AMD might be about to give dual-socket functionality to their Threadripper CPUs

mongeese

Posts: 643   +123
Staff
In a nutshell: At the end of last week, an AMD report describing the features and specifications of the upcoming 5000-series Threadripper Pro CPUs was leaked. It said that all five models would have dual-socket functionality. Shortly after, two Threadrippers appeared in the PassMark database having completed the benchmark in dual-socket mode - but they’re not from the 5000-series.

Instead, they were two of the seventeen-month-old flagship of the 3000-series: the 3995WX. In theory, because the CPU shares its hardware with the dual-socket-capable Epyc 77x2-series, the 3995WX has only ever been prevented from engaging in a dual-socket mode by software limitations.

Working in tandem, the two CPUs achieved a score of 123,631 points; 35% more than the median result of a single 3995WX, and the highest score of any two-CPU pairing in the database.

If the result is genuine, which it seems to be, then there’s only one likely culprit: AMD themselves. It’s simply too unlikely that another party could modify the two CPUs, which cost some $8,800 each, and the requisite motherboard, successfully.

As for why AMD would be experimenting with their old CPUs, our best guess is that the creation of the first dual-socket sWRX8 motherboards for the 5000-series has raised some questions about their backward compatibility. A microcode update could conceivably enable dual-socket functionality in 3000-series CPUs, though AMD doesn’t have much of an incentive to create one.

Credit: Onur Binay

At a minimum, though, AMD does have an incentive to enable dual-socket functionality on the 5000-series. In the past, doing so would’ve cannibalized the Epyc series; it’s one of the main features that differentiate the two product lines. But, as of 2022, the Epyc series will be an entire "generation" ahead of Threadripper and use a newer architecture at a minimum, if not a newer node as well.

Most of the available information about the 5000-series comes from the aforementioned report, which was acquired by Igor’s Lab. Its contents haven’t been verified beyond a few match-ups with other leaks, but Igor’s Lab is a trustworthy source. That said, sometimes specifications are changed in the lead-up to the processors’ announcement.

Possible Threadripper Pro 5000-series Specifications

  5995WX 5975WX 5965WX 5955WX 5945WX
Cores / Threads 64 / 128 32 / 64 24 / 48 16 / 32 12 / 24
Single-Core Boost Clock 4.55 GHz
All-Core Boost Clock 2.70 GHz 3.60 GHz 3.80 GHz 4.00 GHz 4.10 GHz
Base Clock 2.25 GHz 2.70 GHz 2.80 GHz 2.90 GHz 2.94 GHz
L3 Cache 256 MB 128 MB 64 MB
L2 Cache 32 KB 16 KB 12 KB 8 KB 6 KB
TDP at Boost Clock 280 W
TDP at Base Clock 229 W 190 W 171 W 152 W 138 W

There’s now only Pro (with a "W") versions of the processors, according to Igor’s Lab. This year, there are five, up from four; the addition was of the 24-core model.

On the whole, the specifications of these processors aren’t too different from their predecessors. Their all-core clock speeds are a couple of hundred megahertz higher or lower in some cases, but broadly similar. Only their single-core clock speed is a consistent upgrade of 250-350 MHz.

Like the Ryzen 5000-series, the biggest upgrade is under the hood: the Zen 3 architecture. In our testing, it could provide an IPC performance uplift of 10-20% in various applications. It might provide an even larger uplift on higher-core count models that benefit from its impressive inter-core and cache latency, vastly improved over the Zen 2 architecture of the 3000-series.

But it’s more likely to be the dual-socket functionality driving the sales, should it eventuate. It would be interesting to see what 128 unlocked cores can do.

Permalink to story.

 
I don't understand how these would cannibalize epyc server cpu sales. These are HEDP parts. I guess you could stick them in a server rack but costs are basically the same. The form factor doesn't make sense and the cost is basically the same anyway. Also, Epyc CPUs and server products come with commercial level support. That alone is worth millions a year to businesses.

Don't get me wrong, I'd throw a threadripper pro in my server rack at home but everything I do there is nothing more than a hobby
 
I don't understand how these would cannibalize epyc server cpu sales. These are HEDP parts. I guess you could stick them in a server rack but costs are basically the same. The form factor doesn't make sense and the cost is basically the same anyway. Also, Epyc CPUs and server products come with commercial level support. That alone is worth millions a year to businesses.

Don't get me wrong, I'd throw a threadripper pro in my server rack at home but everything I do there is nothing more than a hobby
A few of these would be enough for small-medium companies to build a good inhouse server for rendering and storage. Instead of 10-20k you would pay just 20-30% of the price. These can add up a lot globally. The server market operates at much much higher profit margins and AMD would have to use more of its dies in HEDT instead of servers.
 
A few of these would be enough for small-medium companies to build a good inhouse server for rendering and storage. Instead of 10-20k you would pay just 20-30% of the price. These can add up a lot globally. The server market operates at much much higher profit margins and AMD would have to use more of its dies in HEDT instead of servers.
In the commercial sector down time costs money and companies are willing to pay a premium for commercial level support. If these things go down from some compatibility error that pops up down the road because of a software update they're going to be paying A LOT more than 10-20k for downtime. And I don't know where you're getting your prices from, comparable Epyc CPUs and threadrippers are very closely priced.

You see the same thing with nVidia Quadro vs their RTX series. Companies aren't buying 3090's just because they're cheaper, they're paying for the support that comes with the Quadro series.
 
In the commercial sector down time costs money and companies are willing to pay a premium for commercial level support. If these things go down from some compatibility error that pops up down the road because of a software update they're going to be paying A LOT more than 10-20k for downtime. And I don't know where you're getting your prices from, comparable Epyc CPUs and threadrippers are very closely priced.

You see the same thing with nVidia Quadro vs their RTX series. Companies aren't buying 3090's just because they're cheaper, they're paying for the support that comes with the Quadro series.
Human greed and lack of cashflow can lead to such situations. Not everybody needs the perfect hardware, they can handle a few errors from time to time. It's also simply faster to buy powerful HEDT systems on a normal online market than to order servers (it also requires less manpower to install and operate - simply put it's just cheap).
 
Human greed and lack of cashflow can lead to such situations. Not everybody needs the perfect hardware, they can handle a few errors from time to time. It's also simply faster to buy powerful HEDT systems on a normal online market than to order servers (it also requires less manpower to install and operate - simply put it's just cheap).
if they have a lack of cash flow then they'd never be a significant part of AMDs market. And I don't seem to think you understand this, the premium isn't for the hardware, it's for the software support. It doesn't matter if you save money on hardware if that hardware doesn't work with the software you're using.

I will say this again, go look at Threadripper prices vs Epyc prices, they're almost identical core-for-core. You want to bring up having someone install and set this stuff up. I don't know if you thought of this, both Threadrippers and Epyc CPUs need installed and connected to the network. Frankly, installing an Epyc server might be cheaper because they come with the documentation, support and installation instructions.
 
if they have a lack of cash flow then they'd never be a significant part of AMDs market. And I don't seem to think you understand this, the premium isn't for the hardware, it's for the software support. It doesn't matter if you save money on hardware if that hardware doesn't work with the software you're using.

I will say this again, go look at Threadripper prices vs Epyc prices, they're almost identical core-for-core. You want to bring up having someone install and set this stuff up. I don't know if you thought of this, both Threadrippers and Epyc CPUs need installed and connected to the network. Frankly, installing an Epyc server might be cheaper because they come with the documentation, support and installation instructions.
For every big server farm AMD/Intel sell servers to hundreds of smaller companies. I don't get why you are confused by this.

And I'm obviously talking about pre-built systems/server racks, not CPU newegg prices. Any company that forces a poor IT intern to build something like this from scratch would have never went for a server anyway.
 
Last edited:
For every big server farm AMD/Intel sell servers to hundreds of smaller companies. I don't get why you are confused by this.
I'm not confused by it at all. What do you think I'm confused about? I have 3 1800X's and a 1700x running Linux on my server rack all with different motherboards. It took me a few days to troubleshoot everything on each one. It's one thing when it's computers in my closet I bought off craigslist and ebay, it's another when you have a week of downtime and your employees can't login to do finance reports, order inventory or submit payroll. There is also a decent bit of legal protection here. Companies can get fined heavily if paychecks go out late. In my industry the penalty for late paychecks is 1 days pay, per day for up to 10 days. That's per pay cycle.

Now, if you have the compatible hardware and software the business can point to Dell or HP and say, "hey, we did the right thing, we paid for enterprise support, this on them" and then they can pass the liability onto the manufacturers.

You say I'm confused, but I'm trying to explain the complex nuances of how ThreadRipper Pros wont take away from the enterprise sector(Small business still counts as enterprise) and even giving you examples.

Frankly, I think ThreadRippers are going to be soft locked out of server boards and it's going to be more like AMD's 4x4 systems back in 2005-2006 IIRC.
 
Can I get your source first?
Last I heard 512GB is the current max. Source: AMD TR spec sheet.
He's right, the Pro models (with WX at the end) have eight memory channels, per AMD's site: link. In theory, they could have up to 4 TB in dual-channel mode, but I've only seen single-channel motherboards that support up to 2 TB like this one: link.

You say I'm confused, but I'm trying to explain the complex nuances of how ThreadRipper Pros wont take away from the enterprise sector(Small business still counts as enterprise) and even giving you examples.
Threadripper has additional advantages over Epyc, like much higher clock speeds. And one of the major benefits of Epyc, which is software and security updates, isn't important to everyone.

I think that a small portion of businesses would consider both options. Maybe a university that's doing some small scale scientific modelling would consider Threadripper.

And, as a point of discussion, I'd like to highlight the fact that Nvidia's terms of service for their GeForce cards explicitly prohibit their use in servers, and they won't sell them in bulk to companies that want to use them in data centers.
 
There is this on Wiki. It states the max is upto 4TB. But then I just noticed that is also listed under Epyc.
System DRAM
  • EPYC 7003 "Milan":
    • 8 channels per socket, up to 16 DIMMs, max. 4 TiB
    • Up to PC4-25600L (DDR4-3200), ECC supported
    • SR/DR RDIMM, 4R/8R LRDIMM, 3DS DIMM, NVDIMM-N
  • Ryzen 5000 "Vermeer":
    • 2 channels, up to 4 DIMMs, max. 128 GiB
    • Up to PC4-25600U (DDR4-3200 UDIMM), ECC supported
  • Ryzen 5000 APU "Cezanne":
    • DDR4-3200 or LPDDR4-4266
 
Threadripper has additional advantages over Epyc, like much higher clock speeds. And one of the major benefits of Epyc, which is software and security updates, isn't important to everyone.

I think that a small portion of businesses would consider both options. Maybe a university that's doing some small scale scientific modelling would consider Threadripper.

And, as a point of discussion, I'd like to highlight the fact that Nvidia's terms of service for their GeForce cards explicitly prohibit their use in servers, and they won't sell them in bulk to companies that want to use them in data centers.
I'm getting tired of replying to this thread. ThreadRipper and Epyc CPUs are different products. Your example of small scale scientific modeling is a great example, but ThreadRipper isn't designed for mission critical applications like businesses who need guaranteed software support and 100% uptime.

And nVidia can say whatever they want in their ToS, nothing is stopping anyone from buying 1000 3080's and throwing them in a server farm. What are they going to do, bust down their doors and shut them down? You'd be a fool to do something like that but someone can do it if they really want to. If you need 1000 GPUs then paying a premium for the enterprise support that comes with it is well worth the money. Frankly, aren't GPU mining rigs technically a server? The only difference is the form factor and nVidia has no problem selling consumer GPUs to miners by the thousands.
 
He's right, the Pro models (with WX at the end) have eight memory channels, per AMD's site: link. In theory, they could have up to 4 TB in dual-channel mode, but I've only seen single-channel motherboards that support up to 2 TB like this one: link.


Threadripper has additional advantages over Epyc, like much higher clock speeds. And one of the major benefits of Epyc, which is software and security updates, isn't important to everyone.

I think that a small portion of businesses would consider both options. Maybe a university that's doing some small scale scientific modelling would consider Threadripper.

And, as a point of discussion, I'd like to highlight the fact that Nvidia's terms of service for their GeForce cards explicitly prohibit their use in servers, and they won't sell them in bulk to companies that want to use them in data centers.
Sorry, I read this article thinking they were desktop TR's which are 256 and 512GB.

I don't travel far from desktop parts, so my confusion sparked my curiosity about the WX lineup. I'm gonna finish reading Tom's 3995WX review to get myself more up to speed. Thanks for replying.

I found this in Tom's Review about your thoughts on a 4TB limit:
"The Threadripper Pro chips come with the 'WX' suffix to denote they are designed specifically for the professional workstation market and drop into specialized single-socket WRX80 motherboards featuring the sWRX8 socket."
 
Last edited:
3rd gen Threadripper has 1TB capacity, and the TR PRO processors, that have been out for a year already up to 2TB RAM. Where ya been?
I'm more of a desktop guy and I got the Pro's (WRX80) confused with the non Pro's (sTRX40) which do have 256GB max capacity. The Pro's weren't really my jam until this article. I'm getting up to speed though. Tom's and Anandtech are big helps.
 
Sounds like a great computer for livestreaming/ gaming?
This would be total overkill for livestreaming/gaming. CPU wise you only need a 5600x for gaming, 5800x for best results, a 12 core 5900x is hardly any faster, so one of those would give you enough CPU power to game and livestream without issues.

Nvidia GPUs used by the vast majority of streamers already support on board encoding and recording with OBS for streaming, no other hardwar required. If for some reason you dont want to do so, you can buy 1080p60 capture devices from elgato relatively cheap, many of which run on USB, although an internal card is available. The B550 chipset would be more then sufficient for this setup. If you wanted 4k/1080p240/HDR10/1440p144, you need PCIe 4 and a 4k60 pro card, meaning you need X570 for the extra x4 4.0 lanes.

None of this requires threadripper. Outside of database or rendering work nobody needs threadripper at home.
 
I'm not confused by it at all. What do you think I'm confused about? I have 3 1800X's and a 1700x running Linux on my server rack all with different motherboards. It took me a few days to troubleshoot everything on each one. It's one thing when it's computers in my closet I bought off craigslist and ebay, it's another when you have a week of downtime and your employees can't login to do finance reports, order inventory or submit payroll. There is also a decent bit of legal protection here. Companies can get fined heavily if paychecks go out late. In my industry the penalty for late paychecks is 1 days pay, per day for up to 10 days. That's per pay cycle.

Now, if you have the compatible hardware and software the business can point to Dell or HP and say, "hey, we did the right thing, we paid for enterprise support, this on them" and then they can pass the liability onto the manufacturers.

You say I'm confused, but I'm trying to explain the complex nuances of how ThreadRipper Pros wont take away from the enterprise sector(Small business still counts as enterprise) and even giving you examples.

Frankly, I think ThreadRippers are going to be soft locked out of server boards and it's going to be more like AMD's 4x4 systems back in 2005-2006 IIRC.
Threadrippers have no need to be softlocked, they just don't work hardware wise in server boards.
 
I'm getting tired of replying to this thread. ThreadRipper and Epyc CPUs are different products. Your example of small scale scientific modeling is a great example, but ThreadRipper isn't designed for mission critical applications like businesses who need guaranteed software support and 100% uptime.

And nVidia can say whatever they want in their ToS, nothing is stopping anyone from buying 1000 3080's and throwing them in a server farm. What are they going to do, bust down their doors and shut them down? You'd be a fool to do something like that but someone can do it if they really want to. If you need 1000 GPUs then paying a premium for the enterprise support that comes with it is well worth the money. Frankly, aren't GPU mining rigs technically a server? The only difference is the form factor and nVidia has no problem selling consumer GPUs to miners by the thousands.
It's not about getting support or not. It's about having the option of not getting it. That's what you don't understand right now. If you can get a server that is much cheaper that offers many memory channels and many PCIe lanes then many will do so. The latest Threadripper Pro has moved a lot towards replacing servers (including the hardware cost).
 
It's not about getting support or not. It's about having the option of not getting it. That's what you don't understand right now. If you can get a server that is much cheaper that offers many memory channels and many PCIe lanes then many will do so. The latest Threadripper Pro has moved a lot towards replacing servers (including the hardware cost).
People have that option with the curremt threadripper pro and they just aren't. This probably has something to do with Epyc cpus costing basically the same, core-for-core. Can you do some atleast some basic research before before you reply saying I'm confused

Threadrippers have no need to be softlocked, they just don't work hardware wise in server boards.
Did you even read the article? It's in the first few sentences. They literally have the same socket down to memory channel pinouts. It's a softlock, this whole test was likely run on a 2p server board with a modified bios.
 
Last edited:
People have that option with the curremt threadripper pro and they just aren't. This probably has something to do with Epyc cpus costing basically the same, core-for-core. Can you do some atleast some basic research before before you reply saying I'm confused


Did you even read the article? It's in the first few sentences. They literally have the same socket down to memory channel pinouts. It's a softlock, this whole test was likely run on a 2p server board with a modified bios.
"and they just aren't" - These kind of assumptions make no sense. Do you know something that the rest of the industry doesn't know? I don't know where you live but I've seen TR servers in smaller companies and reason is most the time advice from the IT admin (and I would give the same advice a TR system would fit well instead of a server). Word of mouth is very powerful.

Getting dual socket HEDT will effectively make TR a good choice for many types of internal workloads as you don't need to buy two entire systems instead of one big server. 128 cores of pure raw power, tons of I/O and ECC memory that has a lot of capacity/bandwidth makes for a very compelling purchase.

An barebones server with Epyc 7763 (64 cores), 64TB RAM and 1TB SATA HDD costs around 12k euro (configured on thinkmate, no GPU). You can get a better TR system with a professional GPU for cheaper.
 
Last edited:
"and they just aren't" - These kind of assumptions make no sense. Do you know something that the rest of the industry doesn't know? I don't know where you live but I've seen TR servers in smaller companies and reason is most the time advice from the IT admin (and I would give the same advice a TR system would fit well instead of a server). Word of mouth is very powerful.

Getting dual socket HEDT will effectively make TR a good choice for many types of internal workloads as you don't need to buy two entire systems instead of one big server. 128 cores of pure raw power, tons of I/O and ECC memory that has a lot of capacity/bandwidth makes for a very compelling purchase.

An barebones server with Epyc 7763 (64 cores), 64TB RAM and 1TB SATA HDD costs around 12k euro (configured on thinkmate, no GPU). You can get a better TR system with a professional GPU for cheaper.
Well toms hardware did a review of a similarly spec'd ThreadRipper pro system that has an MSRP of $19,000. It had a pair of nVidia quadro RTX 6000 cards in it which retail for $3400 each so take $7000 off of that and we're in in the same $12000 range as your Epyc server.


The cost of the systems is basically the same, the lack of enterprise support isn't worth it for a business. If you know "professionals" who recommend a TR system for use as a server rather than an actual server then I'd argue they wasted 3 semesters at community college.
 
Back