AMD Ryzen Threadripper 1900X Launch A Threadripper on a Budget

[Cliff Robinson]

AMD Ryzen Threadripper 1900X is all about delivering the Threadripper platform while spending less on the CPU for the HEDT market

The post AMD Ryzen Threadripper 1900X Launch A Threadripper on a Budget appeared first on ServeTheHome.

Continue reading...

 

[gigatexal]

They have to lower the price of the r7 1800X now right? Unless the gulf between the ryzen mobos and threadripper ones is so great they don’t see them as competitors.

I think I’m going to wait until these are a out 6 months out or for the next price cut. That or start putting some coin aside for the 999 top end TR4 part so that I can consolidate my “mini lab”

 

[TType85]

Interesting dilemma here. I have a Microcenter close so I can get these CPU's at a good price.
1900x = $549
7820X = $549

Motherboards are similar in price. Both would give an upgrade path to more cores in the future. (8 cores fits what I need now).

 

[Patrick]

Interesting dilemma here. I have a Microcenter close so I can get these CPU's at a good price.
1900x = $549
7820X = $549

Motherboards are similar in price. Both would give an upgrade path to more cores in the future. (8 cores fits what I need now).

Same core count and price, get the single die unless you need the extra PCIe lanes.

 

[TType85]

Same core count and price, get the single die unless you need the extra PCIe lanes.

Yeah, that's where the TR pulls ahead. For the linux workstation with KVM pass through I am looking at doing it is looking like this:
16x GPU (Passed to Windows)
4x m.2 nvme
8x 10GB PCIe card
4x GPU (GPU for Linux)
1x USB 3 PCIe card (Passed to Windows)

I was going to go with the 12 core, and at $200 more it still may make more sense. The lack of lanes and 4 ram slots knocks the R7 out.

 

[toomuchstuff]

Thanks for giving me a good reason to hold onto this ECC DDR4. Looking forward to these or the EPYC now.

 

[gigatexal]

Yeah, that's where the TR pulls ahead. For the linux workstation with KVM pass through I am looking at doing it is looking like this:
16x GPU (Passed to Windows)
4x m.2 nvme
8x 10GB PCIe card
4x GPU (GPU for Linux)
1x USB 3 PCIe card (Passed to Windows)

I was going to go with the 12 core, and at $200 more it still may make more sense. The lack of lanes and 4 ram slots knocks the R7 out.

I’m lookin forward to your write up on this build. I’d like to do the same. Curious on how you fare with the IOMMUs and GPU passthrough.

 

[TType85]

I’m lookin forward to your write up on this build. I’d like to do the same. Curious on how you fare with the IOMMUs and GPU passthrough.

Looking at the latest Level1Techs linux video, TR is looking a lot better than Ryzen 7. Ryzen with passthrough was disappointing. Getting all the parts may take a little time though. I have to shell out money to run a thicker gauge wire to my pool sub-panel for the spa we got.

 

[zir_blazer]

Looking at the latest Level1Techs linux video, TR is looking a lot better than Ryzen 7. Ryzen with passthrough was disappointing. Getting all the parts may take a little time though. I have to shell out money to run a thicker gauge wire to my pool sub-panel for the spa we got.

Ryzen with Passthrough is dissapointing because there is a decade old bug with NPT (Nested Page Tables) that absolutely kills performance when using QEMU-KVM, which is the most popular option today. TR will probably suffer from it, too, given the fact that even old Bulldozer is affected.
A guy tested with Xen (That has its own set of cons, since for Passthrough is inferior due to no compatibility with GeForces. There is a long history going on with nVidia antagonizing passthrough users) and it worked as expected. Basically, the NPT issue is assumed to be a Software fixable bug, but no one knows if AMD is going to get serious about fixing it soon enough or we have to put some pressure...
I think I also saw someone complaining that ThreadRipper produced a lot of IOMMU errors when trying to use Passthrough. I don't know if there are people that managed to get it working successfully, but didn't really looked into it, either.

Another problem that both TR and EPYC have is that they need very detailed Block Diagrams to get the most out of these platforms. I recently complained that not even Supermicro is providing enough details. Since you have 2 integrated PCIe Controllers in TR and 4 in EPYC, you want that the Video Card is in a slot that is wired to the same die that will provide the Cores for your VM. TR is capable of providing pretty much two independent computers in the same case.
I'm curious if there are pins that exposes the other dies Azalia Buses, as a Motherboard with two integrated Sound Cards could be interesing for this purpose.


I'm personally interesed in benchmarks to see how a 2+2 CCX * 2 MCM works when compared to the single 4+4 CCX from the Ryzen 7 1700 and higher models. Same with the 8-Core EPYC that was 1+1 CCX * 4 MCM. I know that they are going to be inferior due to interdie latency, but is important to measure that so you know how much performance you are sacrificing if you want to get to the wider platforms without breaking the bank by staying in the 8 Core models.

 

[gigatexal]

Ryzen with Passthrough is dissapointing because there is a decade old bug with NPT (Nested Page Tables) that absolutely kills performance when using QEMU-KVM, which is the most popular option today. TR will probably suffer from it, too, given the fact that even old Bulldozer is affected.
A guy tested with Xen (That has its own set of cons, since for Passthrough is inferior due to no compatibility with GeForces. There is a long history going on with nVidia antagonizing passthrough users) and it worked as expected. Basically, the NPT issue is assumed to be a Software fixable bug, but no one knows if AMD is going to get serious about fixing it soon enough or we have to put some pressure...
I think I also saw someone complaining that ThreadRipper produced a lot of IOMMU errors when trying to use Passthrough. I don't know if there are people that managed to get it working successfully, but didn't really looked into it, either.

Another problem that both TR and EPYC have is that they need very detailed Block Diagrams to get the most out of these platforms. I recently complained that not even Supermicro is providing enough details. Since you have 2 integrated PCIe Controllers in TR and 4 in EPYC, you want that the Video Card is in a slot that is wired to the same die that will provide the Cores for your VM. TR is capable of providing pretty much two independent computers in the same case.
I'm curious if there are pins that exposes the other dies Azalia Buses, as a Motherboard with two integrated Sound Cards could be interesing for this purpose.


I'm personally interesed in benchmarks to see how a 2+2 CCX * 2 MCM works when compared to the single 4+4 CCX from the Ryzen 7 1700 and higher models. Same with the 8-Core EPYC that was 1+1 CCX * 4 MCM. I know that they are going to be inferior due to interdie latency, but is important to measure that so you know how much performance you are sacrificing if you want to get to the wider platforms without breaking the bank by staying in the 8 Core models.

Well shoot. I plan on using proxmox for my build but that’s KVM based.

 

[realtomatoes]

Ryzen with Passthrough is dissapointing because there is a decade old bug with NPT (Nested Page Tables) that absolutely kills performance when using QEMU-KVM, which is the most popular option today. TR will probably suffer from it, too, given the fact that even old Bulldozer is affected.
A guy tested with Xen (That has its own set of cons, since for Passthrough is inferior due to no compatibility with GeForces. There is a long history going on with nVidia antagonizing passthrough users) and it worked as expected. Basically, the NPT issue is assumed to be a Software fixable bug, but no one knows if AMD is going to get serious about fixing it soon enough or we have to put some pressure...
I think I also saw someone complaining that ThreadRipper produced a lot of IOMMU errors when trying to use Passthrough. I don't know if there are people that managed to get it working successfully, but didn't really looked into it, either.

would you happen to know if the same NPT issue exist on ESXi?

 

[Patrick]

@zir_blazer working on it. We now have every EPYC SKU.

 

[realtomatoes]

@zir_blazer working on it. We now have every EPYC SKU.

I bet that article is gonna be, wait for it...


Sent from my iPhone using Tapatalk

 

[Patrick]

I bet that article is gonna be, wait for it...

Working on Xeon W as well.

But as of tomorrow, we will have 38% of all 1P and 2P Skylake-SP and EPYC test cases covered by chips in the lab.

We are going to have a lot more of these Intel Xeon Silver 4112 Linux Benchmarks and Review

Have a dozen (or more) systems doing regressions on older E5 V3/ V4 CPUs as well and have most of the Xeon D and soon C3000 series online to compare.

I looked at it, and STH is going to have to move from an article Monday - Friday to 7 days a week in September and October just to keep up.

EPYC we are limited not by CPUs, RAM, or disks. Dates for systems keep getting delayed.

On the Skylake side, in the last 24 hours we added three new 1U skylakes and 2x new motherboards to the stable just to keep pace.

 

[gigatexal]

Curious to see what STH’s take on the NPT slowdown.

 

[zir_blazer]

would you happen to know if the same NPT issue exist on ESXi?

I know nothing about VMware products, my focus is on QEMU-KVM-VFIO and used to be Xen.


The TR 1900X just got even MORE interesing. Check this TechPowerUp article.

At its pre-launch media conference call for the Ryzen Threadripper 1900X, AMD mentioned that the chip has been carved out from the common 4-die EPYC MCM using a "4-0-4-0 diagonal configuration," which led to some confusion as to which cores/dies AMD disabled to carve out the $549 8-core HEDT processor. The company shed some light on this matter, responding to questions from TechPowerUp.

It turns out, that the Threadripper 1900X features an entire CCX (quad-core CPU complex) disabled per active die on the multi-chip module, so the CCX that's enabled has 8 MB of L3 cache; and access to the die's entire uncore resources, such as the dual-channel memory controller, PCIe root complex, etc. With two such active "Zeppelin" dies, the Threadripper 1900X ends up with 8 cores, 16 MB of L3 cache, a quad-channel memory interface, and 64 PCIe lanes.

This put it almost on par to the standard Ryzen! Instead of intradie CCX latency, you got interdie latency, but the 4+0-4+0 looks extremely clean and better than the 2+2-2+2. I think that if you are going to spend 500 U$D in a Processor, the R7 1800X makes no sense with this beauty for just the extra cost of the Motherboard.

 

[gigatexal]

Exciting times indeed. Will be curious to see what STH finds with iommu and such.

 

[Patrick]

@zir_blazer Intra-die on Zen is very good. Inter-die is the painful one. Intra-die from one CCX to another you notice more in cache benchmarks than in real-world performance. Inter-die is another animal since it also introduces NUMA.

2-2-2-2 would have made zero sense given how AMD's architecture works. For the 7251 it is a bit different since they still need to light all of the RAM and PCIe.

Another (very) important aspect to remember is that TR only has one Infinity Fabric link die to die. In EPYC, there is one direct and two two hop routes on each package.

 

[cactus]

@zir_blazer Intra-die on Zen is very good. Inter-die is the painful one. Intra-die from one CCX to another you notice more in cache benchmarks than in real-world performance. Inter-die is another animal since it also introduces NUMA.

2-2-2-2 would have made zero sense given how AMD's architecture works. For the 7251 it is a bit different since they still need to light all of the RAM and PCIe.

Another (very) important aspect to remember is that TR only has one Infinity Fabric link die to die. In EPYC, there is one direct and two two hop routes on each package.

On TR with 2 dies, 2 cores enabled in each CCX to get 8 cores would follow what I believe they did with the R3/R5 lines. Disabling a full CCX per die is interesting for the 1900x and gives you two latency tiers vs three. Are the 8 and 16 core EPYCs doing 1 or 2 cores per CCX or disabling a full CCX per die?