AMD ThreadRipper Gen 2 Reviews

[ScratchCat]

58 minutes ago, cj09beira said:

keep an watch on that page as derbaur is aiming at dual 7601 on ln2 which might be enough for some crazy scores

Approaching 1-5 second render times with Cinebench is going to lead to some crazy variance.

 

Intercore communication is getting insanely expensive, Skylake-X expends 40% of its energy budget on the mesh while Epyc expends ~60% ryzing (only one pun, I promise!) to an epyc (Never trust internet folk) 89% at idle. To put it in perspective, Epyc uses as much power for IF as an 8700K at full load.

Zen may scale well, without costs this scaling however is not.

 

 

[S w a t s o n]

13 minutes ago, sazrocks said:

8419cb achieved... holy crap

http://hwbot.org/submission/3916493_sampson_cinebench___r15_ryzen_threadripper_2990wx_8419_cb

Interesting, lower clock speed but slightly tighter memory timings. Nice find!

[sazrocks]

@The Benjamins Can you add Phoronix to the OP?

 

https://www.phoronix.com/scan.php?page=article&item=amd-linux-2990wx&num=1

[Firewrath9]

what about dual socket 7601? what cinebench score does that have?

[sazrocks]

4 minutes ago, Firewrath9 said:

what about dual socket 7601? what cinebench score does that have?

7252

http://hwbot.org/benchmark/cinebench_-_r15/rankings?start=0#start=0#interval=20

 

though I had thought that der8auer has gotten higher...

[Taf the Ghost]

Looks like Handbrake is the first "common" program, in the space, to not like 64 threads. 

 

 

Guy turned off SMT and got a performance boost in Handbrake. Explains why that one seems so slow in a bunch of benchmarks in the first run.

[leadeater]

18 hours ago, Taf the Ghost said:

Without Octa-channel memory, I think 32 cores is probably just too much for any really memory intensive workload. The first batch of simulated 2970WX (24 core) really seems like it'll be the better buy. 2950X really has come along with the year of improvements on the platform, too.

 

That many cores might just be too much for quad-channel until DDR5 lands in a few years. It's fun, but it's going to be pretty niche.

Glad to see this actually been highlighted by multiple reviews, the amount of flak some people gave me when pointing out memory bandwidth as a problem when increasing cores geez. Time to find that I told you so T-Shirt.

[Doobeedoo]

Sweet. Twice the core increase for exisitig mobos just like that, really awesome. The price is great for such a CPU too.

Obviously for heavily paralleled workload and software that can take advantage of all the cores efficiently, though would be nice to see some heavy multitasking. Doing various different things each being demanding enough. Render, game, stream, record, decode, using other programs along, web browser with many tabs. 

Would like to see how much it can be pushed doing many different things at once. Can something like a rendering PC a streaming PC and gaming PC be like that one machine.

 

[Hyperwerk]

15 hours ago, Swatson said:

Asus would be a prime candidate to make a nice WS board but they dont seem to want to make any TR2 boards and just ship a heatsink and fan.. They are shooting themselves in the foot me thinks

If you want a legit server/WS board you get something from say Supermicro.  No need for overclocking, and one would appreciate the proper management options etc that are available.
If you want overclocking however something from Asus ,MSI or Asrock should be solid. 

[WereCat]

4 minutes ago, leadeater said:

Glad to see this actually been highlighted by multiple reviews, the amount of flak some people gave me when pointing out memory bandwidth as a problem when increasing cores geez. Time to find that I told you so T-Shirt.

Yeah, for that reason alone the 2950X is a lot more attractive to me than the 2990WX.

Still, imo its better than if they used single channel for each die. They just need to better optimise the scheduling now... or Microsoft will have to come up with some patches for that as well since I am not really sure how much Windows is crippling TR2 as in Linux it seems to fare a lot better.

[Taf the Ghost]

6 minutes ago, leadeater said:

Glad to see this actually been highlighted by multiple reviews, the amount of flak some people gave me when pointing out memory bandwidth as a problem when increasing cores geez. Time to find that I told you so T-Shirt.

I'm pretty sure I was with you on the memory being an issue at scale, though it's seeming like Windows is more of a bottleneck than the memory at the moment.

[Taf the Ghost]

3 minutes ago, WereCat said:

Yeah, for that reason alone the 2950X is a lot more attractive to me than the 2990WX.

Still, imo its better than if they used single channel for each die. They just need to better optimise the scheduling now... or Microsoft will have to come up with some patches for that as well since I am not really sure how much Windows is crippling TR2 as in Linux it seems to fare a lot better.

If your workload didn't already do better on the 1950X over the 7900X, the more cores likely weren't going to help. However, the WX line is clearly the King of the in-office render box.

 

https://www.pugetsystems.com/labs/articles/V-Ray-CPU-Rendering-AMD-Threadripper-2990WX-Takes-the-Single-CPU-Performance-Crown-1214/

 

 

When you have to buy 2-socket Xeon Gold systems to match what a ~3000 USD system can do in a lot of render stuff, it's a really hard sell. There's going to be some very specific buyers for this, but they're going to love it.

[Deus Voltage]

18 minutes ago, VegetableStu said:

I did some thinking...

 

I'm still on the memory topology traffic here (I can't claim to know any of this to save my own life), and I'm starting to think the 1-channel per CCX might be a bad idea because of how much RAM is on one stick and how each CCX has memory control for one particular channel (or 2 in regular real-world cases)

 

say a system has 4x4GB, and each CCX only has 4GB to contend with. Windows typically use 3GB at minimum, so if it's writing those sequentially then only one CCX has the closest access to OS files. if it's spread out, the CCXes has to start exchanging notes quite a lot.

 

And that's before considering workloads that use a lot of RAM. 3/4 of the actual RAM data has to be "paged" from the other, and the CCXes would already be busy with managing their own. Ideally having more RAM per channel and managing tasks in a freer CCX might help but obviously that's not how it works because I'm just a guy on the internet without any means of experimentation

 

I'm already drawing parallels of that thought to the 2950X (and even to the point of considering to buy larger RAM), but I'm kinda glad it's pretty robust for what it is currently despite being less ideal than a unified memory controller situation. I kinda hope threadripper gen 3 somehow does away with CCXes having their own memory controllers and relying on something more unified, but the way it works is being able to use multiple single chips and melding it into a many-core single unit, so...

 

unless all the memory controllers answer to a head librarian but that's another thing to add and I'm overthinking this way too much. can't wait for in-depth reviews and experiments and can't wait for next year's generation

One word (well two actually), the Misaligned ButterDonut topology

Here's the link to the research paper, it's called "Enabling Interposer-based Disintegration of Multi-core Processors"

 

 

[Deus Voltage]

1 minute ago, VegetableStu said:

that was the most "there has to be a maths paper about this" moment I'll ever have. O_O TIL

To be honest, I don't really understand any of this stuff. AdoredTV helped me understand some of the technical nuances, but I'm still an ignoramus when it comes to chip architecture.

[Taf the Ghost]

2 minutes ago, VegetableStu said:

that was the most "there has to be a maths paper about this" moment I'll ever have. O_O TIL

LOL.

 

Path Optimization approaches always get weird. Then you let the math guys name the thing. You're lucky it isn't named after GoT characters. 

 

The solution is Active Interposers. Right now, they're pass throughs, but if you produce them on 45nm node for cheap, you get some massive benefits. 

[PlayStation 2]

The 2990WX showcases Zen's biggest flaw by far: how it handles memory.

I love the Zen architecture to death, but honestly, AMD needs to sort it out really fast if they want to make a viably good 32 core CPU for the prosumer market.

[Taf the Ghost]

Just now, Deus Voltage said:

To be honest, I don't really understand any of this stuff. AdoredTV helped me understand some of the technical nuances, but I'm still an ignoramus when it comes to chip architecture.

It's Network Load issues. In a straight-across patterns, you put too much load on the center part of the network. It becomes a bottleneck. Intel probably is already there, but nothing can leverage the middle of the Mesh hard enough, yet, to do cause too many issues. 

 

AMD's research paper is about balancing the Center and the Edges with as few connection nodes as possible. If you look at the picture carefully, it's actually a basic web design with 2 overlapping squares in the middle to handle the higher bandwidth needed in the central part of the network.

[Taf the Ghost]

2 minutes ago, Dan Castellaneta said:

The 2990WX showcases Zen's biggest flaw by far: how it handles memory.

I love the Zen architecture to death, but honestly, AMD needs to sort it out really fast if they want to make a viably good 32 core CPU for the prosumer market.

Well, you can go buy an Epyc 7501P if you really want the full 32 core utility. Or wait for the Zen2 based ones.

[Deus Voltage]

Just now, VegetableStu said:

I can't claim to know anything as well, and sometimes I type in hopes that someone else who knows better would step in to correct me ,_,

That's what I like about LTT, it has a decent amount of people who actually know how stuff works and are willing to smack us right in our metaphorical faces (and steer us on the right path) with hard numbers, facts and logical reasoning. 

[Taf the Ghost]

4 minutes ago, VegetableStu said:

I can't claim to know anything as well, and sometimes I type in hopes that someone else who knows better would step in to correct me ,_,

There's both Design aspects to this, along with technological limitations. AMD's approach will make more sense the further they divide things. Infinity Fabric already has fairly extensive load-balancing ability. Once DDR5 goes mainstream, feeding the cores becomes far less of a problem. 32 Cores on Quad Channel is just a little too tight, which is why Intel's upcoming A-series will have 6 channel for 28 cores.

[Taf the Ghost]

The other detail is that the 2990WX is really a 2950X with 2 extra dies in "leech" mode. They really shouldn't be doing much, if the Scheduler is working properly, until it hits a load that can activate more than 32 threads. In rendering tasks, which do a lot more throughput than quick calling, the linux testing shows up to 70% improvement over the 2950X. That's the real reason we're seeing these SKUs.

 

The other detail is that AMD is also responsive to the market. If the 7900X was the top SKU, AMD probably wouldn't have brought the 4-die models forward.  However, that's not the current market, so AMD wants to put something in a single socket that's worse than the 32 core Epyc but great for HEDT. For a render box, these parts are unmatched by Intel. If you do workloads that can leverage them, you can build two full systems for the price of 1 Xeon box that can match the power of a 2990WX.

 

Intel, in response, is going to sell some really expensive dies into the HEDT space.

[Brooksie359]

13 minutes ago, Taf the Ghost said:

The other detail is that the 2990WX is really a 2950X with 2 extra dies in "leech" mode. They really shouldn't be doing much, if the Scheduler is working properly, until it hits a load that can activate more than 32 threads. In rendering tasks, which do a lot more throughput than quick calling, the linux testing shows up to 70% improvement over the 2950X. That's the real reason we're seeing these SKUs.

 

The other detail is that AMD is also responsive to the market. If the 7900X was the top SKU, AMD probably wouldn't have brought the 4-die models forward.  However, that's not the current market, so AMD wants to put something in a single socket that's worse than the 32 core Epyc but great for HEDT. For a render box, these parts are unmatched by Intel. If you do workloads that can leverage them, you can build two full systems for the price of 1 Xeon box that can match the power of a 2990WX.

 

Intel, in response, is going to sell some really expensive dies into the HEDT space.

So the long story short is everyone should buy the 2900wx and switch to Linux

[Taf the Ghost]

Just now, Brooksie359 said:

So the long story short is everyone should buy the 2900wx and switch to Linux

Well, if you're doing a lot of rendering, it's hard to beat a 2990WX box. You can, but you're shelling out some impressive amount of money.

[orbitalbuzzsaw]

2 minutes ago, Taf the Ghost said:

Well, if you're doing a lot of rendering, it's hard to beat a 2990WX box. You can, but you're shelling out some impressive amount of money.

I mean an OC'd 7980 can keep up pretty well

[Taf the Ghost]

1 minute ago, OrbitalBuzzsaw said:

I mean an OC'd 7980 can keep up pretty well

It depends on the task.

 

https://www.phoronix.com/scan.php?page=article&item=amd-linux-2990wx&num=6

 

In some, the answer is "not even close". Which is exactly why Intel is running out Server socket to "HEDT" in a few months.