Hacker News Reader

Comment by andreime | Hacker News Reader

While listening to the podcast "Moore's law is dead", I heard repeatedly that AMD has no interest in launching new Threadrippers, ever again.

View on HN

That is almost certainly wrong. Threadripper is what AMD is also using for its workstation line of products (eg, https://www.lenovo.com/us/en/thinkstation-p620/ ).

So unless AMD just doesn't want a piece of that pie anymore despite only just entering the market, there's absolutely more Threadrippers coming. No doubt about that at all. AMD isn't going to leave the Xeon-W just entirely unchallenged when AMD is otherwise firing on all cylinders.

What they probably won't do is continue to have different sockets for the "Pro" and "non-Pro" threadrippers, though. But that'd be different from them just not having anymore Threadrippers at all. It'll probably just end up in the same category as Ryzen & Ryzen Pro where it's the same socket, just one is sold to consumers and one is sold to OEMs.

I don't think MLID has said anything besides that there won't be a Zen 3 Threadripper, because Intel still doesn't have anything that comes close to Zen 2 TR.

Now, a Zen 4 Threadripper is extremely compelling because of DDR5 and PCIe 5.0.

To be fair I'd be completely okay with them basically making TR and TR Pro with even more L3 cache and leave it at that. I'd still buy.

That's definitely not happening. The L3 is part of the chiplet, so they aren't going to take Zen 2 and rev it to have more L3. Rather they'd just swap it out with a Zen 3 chiptlet instead, which is far less work for AMD to do. Especially since TR & TR Pro already have way more L3 than anything else (128-256MB L3)

I highly doubt we'll see a vcache version of TR/TR Pro prior to the jump to Zen 4, especially since so far AMD has only officially announced a 3d vcache 5800X which is a single-chiplet product ( https://www.anandtech.com/show/17152/amd-cpus-in-2022-ces )

Yep, I've definitely fallen behind on the info there. Looks like you are right.

I do wonder though, what improvements will they bring to the next Threadrippers? There are some rumours about the start of March 2022, I am very curious what can they unveil. Any theories?

A March 2022 Threadripper product would all but certainly just be a Zen 3 update. So compare the performance gains going from a 3950X to a 5950X or from Epyc 7002 to 7003. Just that but applied to Threadripper (so something like +15% performance)

Thank you. In the HEDT space a 15% improvement is quite good though.

And I am not sure Zen 4 Threadrippers would even be available this year but who knows. Plenty of rumours. I better start saving up money because I want to invest in a workstation that will last me at least 10 years (minus SSD wear). And Zen 4 / PCIe v5 / DDR-5 look preeeeeeeetty good.

I don’t understand the logic of waiting so long to launch an already outdated platform when Zen 4 is nearly ready to go. It’s just so late in the product cycle.

I agree that a March launch would mean it would be on Zen 3.

Zen 4 is on TSMC 6nm and so is the new line of laptop skus running Zen3+, so I'd guess that they are just waiting for sufficient volumes to release another 7nm Zen 3 product. Epyc & Ryzen almost certainly pull much higher volumes for AMD than TR/TR Pro, and although Ryzen at least has been mostly available for the past couple months it was in short supply for quite a while there.

I'd guess they planned on TR / TR Pro being a quicker followup after Epyc 7003 launched, but then chip capacity just wasn't there to justify it. And now it probably is, especially as their new stuff is on a different node.

But from a product perspective if there are still no competitors to the extant 2019 Threadripper SKUs, what's the point of such a minor bump today? I don't buy the capacity argument for the exact reason you mention: it's low volume. That's the whole beauty of the Zen architecture, right? They can shift capacity around after the wafer stage.

The way I see it, Zen 2 Threadripper is still a good enough product for its goals of a shit load of cores. Zen 3 didn't increase core counts, but Zen 4 does. My suspicion is that AMD probably overproduced Zen 2 TR and is letting its supply run down until it can launch another category redefining product built on Zen 4. Plus, it would look even better in their comparison charts.

I could definitely see AMD not bothering with a Zen 3 Threadripper. The current TR socket mess doesn't look fun to update, and there's a new socket right around the corner anyway.

But AMD is going to want to establish consistency with OEMs still, so Intel not having anything competitive right now is merely a reduction in pressure on AMD, not something that eliminates it entirely. A Zen 3 Threadripper Pro as a result still seems more likely than not, even if consumer Threadripper doesn't get a Zen 3 treatment. I'm sure AMD would also prefer to simplify the chiplets its fabricating as well. Why keep making Zen 2 chiptlets for TR Pro specifically and nothing else?

I occasionally watch the same YouTube channel, and I have not gotten this same impression. Just a month ago, he said that at least Zen 3 Threadripper Pro and Zen 4 Threadripper were still planned[1]. Can you point to the piece of content where he suggests this?

[1] https://youtu.be/TX6bREO3Nd4?t=380

I'm very sorry but I can't. I listen to podcasts while I do chores so ... yeah.

edit: So, I'm sure there's some other older thoughts/info presented on his podcasts, but it's the latest thing I heard from him talking about the subject:

https://www.youtube.com/watch?v=b8gVyjvSZkg&t=5713s (12 minute section )

Sorry if this came across as me saying he is leaking there's no desire to launch Threadripper. To my knowledge he did not leak such a thing, but he was talking about Threadripper not making sense https://www.youtube.com/watch?v=b8gVyjvSZkg&t=5713s in a new podcast episode. I have not conveyed this well enough in my original comment.

...What? But why!

Sigh, less and less choice every day.

EDIT: What about this potential leak? https://www.digitaltrends.com/computing/amd-to-release-ryzen...

Probably because they're server cpus that can be sold for more, it's too bad if this is true, I really want more PCIe lanes, though the threadripper 128 lanes are overkill, I'd be fine with 40.

Other than that, performance is plenty on a 5950x for all my imaginable use cases.

Sure but I'd be paranoid about not having ECC RAM whereas Threadrippers support it officially. But I am sure there's a curated list of Ryzen motherboards that do in fact support ECC...

Pretty much. https://hardwarecanucks.com/cpu-motherboard/ecc-memory-amds-...

Much appreciated and I bookmarked it. Thanks.

The Gigabyte B550 VISION motherboards claim to support ECC both on box and on the "key features" page[1]:

  Reliability
  ECC Memory
  To protect against data corruption
  Error Correction Code (ECC) memory corrects errors in your data
  as it passes in and out of memory to ensure reliability for
  critical applications.

[1] https://www.gigabyte.com/Motherboard/B550-VISION-D-rev-10#kf

Thank you. Any observations on whether the CPU performs worse? It's a widely circulated meme that AMD CPUs hugely benefit from as fast a RAM as you can put in, whereas ECC is a bit slower.

ECC is not inherently slower. Registered modules are slower (by the added latency of the register). ECC-or-not and registered-or-not are two orthogonal features of the DIMM and in theory all four combinations are possible. On the other hand non-registered ECC is somewhat rare (and for some reason often ridiculously expensive) and registered non-ECC probably is not manufactured at all (it does not make much sense).

I see. So I should chase the fastest unregistered ECC modules I can find then? I'll do that when the time comes.

I just tried out a Vision D-P (B550) and the IOMMU groups were terrible. All PCIe cards were in one big group, so good luck passing through specific hardware to a VM. It might not matter in a workstation or it might be fixed by a BIOS update, but shrug.

I went with an Asus instead. The Asus manual says "ECC memory support varies by CPU" which is questionable for the 5700G APU I went with, and I didn't feel like digging to find out if harder to obtain ECC UDIMMs would be beneficial. (This was for a router / personal server. My workstation is ECC on a libreboot/KGPE)

> My workstation is ECC on a libreboot/KGPE

Is it all-AMD? If it is, could you link to your setup? I am looking to build an all-AMD workstation this year and I need all the PCIe lanes that I can get.

This would usually be a tangent, but it's quite apropos to the main topic. My workstation is the last amd64 hardware without ME/PSP. It's a few generations old - dual Opteron 6380's with 112G of RAM. I couldn't get it to train with 8 sticks of RAM, but it's solid and reliable with 7. Last time I stuck a Kill-a-watt on it, I think it draws around 140W at idle. An alternative to it would be something like the Raptor Talos.

https://libreboot.org/docs/hardware/kgpe-d16.html

https://www.coreboot.org/Board:asus/kgpe-d16

It's got 5 usable PCIe slots (including the physically-flipped "PIKE" one), I think a total of 40 PCIe lanes from both sockets that I believe you could theoretically split out with the maximum amount of bifurcation, since the BIOS is Free Software.

My new Ryzen build is for a personal server that resides at a lower trust level, and targets around 30W draw. B550 motherboards generally have around 28 PCIe lanes going to expansion slots (20 from the proc, 8 from the chipset). The best you can do for number of slots is to bifurcate the main "graphics" x16 into 3 x4 slots (maybe 4 with a non-APU) using a "VROC" card, but the BIOS has to support that (Gigabyte and Asus seem to generally, but check manual). I believe X570 motherboards have a few more PCIe lanes coming from the chipset.

If you really need more PCIe lanes, I hear Threadripper/EPYC is the way to go. But I don't have any personal experience. If you just need more PCIe slots, you can find PCIe x1 -> four slot switch-based splitters inexpensively on ebay/aliexpress.

I see. Thanks a bunch, definitely bookmarking your comment. I have a bit more requirements for hardware but your setup might actually turn out awesome for a server.

> Last time I stuck a Kill-a-watt on it, I think it draws around 140W at idle

Another tangent (sorry!): do you know of good power meters that don't have to sit right at the outlet? I really need several but I want them to have extensions because I want to glue them on my wall and monitor them in real time, not having to stick my ass in the air while trying to crawl under my desk (where all the outlets are), just so I can see the measurements.

I'd probably just get an appropriate power strip and some Kill-a-watts. Alternatively you could look at smart devices with energy monitoring. The TP-LINK HS110 connects to wifi and measures current, power, and total energy with a community-documented local-network protocol [0], but they have been discontinued and prices shot through the roof. The replacement is KP115 but I have no idea if it still uses the same protocol or not. And I have no idea how accurately any of these devices handle weird (eg non-resistive) current waveforms.

[0] it wants to connect to "cloud" as well, but works fine without giving it Internet access.

I honestly don't care if my machine miscalculates, I don't run any important workloads. All content is in some cloud, all code is on github, all configuration is on github and by definition everything is hashed and stored in multiple locations.

EDIT: miscalculates as rarely as a computer normally does

Well, that's a solid point. I will still consider ECC RAM for my home servers but you are right that for a workstation it is a bit too paranoid.

Cheers, though I would say that I don't like that ECC isn't the standard, there's nothing but segmentation stopping ECC from being mainstreamed other than that it would make a lot of server gear redundant for a lot of applications, but as I said, it really doesn't matter for my application. Something that should be online for extended periods of time though, go ECC

Silicon supply and engineering resources are a major issue for HEDT platforms.

Thread Threadripper uses a massive amount of high quality silicon, Epyc has much better margins.

Supporting HEDT is living hell for Intel/AMD. You're often using all kinds of consumer hardware with otherwise enterprise grade gear. Addressing every little PCIe device issue and catering to a very very small market has very little ROI.

Not really, that's not how Threadripper works thanks to the chiplet design. It's not necessarily "high quality silicon" since it has a higher TDP budget. But regardless remember that in this case the 5800X and 5950X are also competing with Epyc for that same 8c chiplet silicon.

For something like Intel's HEDT then yes, you're absolutely correct. That involves HEDT-specific binning and it eats into Xeon-W profits very directly. But for Threadripper it's not nearly as clear cut. In fact it seems more like it's silicon that failed to validate for Epyc since it's half the IO die of Epyc (for the non-Pro Threadripper anyway).

Otherwise you're talking about AMD just taking Ryzen-quality chiplets, putting more of them on a substrate, and selling them at a huge markup.

Take the Threadripper 3970X as just a simple example here. It's 4x 8c zen2 chiplets that can hit a peak of 4.5ghz at 280W TDP. Meanwhile the Ryzen 3950x is a 2x 8c zen2 chiplet that hits a peak of 4.7ghz at 105W TDP. So 3970x is 2x the silicon of the 3950x, but lower quality silicon, and AMD charged more than 2x for the 3970X.

That's the brilliance of chiplets and Threadripper. It's Ryzen-class chiplets with half of an Epyc IO die (the IO die being 12nm GloFlo means it's not really competing for high-end prices anyway). It's binning they're already doing for they're primary product lines, not a specialized HEDT-specific process.

My understanding (I got this from listening to The WAN Show, but I don't have a timestamp off hand) is that they want to stop selling them to consumers directly (rather than to OEMs), presumably because volume was too low to support it.

Reply on news.ycombinator.com