AMD Threadripper 3970X Performance On Linux After Two Years

[jaxa]

I don't see AMD being worried enough about Alder Lake to explicitly cut prices. They can just phase it out and let the retailers handle getting rid of the old stock. Prices have already fallen under MSRP for everything.

Zen 3 with 3D V-Cache could erase Alder Lake's advantages, while still targeting the same AM4 platform. We already know that AMD claimed a +15% average performance increase in games, with some games getting bigger increases and some getting none. Then we have Microsoft's benchmarking of Milan-X Epyc, which showed up to +80% performance in some applications that were sensitive to memory bandwidth (performance increases would typically be less, such as 50%). I don't know if those gains would translate over to Ryzen, but if they do, a 5950X with triple the L3 cache could blow the 12900K out of the water.

[ddriver]

AMD has no reason to prioritize TR as long as Intel isn't competing in that segment. It doesn't matter that it's not performing as well as it could if your alternatives are Zen2 TR and the aging Xeon Ws from Intel.

That said, Moore's Law is Dead made a video a while back saying that Intel is going to launch new things in the workstation segment next year, so we'll see how long AMD has without competition.

Ummm... intel and amd hedt do not exist in a vacuum. Competition is only one reason why amd may need to launch an improved tr.

TR is facing performance competition from ryzen now, it is one thing to skip a generation because your competitor's got nothing, it is another thing if your own mainstream makes your hedt anemic and pointless.

A 16 core with vcache zen 3 ryzen will probably land around the 24 core zen 2 tr...

[smitty3268]

Ummm... intel and amd hedt do not exist in a vacuum. Competition is only one reason why amd may need to launch an improved tr.

TR is facing performance competition from ryzen now, it is one thing to skip a generation because your competitor's got nothing, it is another thing if your own mainstream makes your hedt anemic and pointless.

A 16 core with vcache zen 3 ryzen will probably land around the 24 core zen 2 tr...

I think one rumor is that AMD is afraid that TR is too good, and can possibly cannibalize sales of Epyc and that's why the Zen 3 TR got delayed. Trying to figure out exactly what form it should come out in, and how much to charge for it.

[jaxa]

I think one rumor is that AMD is afraid that TR is too good, and can possibly cannibalize sales of Epyc and that's why the Zen 3 TR got delayed. Trying to figure out exactly what form it should come out in, and how much to charge for it.

Rather than Threadripper cannibalizing Epyc sales, I think it comes down to Epyc being backordered, better margins on Epyc, and Epyc being more strategically important (a good contract could result in several years of sales to a big customer). Add the lack of effective competition in HEDT from Intel, and there is no incentive to launch a new Threadripper.

If things change, AMD just has to assemble the chiplets.

Ryzen encroaching on Threadripper is interesting, but if Ryzen is going to 16-core Zen 4 (2 chiplets only) and then 8-core Zen 5 + 16-core Zen 4c, the Threadripper users will have 96-core Zen 4. Also, some people will still buy low core count Threadripper if it has more memory channels.

[Slartifartblast]

AMD Threadripper 3970X Performance On Linux After Two Years ?

Shirley it should be Linux performance on AMD Threadripper 3970X after two years, and don't call me Shirley.

[coder]

I think one rumor is that AMD is afraid that TR is too good, and can possibly cannibalize sales of Epyc and that's why the Zen 3 TR got delayed.

That's silly. First, TR is not exactly cheap, much less the Pro version. Second, AMD can't make Milan (Zen3) Epyc fast enough to keep up with demand, which I think is the main issue.

IMO, the main issues are:

1. Lack of top-binned chiplets (I know some of you don't agree with this, and since I have no proof, we'll just take it as pure speculation - no need to rehash that debate).
2. Lack of engineering resources, at least for non-Pro threadripper platform work. Over the past 3 years or so, I think customer engagement @ AMD has probably ballooned much faster than they could staff up to support it.

[coder]

A little off-topic, but it seems to me there's a missing opportunity. AMD's approach of moving the memory controller onto an I/O die actually lets them decouple platform changes (i.e. new socket) from chiplet changes. In the ideal world, they could conceivably introduce a new socket, and still sell a generation of CPUs for both the new & old sockets. And this goes for Threadripper (DDR5? PCIe 5?) just as much as Ryzen.

I understand the engineering work needed to support all of that probably costs more than the market opportunity is worth. However, it's neat to contemplate.

[smitty3268]

That's silly. First, TR is not exactly cheap, much less the Pro version. Second, AMD can't make Milan (Zen3) Epyc fast enough to keep up with demand, which I think is the main issue.

IMO, the main issues are:

1. Lack of top-binned chiplets (I know some of you don't agree with this, and since I have no proof, we'll just take it as pure speculation - no need to rehash that debate).
2. Lack of engineering resources, at least for non-Pro threadripper platform work. Over the past 3 years or so, I think customer engagement @ AMD has probably ballooned much faster than they could staff up to support it.

I'm not saying that's the primary reason. Obviously AMD is going to prioritize Epyc because that's where the most money is, and making inroads into the the server market is big for AMD in the long run. HEDT is something they can afford to ignore.

I'm just saying I believe it's a factor in why it's been so slow. They could have come out with an update this fall if they wanted to - which most likely would have been a response to an Intel HEDT chip that was competitive. But they aren't quite sure how it fits into their lineup right now (do they want 8-channel TR chips with ECC support? Do 64 Zen 3 core chips make sense with only 4 channels? People will likely complain more if AMD reduces the core count vs just not putting out a product at all), and Intel is making it easy on them to just leave alone for now while they focus their efforts elsewhere.