Hacker News Reader

Comment by rektide | Hacker News Reader

AMD has done extremely well with multi-chip(let) modules. Zen cores & zen clusters (on Core Chiplet Die, CCD) are wonderfully small, and a huge amount of the regular stuff cores do is relegated to the IO Die (CCX), which is not as cutting edge.

But wow there's a bunch of power burned on interconnect between CCDs and CCX. And now AMD's new southbridge, Promontory 21, made by Asmedia, is another pretty significant power hog, and the flagship X670 tier is powered by two of these.

There's absolutely a challenge to bring power down. I'm incredibly super impressed by AMD's showing, & they've done very well. But they've been making trade-offs that have pretty large net impacts, especially if we measure at idle power.

View on HN

CCX stands for Core Complex

CCD stands for Core Complex Die (and neither terms refer to the IO die)

Whoof, oops, thanks. I'd been using CCD and IOD as terms until this post, but had a "omgosh, I've been doing it wrong" panic & changed into what we have here. My mistake. Thank you for correcting us back!!

https://hn.algolia.com/?query=rektide%20ccd&type=comment

Manufacturing costs force them to go down the chiplet path. Its actually impressive they can remain competitive at all given TSMCs margin of 67%. [0] If Intel Foundry manages to keep up with TSMC, thats a lot of pricing power advantage over AMD. Or they could make lower-power CPUs that would be uneconomical for AMD.

[0] https://www.macrotrends.net/stocks/charts/TSM/taiwan-semicon...

But don't forget that TSMC's margins for massive bulk purchasers like AMD will be lower than the average.

Yup, obviously they are aware of AMD competition and can adjust their margins accordingly.

Does TSMC even have non-bulk customers at 7nm/5nm/4nm? I would have thought the mask costs are so high that it isn't economical except for the biggest companies.

Cerebras has probably bought less than 100 7nm wafers, total. Though they may have gotten many more sales than the last I checked. Tesla Dojo is probably about the same. I've seen loads of random chips like that on 7nm. I guess it depends what you consider "bulk".

7nm yes, 5nm/4nm not as much, apple pretty much bought out the entirety of 3nm. 7nm is a relatively mature process at this point.

Why is EBITDA margin better in this context than net margin? Wouldn’t manufacturing facilities have a ton of depreciation and amortization that should be incorporated into the costs?

Gross margin is the appropriate measure of per unit cost/profit. Yes, EBITDA is not a good measure at all (for part economics).

Intel can't figure out how to make the next generation of chips. They used to be able to innovate on a steady regular basis, but they're still trying to get 7nm right and they've been working on that for many years while TSMC is on to 3nm. Makes you wonder what happened to that company that they got so far behind.

I'm not up-to-speed on modern chipsets, but WTF is the Southbridge doing that it needs that much power? Is Thunderbolt going through it or something? I think of the Southbridge as a mostly ignorable part of the chipset (up until something goes horribly wrong).

Single word: PCIe.

Lots of fast PCIe lanes eat a lot of energy. A server motherboard contains tons of more PCI devices when compared to a consumer desktop systems, and they are not the cards, but the small units enabling the advanced features in servers, which are embedded on the motherboards themselves.

Thunderbolt is just a PCIe encapsulator of some sort, which can also do plethora of other things.

> Single word: PCIe

This explains Intel’s squandering of PCIe lanes for consumer desktops versus AMD’s generosity.

There's very little difference in the number of PCIe and M2 slots between Intel and AMD on their consumer platforms. The only difference really between AM5 and LGA 1700 motherboards is a lot more AMD boards have one M2 PCIe 5.0 slot, while only the very top Intel boards have this feature.

AMD has 24 PCIe 5.0 lanes directly from the CPU available for user, while Intel has 16 5.0 + 4 4.0. Cheaper motherboards might not expose all of those, or downgrade some to 4.0 to save on on-board components. In addition, both have more on the chipset, which is connected to (additional, reserved) PCIe 4.0 lanes. The best AMD chipsets have 12 4.0 and 8 3.0 lanes, while the best Intel ones have 20 4.0 and 8 3.0 lanes. An important point is that the connection between the chipset and CPU is twice as wide on intel (8x vs 4x).

So overall, AMD has more and faster IO available directly from the CPU, but less lanes from the chipset, and with a weaker connection to the chipset. If PCIe 5.0 drives become available and the transfer speed to storage is important, I'd say AMD is better, otherwise I'd say Intel has more IO.

> very top Intel boards

Where "top" means "most expensive". The Z790 board I recently purchased for around $300 was pretty barebones and lackluster (no TB, meager IO from ports and headers, wattage constrained VRM relative to 13th gen TDP, etc), but it was the least costly way to work with an Intel proprietary technology.

It'll last another four or five years, but it was my first Intel build since the slocket days, and likely my last.

AM5 boards aren't any cheaper than that. I'd say that Intel boards are generally cheaper in fact, though that has narrowed with Z790.

You do get to use the AMD boards for more than two CPU generations though.

Oh, I assumed that PCIe was on the northbridge. Of course PCIe eats energy.

The one contradiction I have to point out here is that server motherboards dont need big southbridges: the cores themselves have gobs of PCIe. 1 and 2P AMD Epyc server cores have 128 lanes of PCIe.

I wish modern chips did a better job of breaking down where power went. It'd be so interesting to know how much power is going to usb controllers, how much is going to PCIe. I'd also hope that they could do things like shut down parts of the chip, if there's no USB or PCIe devices plugged in. But these chips seem to have a pretty high starting place of power consumption. Although maybe it's in part because the first example were flagship motherboards with a whole bunch of extra things peppered across the board - fancy NIC chips, supplementary thunderbolt controllers, sound cards, wifi - so maybe there was just an unusual lot of extra stuff going on. But it has been shocking seeing idle power raise so much on the modern platforms. It feels like there's a lot of room for improvement in power-down.

To me the most impressive part is what incredible job they did in the years of intel dominance. They almost bulldozered themselves.

I see what you did there.

Please explain!

Bulldozer [0] was a major flop. Sounded too good to be true: 2 cores sharing the same FPU because integrated GPUs should be executing these instructions much faster anyway. Maybe it was just ahead of its time and they simply couldn't deliver

But it failed hard which coincided with Intel releasing a major winner with their last planar architecture - sandy bridge.

As a result, AMD spent years circling the drain and their stock dipped below 2 dollars. Some people made good money buying around that time.

[0] https://en.wikipedia.org/wiki/Bulldozer_(microarchitecture)

Edit: oh, I missed the exclamation mark haha. Oh well. Too tired to even feel ashamed

It was a pun: "Bulldozer" as in "(v.) to destroy" and also a former AMD microarchitecture which, as sibling reply mentions, was a flop.

Reply on news.ycombinator.com