AMD Announces Ryzen Threadripper Pro: Workstation Parts for OEMs Only
anandtech.com
anandtech.com
There are higher base frequencies but lower turbo on the new TR Pro vs Ryzen 9:
| name | CPUs | base | turbo|
| 3955WX | 16/32 | 3900 | 4300 |
| 3945WX | 12/24 | 4000 | 4300 |
| 3950X | 16/32 | 3500 | 4700 |
| 3900XT | 12/24 | 3800 | 4700 |
The TR 3970x in particular is amazing for a server. It's 32 core, all running at a base freq of 3.7Ghz. It's amazing to have that kind of high clock with so many cores.
You don't have to wait. The current Ryzen 3000 series processors support ECC.
ASUS, ASRock, and Gigabyte have several Ryzen 3000 motherboard options that support ECC RAM. Example: https://www.asus.com/Motherboards/Pro-WS-X570-ACE/
The real advantage of the PRO processors is that they support 8-channel memory and have more PCIe lanes. The higher TDP spec might allow for higher all-core turbo, too.
Same TDP spec as the existing non-pro threadrippers - 280W.
> The real advantage of the PRO processors is that they support 8-channel memory and have more PCIe lanes.
Those are advantages yes but the other DIMM types means you can also actually exceed 256GB of RAM, which with 64-cores is definitely also a bottleneck on the existing Threadripper lineup
The current 280W TDP Threadrippers are 24, 32, or 64 core.
These new Threadripper PRO parts have 280W TDP even down to 16 and 12 core parts.
However, even if 16GB and 32GB are easy to find, that won't let you get anywhere near the 1TB limit of the first systems to ship or the technical limit of 2TB as other posters pointed out.
It's just not realistic to equate the two as equal -- these new systems are definitely going to unlock use cases that aren't possible with the regular Threadripper.
The limit of TRX40 is 256 GB max, 32 GB individual, and WRX80 supports RDIMMs, so finding larger UDIMMs is pointless, anyway.
Is there a difference in silicon quality, or is it essentially PBO[1] being preapplied? In other words, can you get the same effect by getting the non-pro part and going into your mobo setting and setting the power limit higher yourself?
It was impossible to find an OEM workstation with those specs so I ended up building it and it's been working with no hassles.
The post you were responding to asked when they would get official ECC support. ECC is not officially supported on Ryzen. If it works on your particular board/BIOS, great, but it's not officially validated or tested, so nobody besides you will actually be testing whether it works, and if it doesn't work in some edge case (or at all) then sucks to be you. If ECC doesn't work right on some future BIOS revision that fixes a security issue or some other critical bug, and there is never an updated version that does support ECC, sucks to be you.
That is specifically the problem that people are looking to avoid by asking for official support.
As it stands, you can probably count the number of boards that anyone would bother to validate ECC on on a single hand - Asrock Rack has two server boards, Asus has some "WS" series that might, maybe some of the high-end Asrock consumer boards. And it will never be officially supported by AMD, if there is a critical bug then welp.
As such I think "ECC is supported on AM4" is a bit of an exaggeration or a clever bit of terminology spin from some fans. If you call up AMD and ask, they will tell you they don't go out of their way to disable it (except in APUs) but it's not supported. I think "functional" or "enabled" is a better term.
(Intel validates support on their consumer processors, by the way... i3s and Pentiums officially support ECC. Supermicro/Asrock Rack/etc test those configurations and if they don't work then you can complain and they will fix them.)
> i3s and Pentiums officially support ECC
my experience is that OEMs like Dell won't support ECC except for Xeon configurations. You can get the same machine with an i9 but not ECC memory or the Xeon and ECC.
The desktop and workstation class Xeons are in direct competition with i9 processors in terms of raw performance, which is why Intel uses features like ECC to enforce product segmentation. This isn't the case for the low-end i3 and Pentium branded processors, because the Xeon lineup simply doesn't extend down that far. If you want to build eg. an industrial PC with ECC memory but you only need a low-end CPU, Intel's official solution is for you to get an i3 or Pentium processor because they don't have anything equivalent under the Xeon branding.
But OEMs like SuperMicro and Asrock Rack do officially support this, and sell barebones or pre-configured systems like this.
https://www.supermicro.com/en/products/motherboard/X11SCH-LN...
https://www.supermicro.com/en/products/system/1U/5019/SYS-50...
Workstation: https://www.asrockrack.com/general/productdetail.asp?Model=C...
(and to be fair, Asrock Rack does have a few AM4 server boards and barebones, and these are among the few I would count on to actually validate ECC. It really just is something you have to explicitly go out of you way to find support for in the AMD ecosystem, whereas it's "free" on any C2xx server board in the Intel ecosystem, the default assumption is that it works because it's considered a supported configuration and that's not the case with AMD.)
https://www.asrockrack.com/general/productdetail.asp?Model=X...
https://www.asrockrack.com/general/productdetail.asp?Model=1...
They do occasionally remove some features like IPMI from those cheaper processor options, but the ECC support stays.
The Asus Pro X570 ACE (or similar name) explicitly offers ECC support though, not just unofficially.
In the case of MSI, those boards reportedly also don't support the ECC reporting function. They will try to correct them but they don't actually report them to the OS. MSI's answer: "wow, sucks to be you". AMD’s answer: “wow, sucks to be you”. That's what you get from "unofficial support", just a halfhearted level of effort all around.
https://hardwarecanucks.com/forum/threads/ecc-memory-amds-ry...
(there are really multiple levels of functionality here, "boots with ECC in non-ECC mode", "corrects ECC silently", and "behaves like a server and reports ECC to the OS or lets you reboot the system". You'll note that nobody ever commits to a particular level of support on AM4 based products...)
Furthermore, AMD doesn't support it at a processor level, so they apply no pressure to the mobo companies to actually test the features they claim they support, nor will they fix it if there is ever a critical functionality bug. An AGESA update could break or remove ECC and welp, sucks to be you, they never advertised that as an actual feature.
That is why I think it is a little ridiculous that people phrase that as "supported". Nobody is validating that it actually works, nobody at AMD will stand behind the feature, and will in fact tell you that they don't support it.
It works, probably, on certain combinations of hardware and BIOS. It is not supported by anyone other than yourself and your own time.
You do realize that false advertising is illegal? Of course realistically that would be a very uphill battle to pursue but that doesn't make it legal.
For example my motherboard has built in audio. The board manufacturer makes some fairly vague claims about what precisely that means but it is clear that there are ports on the board and that they will provide some base level of functionality in conjunction with a supported OS. To the best of my knowledge the CPU manufacturer makes absolutely no claims about that feature. Nonetheless, it is reasonable to expect it to work.
Of course it would be nice if AMD required ECC support as part of the platform. Then all the boards would be required to support it and I wouldn't have to bother reading their spec sheets.
> there are really multiple levels of functionality here
This is the real issue. At least historically, some boards "supported" ECC memory to the extent that they could operate with it inserted; they didn't actually do any error correcting though. Of course in the case I'm aware of (MSI) they specifically stated that. If you as a consumer glossed over their claims and missed that, legally that was on you.
Other vendors specifically stated that their boards officially supported actual ECC functionality. It's not safe to assume precisely what that means though (silent vs reported) unless the manufacturer makes an official claim.
ECC is officially supported on Ryzen CPUs. The feature is present in the chip, and you can RMA it if it doesn't work.
ECC support is not _validated_ on the AM4 platform. That means that a motherboard vendor can market a motherboard that lacks ECC support as being AM4-compatible, and not get a nasty letter from AMD's lawyers. By extension, it also means that ECC support becomes a feature of the motherboard, rather than something you can take for granted that the entire AM4 platform supports.
From the user perspective, it does mean that if you're looking to buy a Ryzen-based platform with ECC support, you need to check to see if the motherboard supports ECC. But once you've made that selection, you will have a fully-functional ECC-capable platform with full support.
> As it stands, you can probably count the number of boards that anyone would bother to validate ECC on on a single hand - Asrock Rack has two server boards, Asus has some "WS" series that might, maybe some of the high-end Asrock consumer boards. And it will never be officially supported by AMD, if there is a critical bug then welp.
All ASRock AM4 boards (that I'm aware of) support ECC, and most ASUS boards do. I can't speak for other brands, but ECC is pretty broadly supported (the motherboard manual will say whether it's supported it or not).
What separates the motherboards you mentioned is that ECC support is _marketed_ as being a feature.
As a data point, I'm currently running an ASRock B450M Pro4 motherboard with a Ryzen 9 3900X and 64 GB of Kingston KSM26ED8/16ME DDR4-2666 ECC DRAM, and it is working with full ECC support.
The usual distinction is that Threadripper regular takes unbuffered ECC, whereas Threadripper Pro and Epyc take buffered ECC. But they're both ECC.
The TDP is (in part) because the Epyc IO die consumes roughly 100W by itself. You can still cool a Threadripper on air though.
I'm pretty disappointed in this, but I suppose the market for such workstations leans more towards "large enterprise that needs reliability" than "hobbyist that needs low cost/likes to customize their rig."
What I'm surprised about is that they're branding this "Threadripper Pro" instead of branding it as a high end Epyc workstation chip. Threadripper is a watered down Epyc chip to start with, and all they've done here is raise base clocks, lower boost, and turn on the additional IO that Epyc already had which was disabled in TR.
Are you saying that these chips might last only like 2-3 years and one day just burn off on boot, even if cooled the best way possible?
Unclear what the price differential will be, but consumers can buy EPYC chips just fine, so if you want something home-built, that's a path. The main issue is finding a good motherboard; going Xeon-at-home is pretty straightforward, but EPYC boards tend to be really basic server-grade stuff designed for a datacenter.
I'm quite OK with the limit of 1TB of RAM but the forced buy of a Quadro card is an absolute nope from me.
I hope that when other vendors are allowed to build TR Pro machines that some of them will actually give you a bit more choice.
I think the main thing here with the list of choices in the need to provide ISV certification.
But won't opening the box to add your own GPU void warranty? This workstation kind of gives off that vibe of a professional machine that's pre-built for you and you're not supposed to open it.
Hope I'm wrong.
The design of the system looks like it should be easy to open. A system like this is designed to be serviced and upgraded. Often, customers pay for onsite support to take care of those tasks, but I very much doubt that adding in a PCIE device would void the warranty.
The real question again is about ISV certification. If you are using a consumer GPU for professional applications, then you won't necessarily have the same kind of support from the software vendor if the application doesn't work correctly.
Thanks a lot for the helpful explanations.
Just curious, can you use more than the 256GB limit on the regular Threadrippers? If so, what kinds of applications do you use that would consume that kind of memory?
Basically I am looking to shoot many birds with one stone when I decide to buy my next Linux workstation, even if it ends up costing $30,000+. And for the kinds of research work I am interested in, a ton of RAM seems to be a boon.
But it's more like some paranoid future proofing sentiment than anything else right now.
If you wanted a TR Pro workstation with a consumer graphics card instead of a Quadro, you could work with a third-party reseller. You and the reseller would have to verify compatibility yourselves (e.g. will the card fit in the chassis, will it have the necessary airflow, will it have the needed power connectors and will they reach, etc.), but it's doable.
I guess it would be possible to add a graphics card of your choice to this workstation as well.
I imagine it would be nice to have a mini-AWS at home. Not sure how it would feel like to have a 1K Watt space heater permanently running though.
[1] https://hothardware.com/reviews/amd-ryzen-threadripper-3990x...
This is probably a small misprint; EPYC only supports 2TB per socket, or 4TB on multi-socket boards.
In that regard, TRPro looks like its just a re-branded EPYC chip.
Take GPU like this one RX 570 https://www.newegg.com/sapphire-radeon-rx-570-100412p4gitxl/...
it's still good enough for many games, for me it's fine even with The Witcher 3, sure, not the top tier, but I also looked for energy efficient PC. As low energy Ryzens are hard to get, I took 5 3600 and compromised on GPU, storage and mobo.
I haven't followed the Ryzen versions like the 3600E you mentioned, so I'm not sure how much binning is done there to get power consumption even lower, but I'd guess that you can get most of the effect with a regular Ryzen and manually setting the TDP to 45W.
any particular reason why? power costs? cooling?
Threadripper 3960x will have to do then :(
Apart from the OS, of course.
If you're in category 1, you have a job to do today. Apple's roadmap isn't especially relevant as your job isn't going to wait 6 months to 2 years. If they are planning for the future, the decision tends towards the "safe bet". Mac Pro on Intel is a known quantity that works with their software and workflows today. Future Apple Silicon Mac Pro is a big unknown. They may choose to buy more Intel Pros today while they can, they may try the new ones when available. Or they might investigate switching platforms. I don't think sales here will be impacted greatly.
Category 2 users may decide to wait to purchase a new Apple Silicon whatever when it comes out because they always want the latest and greatest. Whatever they decide to buy is cash in Apple's pocket.
Either way, Mac Pro sales are a drop in the ocean in terms of Apple's revenue.
Indeed. It's their "supercar". It's there to show how impressive an Apple can be and, in that, it doesn't fail - it's a great machine and a solid example of how a well designed computer should be easy to maintain.
I seriously doubt the Lenovo can be opened in less than 10 seconds.
Here's a video of them opening up.
https://www.youtube.com/watch?v=xQGfVFpMpuQ
Here's a video of the case I used in my last build, released 2017.
The side just latches off with minimal effort, this feels even more accessible than having to pull the entire cover off. Both the define r6 and the mac pro have extra steps to get to additional components, like the latch to remove the memory cover on mac pro or the latch to access the top fan mounts on the define r6.
...while remaining functional afterwards
Alienate them like every major transition Apple has ever made has (Classic MacOS->MacOSX, 68K->PPC, PPC->x86, NuBus->PCI, mono->color, all-in-one->expandable, etc., etc., etc.)?
The thing about predicting the end of the world is that if you say it long enough and consistently enough, one day you'll be right. Be so kind as to let us know when Apple goes out of buisiness, will you?
Especially if the Apple Silicon Mac Pro's paste the Intel ones. Companies will change them out in a heart beat if they deliver solid productivity gains. Time is money is more than a pithy saying.
Another option is to fill all the blanks in GNUStep and make a macOS compatible layer sitting on top of another OS. That'd be really helpful.
Are you sure? I'm pretty sure it says that can come preconfigured with Linux :)
If you're custom building PCs, I really doubt you're under-served by their existing lineup. Any time savings you'd get working on a TR Pro machine over regular TR or Ryzen are probably negated by the time you spent fiddling around with assembling the computer and configuring it.
IMO, there are too many computer enthusiasts out there fixated on having powerful machines as a sort of status symbol. It ends up being a waste of good computer parts.
It has higher base frequency than Epyc, typically a bit cheaper than Epyc and it now has up to 2TB of ECC RAM support.
Not complaining. The perf/$ is amazing.
I hope one of them will.
There might be certain use cases where Threadripper Pro might displace Epyc, but for the most part, they aren't competing in the same markets.
+ Reliability and enterprise support
- Ryzen 3, 5, 7 which was fine
- Ryzen Threadripper because for some reason we couldn't just call it Ryzen with a bigger number (maybe they anticipated needing the 9 later)
- Epyc for servers that supports ECC memory
- Now we have Ryzen 9 because Intel made Core i9 (I guess this makes sense for people comparison shopping)
- Now we also have Threadripper Pro just because we can, not even as a response to Intel
Epyc is the only distinction that needs to be made IMHO. The rest should just be Ryzen 3,5,7,9,11,13... for desktop parts. It's fine if Ryzen 11 and 13 had different sockets (like TR does). I don't need TR and I don't need TR Pro to denote product lines alongside the numbering scheme.
I know this is a minor issue, but it shouldn't be that hard. Yeah, yeah, "there are only two hard problems in computer science..." But look at the crap that consumers have to deal with around USB naming and WiFi naming. Don't make CPU naming follow the same trend.
Intel uses a different socket for the i9-10980XE than they do for the i9-10900. They have done this in the past, it doesn't seem like it is an issue for them to put Extreme Edition (their halo line) into the same naming convention. They don't even make a new tier for it, they just stick it at whatever tier is highest.
You probably don't want the i9-10900X, it's worse in every way except it works on the same socket as other Skylake-X parts. It has the same core count, on mostly an older version of the same architecture, at lower clocks, for higher power consumption, in a more expensive platform than the i9-10900X. In basically every use case, the extra clocks on the i7-10700K even make it superior for multi-threaded tasks despite having 25% less cores.
So there's no simplification here. You can't look at i9-10900X and i7-10700k and go "Bigger numbers better, so simple". You're comparing the worst HEDT CPU to the upper end of consumer CPUs. The only reason to buy it is if you consider you're buying HEDT for the platform (you need the extra PCIe lanes). That information is not in Intel's model names, but AMD put it in theirs, which seems better at explaining that difference (also AMD don't have a third gen 16C/32T threadripper SKU to complete with their own 3950X).
I legitimately do not understand writing your post directly addressing AMD as if they're actually listening to you.
> Now we have Ryzen 9 because Intel made Core i9
Or maybe because it was the natural +2 progression
I disagree. AMD did the world a favour by going with the Threadripper brand name instead of the lazy path of mindlessly bumping a number and tagging another brand. The different socket alone is a clear and unambiguous reason to go with a distinct brand, but the clear bet on maxing out memory channels and PCI lanes at the expense of far higher power consumption makes them an entirely different product targetted at an entirely different market segment.
What you need to know is really what socket/chipset the thing works with, what architecture it is (sky/whatever lake, Zen 2, etc), how many cores, what features (smt, avx, ecc, gpu, etc), frequency, and power limits.
So mamy people say "i have an i7" as if that means anything over time. An early i7 is a 4-core part, where 4-core parts are entry level now, and of course a 10 year old part is behind a lot on IPC compared to a recent processor.
i* = ryzen *
threadripper = skylake-X and the like (HEDT processors)
threadripper-pro = xeon-w
epyc = xeon
I omitted various embedded/low-end/iGPU/mobile lines. The only thing where AMD has nothing comparable is the low power atoms. On the other hand intel has nothing to offer in the console market
But wait, that's just one of the confusions of Intel chips. Do they not also have workstation chips... and server chips... which are both called Xeon?
But... what about Core i5/i7/i9 vPro you say? Sorry yeah I forgot those were a thing. (And Intel still sells Pentium and Celeron... and apparently maybe still Atom and Itanium... https://www.intel.com/content/www/us/en/products/processors.... )
I don't know if Intel or AMD are doing this stuff with intention to confuse, inform or just market. AMD has mixed generations between CPU, APU and mobile APU (Zen 2 as 3xxx desktop CPU but 4xxx APUs for example)... but I would definitely say the product lines are pretty clear.
EPYC - Servers
Threadripper/Pro - Workstations
Ryzen - Consumer