
Shortsighted beancounter decisions of lowest professional standard.
To see Intel in such a desparate state is just sad.

The miracle here is that the delta between TIM and SOLDER is only 100 - 200mhz, and only becomes a factor in extreme overclocking situations. The fact is, the Threadripper memory controller consumes quite a bit of power the higher you go on the ram frequency scale. I'm surprised you're not happy a modern day cpu can intelligently take advantage of thermal (and power) headroom to clock itself faster and complete tasks quicker. That is what turbo is, and if it's a bad thing, then both AMD and Intel are guilty of it. The irony here though is, the "factory oc configurations" also have the bigger overclocking headrooms, TIM be damned.

 

[MarkPost]

The miracle here is that the delta between TIM and SOLDER is only 100 - 200mhz, and only becomes a factor in extreme overclocking situations. The fact is, the Threadripper memory controller consumes quite a bit of power the higher you go on the ram frequency scale. I'm surprised you're not happy a modern day cpu can intelligently take advantage of thermal (and power) headroom to clock itself faster and complete tasks quicker. That is what turbo is, and if it's a bad thing, then both AMD and Intel are guilty of it. The irony here though is, the "factory oc configurations" also have the bigger overclocking headrooms, TIM be damned.

To me, OC isnt the key factor here. The key factor is heat. I'm less than confortable with my CPU reaching +85 and +90ºc while encoding or rendering some stuff for hours. Its just not sustainable. We have the AVX example. Its great I can oced my system to "x" frec but if I use some app using avx, I need to apply an offset, due to heat reached, so I lose a good portion of perf.

 

[Zucker2k]

To me, OC isnt the key factor here. The key factor is heat. I'm less than confortable with my CPU reaching +85 and +90ºc while encoding or rendering some stuff for hours. Its just not sustainable. We have the AVX example. Its great I can oced my system to "x" frec but if I use some app using avx, I need to apply an offset, due to heat reached, so I lose a good portion of perf.

Compared to what?

 

[krumme]

The miracle here is that the delta between TIM and SOLDER is only 100 - 200mhz, and only becomes a factor in extreme overclocking situations. The fact is, the Threadripper memory controller consumes quite a bit of power the higher you go on the ram frequency scale. I'm surprised you're not happy a modern day cpu can intelligently take advantage of thermal (and power) headroom to clock itself faster and complete tasks quicker. That is what turbo is, and if it's a bad thing, then both AMD and Intel are guilty of it. The irony here though is, the "factory oc configurations" also have the bigger overclocking headrooms, TIM be damned.

If i look at bwe and temp i cant see the 100-200 difference to skl x. And when people delid temp goes down like crazy. Besides temp is efficiency and probably sklx worse efficiency vs bwe is also due to higher temp. So it matters. Especially as Intel oc the cpu for you.

Intel created a cpu that looks like a dragster running on nitrox - delidded. Yes it overclocks better but so what? Performance (efficiency cost...) is what matters.
Its a cpu without a purpose the day tr arived and the factory oc just made it worse. We need cheaper and more efficient high core skl x and cl 6c for the gaming. Or buy a used bwe for that.

 

[MarkPost]

Compared to what?

to my max oc. If my max oc is 4.5-4.6, I need to apply a good offset to avoid reaching unsustainable temps, so I lose a good portion of avx theoretical advantage

 

[wildhorse2k]

Did anybody find an explanation why are the Threadripper latencies much higher than expected in comparison to standard Ryzen?

http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1950x-cpu,5167-2.html

CPU | Cross-CCX Average Latency | Cross-CCX Average Latency
TR 1950X Creator Mode DDR4-3200 | 160.1ns | 216.9ns
Ryzen 7 1800X | 122.96ns |

I would have expected the Cross-CCX Average Latency to be about the same. They used 4x 8GB G.Skill RipJaws V DDR4-3200 for TR although they didn't publish timings.

 

[scannall]

I'm seriously wondering, now that we've seen "engineering samples" of ThreadRipper de-lidded, with FOUR die, not just TWO... could they be failed / downbinned Epyc CPUs? Will the TR4 socket support more than two dies? I realize that the memory channels and PCI-E lanes of two of the dies will not be connected, supposedly, but couldn't the CPU dies remain active, and just use the Infinity Fabric to communicate with "non-local" memory / PCI-E space?

I was just thinking towards the future, if AMD could offer a 32-core / 64-thread ThreadRipper, without releasing a new socket?

Only if each CCX had 8 cores instead of 4.

 

[ub4ty]

Did anybody find an explanation why are the Threadripper latencies much higher than expected in comparison to standard Ryzen?

http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1950x-cpu,5167-2.html

CPU | Cross-CCX Average Latency | Cross-CCX Average Latency
TR 1950X Creator Mode DDR4-3200 | 160.1ns | 216.9ns
Ryzen 7 1800X | 122.96ns |

I would have expected the Cross-CCX Average Latency to be about the same. They used 4x 8GB G.Skill RipJaws V DDR4-3200 for TR although they didn't publish timings.

Yeah, they're dropping the ball on being clear on the details of the setup and also I fail to understand why they didn't test 3600Mhz ram.

They make a grand statement :

Both of AMD’s CPUs are designed for a maximum power consumption of 180W at their default settings. If the memory’s overclocked, then the CPU gets 15W less, which might affect performance in usage scenarios that employ all of the cores and, consequently, get too close to the limit.

And then provide no detailed information/analysis hash out what exactly they're talking about and under what scenarios... *sigh.

I would like to see a solid review of Threadripper's performance, power usage, and latency across various RAM modules from 3000, 3200, 3600 and possibly 4000.

 

[nathanddrews]

Did anybody find an explanation why are the Threadripper latencies much higher than expected in comparison to standard Ryzen?

http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1950x-cpu,5167-2.html

CPU | Cross-CCX Average Latency | Cross-CCX Average Latency
TR 1950X Creator Mode DDR4-3200 | 160.1ns | 216.9ns
Ryzen 7 1800X | 122.96ns |

I would have expected the Cross-CCX Average Latency to be about the same. They used 4x 8GB G.Skill RipJaws V DDR4-3200 for TR although they didn't publish timings.

I think using Creator Mode in the Tom's Review is your answer there. See AnandTech's review:

Spoiler: Creator vs Game vs Latency vs RAM Speed

This is like a 40% reduction in latency going from Creator DDR4-2400 to Game DDR4-3200.

EDIT: Sorry, I mistook CCX latency for cache latency. Pls don't hurt me.

 

[Ajay]

Did anybody find an explanation why are the Threadripper latencies much higher than expected in comparison to standard Ryzen?

http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1950x-cpu,5167-2.html

I would have expected the Cross-CCX Average Latency to be about the same. They used 4x 8GB G.Skill RipJaws V DDR4-3200 for TR although they didn't publish timings.

Since they are really testing inter-thread communication - who really cares. This won't have much effect on client applications. Different situation for Enterprise applications. The more important value for clients is near and far memory access times, and we've seen that they don't make much difference in actual apps.

 

[moinmoin]

Regarding latencies: Updated core to core latencies by PC Perspective

Please note that this graph has all systems run at DDR4 2400 regardless of capability:

Threadripper comparison between 2400 and 3200:

So in that case a 50% increase in memory speed led to a ~20% reduction in core to core IF latencies.

For RAM latencies and bandwidth we can probably look at Epyc's numbers: ServeTheHome had a rundown of those

Those are taken with DDR4 2666, note that Threadripper being only two dies and single socket only direct access (orange) and one die hop (green) are relevant:

 

[IEC]

It's irritating when reviewers list the rated MT/s of their DDR4 kits, but don't list the rated timings, or the actual timings used.

Ryzen benefits quite a bit from tuning the timings (and subtimings), and the gap between CL16 DDR4-3200 and a tuned CL14 kit can be quite substantial.

Even fine tuning 3200 CL14 B-die kits can yield double digit differences:

 

[nathanddrews]

AnandTech Game Mode Deep Dive
http://www.anandtech.com/show/11726...-game-mode-halving-cores-for-more-performance

Spoiler: Summary