AMD Announces Zen 2/Ryzen 3000 at Computex

[Ampera]

I'm sure everyone's seen it, but by golly do we need to get some new threads up in here!

AMD, at a, beautifully, orchestrated showcase in Taipei just last night (for me), announced their new Ryzen 3000 CPUs. They also announced their 64-core Rome EPYC CPUs, and Navi, their "new" GPU series, which by the numbers doesn't look so amazing to me, and nobody really cares about.

What does look amazing is their new lineup of Ryzen CPUs, including two 8 core 16 thread CPUs, one at 65W SIXTY FIVE WHATS, and another at 105, with 3.6 -> 4.4 and 3.9 -> 4.4 clocks respectively.

Even though that was pretty incredible, not as much as the Ryzen 9 3900X, which albeit not a 16 core CPU like many were expecting, is a 12 core 24 thread monster running at 3.8 -> 4.6 Ghz, at again a 105W TDP.

Other insanities include a reported 40 PCIe 4.0 lanes, with (there's some muddyness about the implementation, but potentially), 80 PCIe 3.0 lanes worth of throughput. There's also a couple of Ryzen 5 CPUs that were supposedly announced, but elsewhere, not in the showcase. Prices for the Ryzen 7 3700X, Ryzen 7 3800X, and Ryzen 9 3900X are 329, 399, and 499 USD respectively.

12 cores 24 threads for half a grand, and apparently it seriously hurts Intel's upper end. We don't have independent benches yet, but it's looking like an all around bad day for Intel, who is still on a, we hope status, for 10nm's dropping in 2020, if it ever drops at all.

Not all is great for AMD, however, particularly its motherboard partners, as these new insane CPUs are apparently going to have iffy support statuses on motherboards from companies like MSI, although not ASUS, who state that all boards from Zen 1 launch to now are going to be perfectly compatible (presumably with BIOS updates) with all Ryzen 3000 chips.

So what are the thoughts? Has AMD repeated 2003 by releasing tech that completely destroys Intel in capabilities, with apparently little recourse? People have also had concerns about AMD's stability, and while I own many many perfectly stable, beautifully running AMD machines, I also own 4 (yes 4) completely dead nForce 4 939 motherboards I had /bought/ to use in a project. Time might only tell.

[Ice-Tea]

Not all is great for AMD, however, particularly its motherboard partners, as these new insane CPUs are apparently going to have iffy support statuses on motherboards from companies like MSI, although not ASUS, who state that all boards from Zen 1 launch to now are going to be perfectly compatible (presumably with BIOS updates) with all Ryzen 3000 chips.

Probably not *perfectly* compatible, as PCIe 4.0 won't be available for "far away" slots. To be honnest, I'm not sure I think it is such as big deal. The high-end first gen AM4 board probably will get updates, the lower-tier maybe not which makes sense as the power delivery there has less headroom. But even if your high-end mobo doesn't allow the upgrade, you can probably buy a lower-tier new gen board that offers the same functionality as features trickle down...

And the notion that at least *some* board will allow upgrades rips a new one over at Team Blue.

[Ice-Tea]

Also: it's kinda funny how there's a hole the size of Uranus in this plot (the 16c/32t part) that is essentially a booby trap waiting to be sprung. As soon as Team Blue gets something out the door that might, maybe, pose a threat AMD launches it watches Intel go down in flames.

They know it, we know it, everybody knows it and there's nothing anyone can do about it 

Also: still rocking a 4750K and have no intention of replacing it.

[Ampera]

If you mean 4790k, that's like the DX2-66 of this generation, everyone has one.

Of course, this is the sort of stuff you have to start complaining about when you're at the top of the mountain, looking down at Intel who still thinks it's okay to disable hyperthreading on even their higher end chips. Intel would have just put a new software incompatible chipset on it, and a socket that just has one incompatible pin that "makes all the difference".

Still, I thought it was worth mentioning.

[BravoV]

Still have my 1700 and feel content about it as I don't do heavy computing or running tons of VMs, and no gaming as well. 

[Ice-Tea]

Also: it's kinda funny how there's a hole the size of Uranus in this plot...

Imagine someone reading this thread using a screen reader .

Eh, well, you gotta enjoy the simple pleasures in life.

And yes 4790k.

[olkipukki]

warm up a Zen topic 

AMD, at a, beautifully, orchestrated showcase in Taipei just last night (for me), announced their new Ryzen 3000 CPUs. They also announced their 64-core Rome EPYC CPUs, and Navi, their "new" GPU series, which by the numbers doesn't look so amazing to me, and nobody really cares about.

I am after EPYC or Threadripper for workstation.

AMD has released not so long time ago, a first high-clocked EPYC 7371.
Trouble is I cannot find a suitable workstation MB. 

Hope that Rome EPYC release bring new stuff on market and a bit of clarify related to workstations.

[Ampera]

Threadripper was really what was intended for workstation use, as the difference between EPYC and Threadripper is really just superficial and marketing. They are mostly the same chips, using very similar, if not identical sockets, designed for different purposes.

I won't say that a dual 64-core EPYC workstation doesn't sound like a wet dream, but perhaps you should look into if it's actually a server you want, if you really have a use for all that power.

[olkipukki]

Threadripper was really what was intended for workstation use, as the difference between EPYC and Threadripper is really just superficial and marketing. They are mostly the same chips, using very similar, if not identical sockets, designed for different purposes.

I have a feeling that Threadripper is half-backed workstation CPU - max 128G support (really?) and not clarify regarding ECC.
It looks more like Intel X-Series competitor... EPYC, on another end, used to be server CPU only until 7371 release...

I won't say that a dual 64-core EPYC workstation doesn't sound like a wet dream, but perhaps you should look into if it's actually a server you want, if you really have a use for all that power.

This is a new workstation to replace existing one and I would like keep it in my possession for 5-7 years minimum, with an option to expand when necessery.

[wraper]

Threadripper was really what was intended for workstation use, as the difference between EPYC and Threadripper is really just superficial and marketing. They are mostly the same chips, using very similar, if not identical sockets, designed for different purposes.

I have a feeling that Threadripper is half-backed workstation CPU - max 128G support (really?) and not clarify regarding ECC.
It looks more like Intel X-Series competitor... EPYC, on another end, used to be server CPU only until 7371 release...

I won't say that a dual 64-core EPYC workstation doesn't sound like a wet dream, but perhaps you should look into if it's actually a server you want, if you really have a use for all that power.

This is a new workstation to replace existing one and I would like keep it in my possession for 5-7 years minimum, with an option to expand when necessery.

Ryzen CPUs have unofficial ECC which will work with unbuffered RAM modules if supported by motherboard (unofficially as well). If you want official ECC, pay for EPYC, AMD will not devour market for their own more expensive products. Say thanks they do not disable ECC like intel does.

[olkipukki]

Threadripper was really what was intended for workstation use, as the difference between EPYC and Threadripper is really just superficial and marketing. They are mostly the same chips, using very similar, if not identical sockets, designed for different purposes.

I have a feeling that Threadripper is half-backed workstation CPU - max 128G support (really?) and not clarify regarding ECC.
It looks more like Intel X-Series competitor... EPYC, on another end, used to be server CPU only until 7371 release...

I won't say that a dual 64-core EPYC workstation doesn't sound like a wet dream, but perhaps you should look into if it's actually a server you want, if you really have a use for all that power.

This is a new workstation to replace existing one and I would like keep it in my possession for 5-7 years minimum, with an option to expand when necessery.

Ryzen CPUs have unofficial ECC which will work with unbuffered RAM modules if supported by motherboard (unofficially as well). If you want official ECC, pay for EPYC, AMD will not devour market for their own more expensive products. Say thanks they do not disable ECC like intel does.

AMD might eventually fill a niche that 100% taken by Xeon W-21xx and upcoming X-32xx.
EPYC has only one option worth consider for workstation-class, this is not convince me to jump on AMD ship yet.

[Ampera]

I have a feeling that Threadripper is half-backed workstation CPU - max 128G support (really?) and not clarify regarding ECC.
It looks more like Intel X-Series competitor... EPYC, on another end, used to be server CPU only until 7371 release...

This is a new workstation to replace existing one and I would like keep it in my possession for 5-7 years minimum, with an option to expand when necessery.

The market segment is HEDT (high end desktop computer). There is ECC support on Ryzen, if you can get a board that supports it, as wraper said.

As for the memory limit, https://en.wikichip.org/wiki/amd/ryzen_threadripper/2990wx states that the memory controller supports up to 2TB for Threadripper over 4 channels. I didn't hear a 128GB limit, and I do not believe that boards even impose a limit upon that, though I could be wrong. Perhaps that limit is for AM4 boards/chips, but https://en.wikichip.org/wiki/amd/ryzen_7/2700x the 2700X is listed as 64GB. Maybe some other chips have more, maybe they don't I haven't looked.

Spend what you want, if you like EPYC, that's awesome, I'd definitely like a system like that, my point is that the price to performance is probably going to be in the shitter for a system like that, so Threadripper might be a more down to earth option. Power to you, regardless.

[olkipukki]

There is ECC support on Ryzen, if you can get a board that supports it, as wraper said.

We have issue problem #1 here, why AMD cannot publish it on own website in the product page ECC: Yes/No/Depends from mobo?

As for the memory limit, https://en.wikichip.org/wiki/amd/ryzen_threadripper/2990wx states that the memory controller supports up to 2TB for Threadripper over 4 channels. I didn't hear a 128GB limit, and I do not believe that boards even impose a limit upon that, though I could be wrong. Perhaps that limit is for AM4 boards/chips, but https://en.wikichip.org/wiki/amd/ryzen_7/2700x the 2700X is listed as 64GB. Maybe some other chips have more, maybe they don't I haven't looked.

Here, we have issue problem #2, similar to #1, AMD didn't publish the details at all.

Why there's so secretive or perhaps carelessly?

To be honest, I don't know exactly max support memory for Threadripper, just looked into, for example, the following mobo:

https://www.asus.com/us/Motherboards/ROG-ZENITH-EXTREME/specifications/
https://www.msi.com/Motherboard/MEG-X399-CREATION/Specification
https://www.asrock.com/MB/AMD/X399%20Taichi/index.asp#Specification
etc.

Can you find any board that support more than 128GB?

Also, UDIMM ECC is not wise to use, limited availability and choice compare to RDIMM.

Spend what you want, if you like EPYC, that's awesome, I'd definitely like a system like that, my point is that the price to performance is probably going to be in the shitter for a system like that, so Threadripper might be a more down to earth option. Power to you, regardless.

Not yet, still in Intel club, but EPYC/Threadripper is a contender

[wraper]

We have issue problem #1 here, why AMD cannot publish it on own website in the product page ECC: Yes/No/Depends from mobo?

Because it is there but not officially supported  . "No" simply won't be true. They won't say "yes". And "depends" is simply confusing. If you want to use ECC, it's on you. No guarantees from AMD. It's certainly not a feature they want to advertise.

[wraper]

Also, UDIMM ECC is not wise to use, limited availability and choice compare to RDIMM.

There is plenty of it. It's just if you after used stuff, it's harder and more expensive to get. On the other hand it's just as fast as usual non ECC RAM, and does not have performance penalty of registered modules.

[olkipukki]

Can you find any board that support more than 128GB?

OK, found one, TYAN S8020 (S8020AGM2NR-EX) https://www.tyan.com/Motherboards_S8020_S8020AGM2NR-EX if you can find 32GB UDIMMs...

[olkipukki]

Also, UDIMM ECC is not wise to use, limited availability and choice compare to RDIMM.

There is plenty of it. It's just if you after used stuff, it's harder and more expensive to get. On the other hand it's just as fast as usual non ECC RAM, and does not have performance penalty of registered modules.

Would RDIMM bandwidth will higher than UDIMM for N (?) number of DIMMs per channel?

Anyway, RDIMM market much bigger...

[olkipukki]

We have issue problem #1 here, why AMD cannot publish it on own website in the product page ECC: Yes/No/Depends from mobo?

Because it is there but not officially supported  . "No" simply won't be true. They won't say "yes". And "depends" is simply confusing. If you want to use ECC, it's on you. No guarantees from AMD. It's certainly not a feature they want to advertise.

This is a bizarre approach.

[wraper]

We have issue problem #1 here, why AMD cannot publish it on own website in the product page ECC: Yes/No/Depends from mobo?

Because it is there but not officially supported  . "No" simply won't be true. They won't say "yes". And "depends" is simply confusing. If you want to use ECC, it's on you. No guarantees from AMD. It's certainly not a feature they want to advertise.

This is a bizarre approach.

What bizarre? You want them to go intel path and just disable it?

[Ampera]

Guys, you know you can edit posts, right? No need to do multiple posts in a row.

If having everything it does on the tin what says it does it is a selling feature, then by golly is Intel your company, they love telling you exactly what you can't do with your product because you didn't pay enough for them to unlock it for you.

I'm not sure how memory is implemented on TR4 boards, but I'd look into if the memory size limitations actually means anything. I'd not be surprised if it didn't.

As for RDIMM vs UDIMM, I wouldn't be complaining about needing UDIMM ECC, you wanted a workstation, that /is/ workstation memory. Again, are you sure you do not want a server that you can offload tasks to, because then EPYC as a platform makes more sense, at least in my mind. As for the price gap, it's not massive, and it's not like you're trying to scrim every single ounce of power out of a 500 dollar budget. This stuff costs money.

AMD publishes the stuff they want you to know about, and I'd say they already give you /more/ than what Intel offers in the same price bracket, for what they advertise. Them not purposefully disabling features like ECC, and just not advertising it is actually something I wish Intel did, because it makes me feel more like I'm buying a machine that can be used to its full potential, and not a lame horse.

[Mr. Scram]

The problem with ECC being unsupported is that you use ECC for some extra peace of mind. Having it not supported means taking a gamble you were actually trying to avoid. Of course, none of that is on AMD.

[Ampera]

The problem with ECC being unsupported is that you use ECC for some extra peace of mind. Having it not supported means taking a gamble you were actually trying to avoid. Of course, none of that is on AMD.

Personally I can only see it being a practical benefit when you start to scale computing. Maybe having a load of workstations in an office counts as scaling, but I've run my desktop PC without ECC memory, and I've run PCs with ECC memory, and I'm not finding myself loosing critical information or system stability by any means, on either computer. Anecdotally, I wouldn't expect to see anything unless I have hundreds of servers running 24/7, and with incredibly critical data that can't have so much as a bit changed anywhere ever.

As for AMD, ECC is either supported or it's not afaik, having undocumented support wouldn't, in my mind, cause it to be less capable of correcting random errors.

[Mr. Scram]

Personally I can only see it being a practical benefit when you start to scale computing. Maybe having a load of workstations in an office counts as scaling, but I've run my desktop PC without ECC memory, and I've run PCs with ECC memory, and I'm not finding myself loosing critical information or system stability by any means, on either computer. Anecdotally, I wouldn't expect to see anything unless I have hundreds of servers running 24/7, and with incredibly critical data that can't have so much as a bit changed anywhere ever.

As for AMD, ECC is either supported or it's not afaik, having undocumented support wouldn't, in my mind, cause it to be less capable of correcting random errors.

It's not just about the errors you see. If you run an CAD application and are doing calculations on which lives depend, you want to eliminate bits falling over. Whether it's sensible depends on your workload. Of course, as the amount of data we process increases the statistical likelihood of issues increases and there may be a point at which ECC becomes inevitable. We're already seeing this issue in storage, which is why certain RAID levels aren't really viable any more.

[TheAmmoniacal]

ECC works fine on Ryzen, official support from AMD is meaningless. There are many "workstation" motherboards for Ryzen that specifically advertise ECC memory support, and many ECC modules are on the QVL.

[Mr. Scram]

ECC works fine on Ryzen, official support from AMD is meaningless. There are many "workstation" motherboards for Ryzen that specifically advertise ECC memory support, and many ECC modules are on the QVL.

Official support is far from meaningless if you actually have a use for it.