AMD ThreadRipper Gen 2 Reviews

[The Benjamins]

16 minutes ago, Taf the Ghost said:

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.

This article shows issues with Windows and the 2990wx, a windows patch can help it a lot.

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

[Taf the Ghost]

9 minutes ago, The Benjamins said:

This article shows issues with Windows and the 2990wx, a windows patch can help it a lot.

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

Handbrake and any game seem to be handling the situation incorrectly. Plus WinRaR/7zip.

[leadeater]

4 hours 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.

Really the problem it shows is more unique to the 2990WX, the EPYC processors with 8 memory channels doesn't have issues like the 2990WX does. The anandtech review actually has an EPYC in it so if you look at that you'll see it doesn't have wild performance drops like the 2990WX does.

[MMKing]

I believe what happened when AMD decided to release the TR4 platform with only quad channel memory support, was that they did not believe Intel would respond with the 10-18 core lineup. So we are in this weird situation, where the top end CPU in the threadripper series cost 1800 dollars and paying for 6-8 memory chips is not a big deal, but on the lower end they are selling the 12 core option for 650 dollars, where paying for 6-8 chips would almost cost as much as the CPU itself. Quad channel was the ideal middle ground for 8-12-16 core, not so much when you add 24 and 32 on top.

 

If AMD want to release another 32 core CPU for Zen2 is support for more memory channels. Ideally for the consumers and AMD, DDR5 is released for the Zen2 Threadripper platform. The consumers get better performance, and AMD can move on to a new platform without backlash.

[Marinatall_Ironside]

1 hour ago, leadeater said:

Really the problem it shows is more unique to the 2990WX, the EPYC processors with 8 memory channels doesn't have issues like the 2990WX does. The anandtech review actually has an EPYC in it so if you look at that you'll see it doesn't have wild performance drops like the 2990WX does.

AMD could have also kept compatibility with existing X399 motherboards by wiring up one memory channel to each die. This way, you can get quad-channel memory with probably zero issues regarding support.

 

Sure, the per-die memory bandwidth would be halved, but when you hit all the dies with a memory-intensive task, it should have higher sustained memory bandwidth and not completely go to the crapshoot like what we're seeing now.

[leadeater]

46 minutes ago, Jurunce said:

AMD could have also kept compatibility with existing X399 motherboards by wiring up one memory channel to each die. This way, you can get quad-channel memory with probably zero issues regarding support.

 

Sure, the per-die memory bandwidth would be halved, but when you hit all the dies with a memory-intensive task, it should have higher sustained memory bandwidth and not completely go to the crapshoot like what we're seeing now.

That would require new motherboards as the memory slots aren't wired that way, short of doing some special magic in the pin outs of the CPU or an extra interposer to remap the slots that would break drop in compatibility with existing motherboards.

[leadeater]

1 hour ago, MMKing said:

I believe what happened when AMD decided to release the TR4 platform with only quad channel memory support, was that they did not believe Intel would respond with the 10-18 core lineup. So we are in this weird situation, where the top end CPU in the threadripper series cost 1800 dollars and paying for 6-8 memory chips is not a big deal, but on the lower end they are selling the 12 core option for 650 dollars, where paying for 6-8 chips would almost cost as much as the CPU itself. Quad channel was the ideal middle ground for 8-12-16 core, not so much when you add 24 and 32 on top.

 

If AMD want to release another 32 core CPU for Zen2 is support for more memory channels. Ideally for the consumers and AMD, DDR5 is released for the Zen2 Threadripper platform. The consumers get better performance, and AMD can move on to a new platform without backlash.

AMD would have known about the 18 core stuff, that wasn't unplanned by Intel. The 28 core one coming is a different story though, price for that is going to be extra spicy though lol.

 

I don't think AMD planned to have 4 active die SKUs on X399 and that only exists due to Intel planning on releasing HEDT CPUs above the current 18 core option.

[MMKing]

34 minutes ago, leadeater said:

AMD would have known about the 18 core stuff, that wasn't unplanned by Intel. The 28 core one coming is a different story though, price for that is going to be extra spicy though lol.

 

I don't think AMD planned to have 4 active die SKUs on X399 and that only exists due to Intel planning on releasing HEDT CPUs above the current 18 core option.

All the leaks and rumors leading up to the 18 core reveal at Computex 2017, showed that Intel's new top consumer CPU would only have 12 cores. Literally no one in the tech press even mentions the Intel 18 core in any of the Threadripper 1 articles.

 

I misspoke regarding the notion that AMD did not believe Intel would release a 10-18 core option, what i meant was 12-18 core. After doing some research, i see that i should have written 14-18. My point being that AMD believed a 16 core CPU would hold the cost and performance lead in the consumer market until Zen2. So there was no reason to design the TR4 platform to handle 32 cores / 4 dies on the infinity fabric. The ''WX'' name scheme is an effort on AMDs part to differentiate the 24 and the 32 core, making it clear they are very much niche CPUs.

 

Here is the slide leaked to the Anandtech forums May 2017, 4 days before the Threadripper announcement.

https://forums.anandtech.com/threads/intel-skylake-kaby-lake.2428363/page-417

Spoiler

 

Here is the Computex slide. Notice there is no information at all regarding the 7920x, 7940x, 7960x or the 7980XE. I don't think even Intel knew anything about their own product at this time, it was simply too early in development. Which is part of the reason why the 12-18 core options are so similar to various Xeon CPUs, because they are. Remove a few lanes, support 4 instead of 6 channels, drop ECC, allow overclock and add Turbo Boost technology, then you got a consumer CPU.

Spoiler

 

''I don't think AMD planned to have 4 active die SKUs on X399 and that only exists due to Intel planning on releasing HEDT CPUs above the current 18 core option.''

 

This we fully agree on. We don't seem to agree on the reasoning however.

[leadeater]

4 minutes ago, MMKing said:

All the leaks and rumors leading up to the 18 core reveal at Computex 2017, showed that Intel's new top consumer CPU would only have 12 cores. Literally no one in the tech press even mentions the Intel 18 core in any of the Threadripper 1 articles.

 

 

Here is the slide leaked to the Anandtech forums May 2017, 4 days before the Threadripper announcement.

https://forums.anandtech.com/threads/intel-skylake-kaby-lake.2428363/page-417

  Reveal hidden contents

 

Here is the Computex slide. Notice there is no information at all regarding the 7920x, 7940x, 7960x or the 7980XE. I don't think even Intel knew anything about their own product at this time, it was simply too early in development. Which is part of the reason why the 12-18 core options are so similar to various Xeon CPUs, because they are. Remove a few lanes, support 4 instead of 6 channels, drop ECC, allow overclock and add Turbo Boost technology, then you got a consumer CPU.

  Reveal hidden contents

 

We might not have known but competing companies generally know what each other is doing far in advance, there is also the fact that Intel has always traditionally based HEDT CPUs off the Xeon architectures so if there is something over there that can do 18+ cores then it can exist on HEDT. Intel doesn't normally take the biggest HCC Xeon die and bring it down to HEDT though, the coming 28 core SKU is the first time since Intel made that split in dies on Xeons.

[GOTSpectrum]

1 hour ago, leadeater said:

We might not have known but competing companies generally know what each other is doing far in advance, there is also the fact that Intel has always traditionally based HEDT CPUs off the Xeon architectures so if there is something over there that can do 18+ cores then it can exist on HEDT. Intel doesn't normally take the biggest HCC Xeon die and bring it down to HEDT though, the coming 28 core SKU is the first time since Intel made that split in dies on Xeons.

Its going to be real interesting to see the price... Its monolithic, Intel are limited on just how cheap they can make it, add that too they dont want it so cheap that people are willing to risk buying that instead of the xeon for mission critical stuff. 

[Taf the Ghost]

5 hours ago, leadeater said:

AMD would have known about the 18 core stuff, that wasn't unplanned by Intel. The 28 core one coming is a different story though, price for that is going to be extra spicy though lol.

 

I don't think AMD planned to have 4 active die SKUs on X399 and that only exists due to Intel planning on releasing HEDT CPUs above the current 18 core option.

I actually think AMD did plan for moving to the 4-active dies, the issue is just Windows on launch. Much in the same way Threadripper 1 had issues in a bunch of programs because of the design. Much in the way Ryzen had issues in a bunch of programs because of the design.

 

This is really the first time the Windows Scheduler has been really forced to deal with a big.LITTLE design topology, and it's thrown up a few wild pitches as a result. Considering all of the problematic testing results go away when you disable the extra 2 dies, it's clearly just a Scheduler problem. I actually think the Memory Bandwidth isn't much of an issue, except that you're not getting as much out of TR2 as you can an Epyc CPU. Which isn't surprising giving the price & platform differences. With a properly working scheduler, for anyone with the type of workloads that can leverage 32 cores, the 2990WX is unbeatable at the price point. You have to pay Intel & partners a lot more to meet the rendering performance.

[Taf the Ghost]

4 hours ago, leadeater said:

We might not have known but competing companies generally know what each other is doing far in advance, there is also the fact that Intel has always traditionally based HEDT CPUs off the Xeon architectures so if there is something over there that can do 18+ cores then it can exist on HEDT. Intel doesn't normally take the biggest HCC Xeon die and bring it down to HEDT though, the coming 28 core SKU is the first time since Intel made that split in dies on Xeons.

 

3 hours ago, Ben Quigley said:

Its going to be real interesting to see the price... Its monolithic, Intel are limited on just how cheap they can make it, add that too they dont want it so cheap that people are willing to risk buying that instead of the xeon for mission critical stuff. 

Intel is bringing list-price 3k to 10k dies from the Server space to HEDT. They're going to be incredibly expensive. Standard Intel pricing would put the 28c between 3999 and 4499. Plus you need 6-channel memory for it.

[JuNex03]

2 hours ago, Taf the Ghost said:

 

Intel is bringing list-price 3k to 10k dies from the Server space to HEDT. They're going to be incredibly expensive. Standard Intel pricing would put the 28c between 3999 and 4499. Plus you need 6-channel memory for it.

Ouch 3999>4499 for a chip?

[Taf the Ghost]

3 minutes ago, JuNex03 said:

Ouch 3999>4499 for a chip?

The Xeon Platinums, that the 28c are, go for 9000-12000 MSRP. Normal sale price is less by a bit, but, yeah, Intel charges a lot for them.

[leadeater]

2 hours ago, Taf the Ghost said:

The Xeon Platinums, that the 28c are, go for 9000-12000 MSRP. Normal sale price is less by a bit, but, yeah, Intel charges a lot for them.

Well technically people who buy those don't actually pay that. We don't pay list for the quad socket E7-8890v4 servers we get, same will apply to Xeon Plat equiv 8176.

[GOTSpectrum]

5 hours ago, Taf the Ghost said:

 

Intel is bringing list-price 3k to 10k dies from the Server space to HEDT. They're going to be incredibly expensive. Standard Intel pricing would put the 28c between 3999 and 4499. Plus you need 6-channel memory for it.

Exactly, But what i'm wondering is will the price difference be worth it, as in compared to TR2. I think its should be a hard sell, but it probably wont be because Intel. 

[Taf the Ghost]

16 minutes ago, leadeater said:

Well technically people who buy those don't actually pay that. We don't pay list for the quad socket E7-8890v4 servers we get, same will apply to Xeon Plat equiv 8176.

Right, the integrated price is a lot less, which does still lead me to wonder why they list MSRP at that level. Accounting reasons, I'd suppose.

[Taf the Ghost]

3 minutes ago, Ben Quigley said:

Exactly, But what i'm wondering is will the price difference be worth it, as in compared to TR2. I think its should be a hard sell, but it probably wont be because Intel. 

I can see some Render Box approaches that could use the upcoming A-series. Especially the guys with actual AVX512 loads to run. It'll also take more memory (192 Gb vs 128 Gb), so I can see a use case for a ~6k USD AVX Box. A lot easier to convince to get for certain Education or smaller Business stuff, compared to a much more expensive 2U server.

 

I expect a 28 core Intel to be around 25% faster than a 32 core Threadripper 2990WX in normal render work loads. In heavier AVX, it's going to be higher, but it's also going to use a good chunk more power and probably be double the cost per box.

[GOTSpectrum]

3 minutes ago, Taf the Ghost said:

I can see some Render Box approaches that could use the upcoming A-series. Especially the guys with actual AVX512 loads to run. It'll also take more memory (192 Gb vs 128 Gb), so I can see a use case for a ~6k USD AVX Box. A lot easier to convince to get for certain Education or smaller Business stuff, compared to a much more expensive 2U server.

 

I expect a 28 core Intel to be around 25% faster than a 32 core Threadripper 2990WX in normal render work loads. In heavier AVX, it's going to be higher, but it's also going to use a good chunk more power and probably be double the cost per box.

The real question is, just how much more power, If you can get two 32core TR2s for the price and power budget of one intel box and you dont have much need for AVX.... AMD being competitive has made some really strange things happen haha. 

[Taf the Ghost]

1 minute ago, Ben Quigley said:

The real question is, just how much more power, If you can get two 32core TR2s for the price and power budget of one intel box and you dont have much need for AVX.... AMD being competitive has made some really strange things happen haha. 

At stock, 2x 2990WX should do about 10k Cinebench R15 points. The top 28c should come in around 6200-6400. So it really depends on what the price comes in at. On the Power Draw, it's pretty likely to be about even between 2x 2990WX under load compared to the A-series 28w, at stock with both pegged. On the assumption you have adequate cooling for the Intel to turbo up pretty high.

[leadeater]

6 hours ago, Taf the Ghost said:

I actually think AMD did plan for moving to the 4-active dies, the issue is just Windows on launch. Much in the same way Threadripper 1 had issues in a bunch of programs because of the design. Much in the way Ryzen had issues in a bunch of programs because of the design.

 

This is really the first time the Windows Scheduler has been really forced to deal with a big.LITTLE design topology, and it's thrown up a few wild pitches as a result. Considering all of the problematic testing results go away when you disable the extra 2 dies, it's clearly just a Scheduler problem. I actually think the Memory Bandwidth isn't much of an issue, except that you're not getting as much out of TR2 as you can an Epyc CPU. Which isn't surprising giving the price & platform differences. With a properly working scheduler, for anyone with the type of workloads that can leverage 32 cores, the 2990WX is unbeatable at the price point. You have to pay Intel & partners a lot more to meet the rendering performance.

I doubt it, Zen and IF we designed around each node having it's own memory controller and PCIe controller which the 2990WX breaks. Also it's a rather poor decision to make the 2990WX UMA only and no NUMA mode, there are two ways they could have done that too. NUMA mode on the 2990WX could be Node 0 both dies with memory and Node 1 both dies without memory (Windows preferences Node 0 first) or made Node 0 one set of linked dies and Node 1 the other set, hell having the option to set these 3 modes/combinations would have been amazing but no UMA only.

 

Also disabling the extra dies doesn't show it's a scheduler issue, it just removes the issue completely so you're not actually testing anything. A better test would be to disable all or all but 1 core on the primary dies then run the benchmarks with 8 or 8+1 (16 or 16+2) remote cores. When you're so memory constrained it effects all the cores greatly so that's why results lower than 2950X show up.

 

Now I'm not saying it's not a scheduler issue but rather it's not only a scheduler issue, EPYC 7601 scores in cases slightly higher than the 2990WX with lower clocks. However on that topic there is a very big problem I've noticed with Phoronix and there testing, IT'S NEVER THE BLOODY SAME! Like omg I went to compare their EPYC 7601 review to the 2990WX and totally different blender scenes, come on standardize.

[leadeater]

9 minutes ago, Taf the Ghost said:

The top 28c should come in around 6200-6400

If you're talking that Intel 28 core then stock probably won't be that high, the LTT video showing the custom water loop one was getting 6k and that was also OC (just not stupid OC).

[Taf the Ghost]

6 minutes ago, leadeater said:

If you're talking that Intel 28 core then stock probably won't be that high, the LTT video showing the custom water loop one was getting 6k and that was also OC (just not stupid OC).

True. Though we don't know what final clocks they'll allow as a baseline turbo.

[Taf the Ghost]

10 minutes ago, leadeater said:

I doubt it, Zen and IF we designed around each node having it's own memory controller and PCIe controller which the 2990WX breaks. Also it's a rather poor decision to make the 2990WX UMA only and no NUMA mode, there are two ways they could have done that too. NUMA mode on the 2990WX could be Node 0 both dies with memory and Node 1 both dies without memory (Windows preferences Node 0 first) or made Node 0 one set of linked dies and Node 1 the other set, hell having the option to set these 3 modes/combinations would have been amazing but no UMA only.

 

Also disabling the extra dies doesn't show it's a scheduler issue, it just removes the issue completely so you're not actually testing anything. A better test would be to disable all or all but 1 core on the primary dies then run the benchmarks with 8 or 8+1 remote cores. When you're so memory constrained it effects all the cores greatly so that's why results lower than 2950X show up.

 

Now I'm not saying it's not a scheduler issue but rather it's not only a scheduler issue, EPYC 7601 scores in cases slightly higher than the 2990WX with lower clocks. However on that topic there is a very big problem I've noticed with Phoronix and there testing, IT'S NEVER THE BLOODY SAME! Like omg I went to compare their EPYC 7601 review to the 2990WX and totally different blender scenes, come on standardize.

Guy disabled SMT on the 2990WX and Handbrake stopped being a spaz, so I think that one specifically is specifically an issue with the number of threads & Windows.

 

Games probably need more of just being hit with Process Lasso, so they're kept off the leech dies.

[leadeater]

10 minutes ago, Taf the Ghost said:

Guy disabled SMT on the 2990WX and Handbrake stopped being a spaz, so I think that one specifically is specifically an issue with the number of threads & Windows.

 

Games probably need more of just being hit with Process Lasso, so they're kept off the leech dies.

I also just found this, maybe Phoronix just isn't running the correct versions of software for the test?

 

 

 

18.01 on Windows seems to work fine...