The Zen Thread

Status
You're currently viewing only Mental Gear Reduction's posts. Click here to go back to viewing the entire thread.
Wonderful engineering at individual component level, shithouse effort at combining into a coherent well performing complete machine, especially at the infinity fabric / hypertransport memory. You have to wonder if it is even capable of handling DDR5 8000 when it turns up. The IO die needs a serious overhaul (its what limited zen4 to mostly DDR4 3200 also).
This strikes me as a pretty weird comment. AMD is doing better with their CPUs than anyone but Apple, but you're throwing up a wall of snark here?

I mean, sure, the 9XXXs, so far, aren't living up to the marketing claims, but they're not having stability problems or anything. They're just not as fast as the marketroids said. If they'd marketed them as "a few percent faster than Zen 4, with large AVX-512 improvements, and much better heat efficiency", I think the market would have been pretty happy with them.

The problem seems to be more with poorly managed expectations, rather than actual silicon failures, while their chief competitor is both figuratively (250+W??) and literally (poof!) burning up their chips to try to stay competitive.
 
GN's review literally has a chapter entitled "Stability Problems".

The only Zen launch that has ever gone well is Zen+. All the rest have been a shitshow of one form or another.
He does mention that, while they struggled pretty hard with RAM timings, the board they were using had a BIOS update that they hadn't tested, so it might be fixed already. And they were able to get it working well, though it took far more effort than it should have.

If the memory issues are fixed with the new BIOS, then I'd call this a pretty painless launch, overall. People are mad about the performance claims being overstated, but that's a marketing problem, not one with the silicon. If, however, the RAM problems aren't fixed quickly, then I'd be more inclined to go with the 'shitshow' description.

So far, it's still a lot better than Zen 3. The USB problems there were unfixable by users, and it took such a long time for them to finally get their crap together.
 
Which means that fixing the scheduling hasn't addressed underlying architectural issues and IO die issues causing Zen5 to underperform comapred to AMD's PR claims (and as a reminder AMD stated Zen5 was a significant re-architecting of Zen).
There aren't really "architectural issues" here. The chips work fine, except for some reported memory timing issues, which will probably be fixed soon if they haven't already been.

The real issue here is the marketing. AMD's marketing department needs some immediate management changes, and hopefully Dr. Su will be smart enough to make them.

As I said earlier, everyone would be fine with the 9000-series if AMD had sold them as 'marginally faster with most code, major AVX-512 improvements, major efficiency improvements'. They wouldn't have commanded a premium price, but folks wouldn't be particularly upset after buying one, either.
 
  • Like
Reactions: rodalpho
Yeah, and as I said in the risk-v thread about x86 going away as an option, that will happen around the same time the sun evolves into a red giant. Leopards don't change their spots, tigers don't change their stripes and AMD won't change their sleazy marketing, no matter who is in charge of the company (and Lisa Su has been CEO since 2014 and has done nothing to rein marketing in thus far).
They were fine in the Zen 3 and 4 era.

You really seem to have a bug up your butt about AMD.
 
It costs many millions of dollars just to tape out a chip, let alone produce it, and there was virutally no performance benefit to doing so.
There's massive AVX-512 speedups, plus substantial efficiency gains. Spending only 60% of the power to do the same thing is pretty damn attractive, especially compared to their main competitor. And under Linux, there are apparently big wins, and Linux dominates computing except on the desktop.

The chips seem pretty good to me, you just have to know what you're buying them for. They seem more attractive than most of their 7XXX counterparts, simply because they run so much cooler. That's a fantastic feature. People who buy them for Windows aren't going to find them much faster than the 7-series, but they're definitely not worse, so they seem like a safe buy in most cases. (except gaming, where you still want the 7800X3D.)

If a 7XXX chip is cheaper and you're on Windows, you'll get better bang-per-buck there at the moment, but you'll spend more effort cooling it.
 
Last edited:
So they are gearing up AVX-512 right at the same time intel is preparing to deprecate it for AVX10, yay I guess. Outside of RPCS3 there is nothing on my system that uses AVX-512 and I'd be shocked if RPCS3 didn't move to AVX10 since the devs did an article about how they use AVX-512 and they actually use very little of it.
The efficiency gains alone make the chips worthwhile.

You seem incapable of admitting that AMD can ever do anything right. Did they hurt you or something? Do we need to get a doll?
 
Define "pricey" ?

Most of the ones I saw were $500+, but I was also looking for units with the better built-in graphics, which AMD charges a lot for.

I think there'd be a big market for RDNA APUs on the cheap, but perhaps their fab allocations aren't presently worth using up on low-margin chips.

I'll keep an eye out, I bet an RDNA2(3? 3.5?) NUC would make a pretty sweet little Linux box. I have to dual boot now, and don't usually bother.
 
Sorry, will be more precise. Zen 5 is not a bad product, the desktop Zen 5 CPUs are.
But they aren't. They're at least as fast as the 7XXX chips, usually a little quicker, while using only 60% of the power. And you get major wins in Linux and with AVX-512 code, both of which are niche use cases, but definitely do exist.

AMD marketed them as better than they are, but that doesn't make the chips bad, it makes AMD marketing deceptive. Zen 5 didn't live up to expectations, but expectations were set incorrectly high.
 
  • Like
Reactions: steelghost
Linux isn't the type of OS that would turn off a setting called "Memory Integrity" to gain speed at the expense of reliability. I will need a more in-depth explanation of this hypothesis before I can believe it.

(The "conditional branch security mitigation" really is a very plausible explanation as far as processor architecture and the timeline of historical security vulnerabilities are concerned. But yes, that cannot be the whole picture, because the Windows patch clearly does additional things to improve performance further.)
I haven't looked it up to be certain, but I think memory integrity on Windows is using a hypervisor to hide some of the most critical Windows components in a special internal VM. That slows the rest of the computer down some, as calls into that trusted zone take a great long while.

IIRC, turning off memory integrity gets you about a 10% overall boost to raw CPU performance. In exchange, you're less protected against malware.

I don't think Linux has anything analogous. It's all just kernel memory.
 
On these or newer CPUs, the overhead should be nowhere near 10%.
I mean, I was looking into that relatively recently (but not recently enough to be absolutely certain my memory was correct) to see if I wanted to turn it off on my machine, and the links I found about it claimed you'd get about a 10% boost if you disabled memory integrity, and they're recent-ish articles. I'm sure they haven't tested on Zen 5, but the articles were written using Zen 3 as examples.

I can probably track down whatever I read if you're really doubtful.
 
Last edited:
Micro Center continues their AMD relationship with the 7600X3D, a Micro Center exclusive:
https://www.tweaktown.com/news/1002...-has-big-micro-center-shaped-catch/index.htmlGoes on sale September 5 for $299. There's a lot of complaints about that price, but it's Micro Center - they'll probably bundle it for far less.
IIRC, pretty much everyone thought the list price on the 5600X3D was also very high, so they're probably just duplicating what worked before.

And that fake CPU is really impressive. Makes me wish there was decent computer retail around here, because Internet purchases just got riskier.
 
It's remarkable that you can install Windows 11 twice on the same hardware from the same ISO in the same way yet get double digit differences in performance. Maybe a race condition in the installer where if one thread finishes first something gets skipped? Would be interesting to see a diff of the two disks.
They do say that adding the scheduler patch fixes it either way, so you're not stuck with Windows sometimes being kinda broken.
 
I wonder if the X3D hope has created the Osbourne Effect... I haven't watched this yet, it just popped up when I cranked up my machine for the day.
One thing he's missing is that the 9000s are about the same speed as the 7000s, but use about 60% of the power to get there.

So they're worth some premium, just nowhere near what AMD wants to charge.
 
Phoronix ran their benchmarks with the 105W cTDP on the 9700X and 9600X. A small jump on average, but some big jumps on specific workloads:
So, for the most part, it's not really worth raising their TDP: +35% power for 3 to 6 percent performance. AMD is delivering them at the right wattage, and there isn't a lot more performance in that silicon.

It's interesting that the 9700x just barely edges the 14900k at the higher TDP, though. And still doing it with one hell of a lot less power.

These aren't bad chips, they're just not much an improvement on Zen 4. You get efficiency and AVX-512. When the prices get closer to Zen 4, they'll be perfectly good purchases.
 
while Zen based boxes are often (well?) above 30W. The difference is real.
35 watts, 24x7x30, is about $4.28/mo at 17 cents per KWh. So that's enough to notice on the power bill.

However, being reasonably sure that CPU won't melt is a big deal, at least for me, and then running so much cooler under load offsets some of that, in addition to making the computer quieter when you're using it heavily. The more often you load the computer with heavy computational tasks, the faster an Intel chip would catch up on total power bill.

I suspect I probably pay about a $3 delta by running AMD (just kinda handwaving it), but I'm happy enough with the Zen 3 platform at this point that I'm unbothered by the expense. I spend a great deal more money on things that please me much less.
 
  • Like
Reactions: Baenwort
Thank you. Much appreciated.

I get 3.15 on the column Pkgwatt on this 4650GE just polling turbostat during idle'ish. Mint-22 Kernel 6.8.0-44
FWIW, on a 5600X mostly-idle server, running one VM, I'm showing about 13 PkgWatts.

edit: a 5800X is pretty close to the same, also not doing much. More variance, because it's got more stuff idling. It usually reads about 12.8, a wee bit lower than the 5600X, weirdly enough. Peak was 13.8 in the 30 seconds or so I looked. It doesn't have any VMs, all native stuff, so that might be why it's a scoche lower.
 
Last edited:
Yes I understand. My point is that Zen 5 isn’t a bulldozer.
Not technically, but they seem to be in terms of sales. They're too expensive and people aren't buying them. There's nothing really wrong with the chips, but they're staying on shelves like Bulldozer did, because Zen 4 is the same speed for less money.

edit: as soon as they get cheaper, they'll be fine. No reason not to buy them. The only thing actually wrong with the 9000 series is the price. They're a tiny improvement over the old chips for most people, and a major improvement for a niche audience. I don't think there's any case where they're worse than their direct predecessors.

They're not broken, and not a technical failure. They're just priced wrong.
 
Last edited:
The best AM4/value CPU for gamers is the 5700X3D. Basically the same thing but much cheaper and maybe 5% slower.

Actually maybe tbe 5600X3D, the same thing with only 6 cores. But anyway not the 5800X3D.
Well, okay, I should have said 'fastest', rather than 'best'. Still a very strong product, and it seems very strange to me to stop making it.
 
The gamers who are complaining that Ryzen 9000 is too small an improvement for too much money must be very rich.
People are just, correctly mind, complaining that the new generation is a shit deal compared to the old generation. If the new chips were priced about the same as the old ones, nobody would much care, and the market would gradually transition.

Your mischaracterization is way off base.
 
Sure. I don’t know why that’s an important point to make, but sure. “No bad products, just bad prices”.
Even in this thread, there's been a ton of hate toward these products. I'd have to go back to look, but I think you may have been one of several who were being very critical. Collectively, you all were kind of labeling them as shitty, lousy product failures.

They really aren't. They're just as good as the equivalent Zen 4 models, and better in some areas. If they were priced like Zen 4, they'd be fine. It seems like everyone's mostly caught up on that now.

AMD's deceptive marketing, on the other hand? That was a shitty, lousy product launch failure for sure. There should be demotions or firings within AMD.
 
I thought we were talking about the AMD GPUs?

Edit: To be clear, when you’re talking about the data center and AI, we are talking about AND GPUs.
Right, and they make very little with their datacenter GPUs. Almost all their datacenter revenue, at least to my knowledge, comes from Epyc chips being so much better than their Intel equivalents.
 
Funny thing - checking pricing in a few boutique PC outlets (Cyberpower, Falcon, Origin...) I am seeing the 9700X selling already at the same prices as the 7700X, and in some cases cheaper than the 7800X. Cyberpower has the 9700X just $26 above the 7700X, for instance, which seems entirely reasonable.

The market has spoken, basically.
Were I buying now, I'd pay a little more of a premium than that, actually, since the improved heat efficiency sounds great. A $26 delta seems more than reasonable.

My CPU-intensive applications are almost purely games, though, so if I buy in this generation, it would likely be a 9800X3D.
 
I am at a wait and see mode, but i am not likely to build a Zen 4 or 5 machine anytime soon unless i stumble over an open box return or such in the future for pennies which is how i am usually buying computers, when the shop clearly has made an error and i just snag it.

The hardware is simply too expensive in general, and i do not mean the CPU's, i think they're for the most part cheap for how much work that goes into them, but motherboards, decent PSU's and graphic cards. I would be satisified really with a really good APU at this point. No, the 8600G doesn't count, i mean a strong non-gimped APU.
Those $700 Playstations don't look so bad in that comparison, do they?
 
They do, as they pretend to be a PC in my opinion, and that is including the games. The console-approach to gaming back in the day was gameplay and fun. The PS3/Wii excelled at it with different addons. Guitar Hero, dance mats, motion sensors and fun, what a console SHOULD be i think which never makes either Microsoft's or Sony's current offering anything remotely interesting.

Yes, i am yelling at clouds :)
Well, I was more talking about the $700 price than the feature set. Other than that, I mostly agree with you, and I'm sure if we both yell at the clouds, things will change.
 
  • Haha
Reactions: Made in Hurry
I really want to ask again...who besides Factorio players are really playing games that are CPU constrained and not GPU constrained?
BG3 likes lots of CPU, and I'm fond of all sorts of sims like Dwarf Fortress. And I haven't fired it up in a long while, but I put a hell of a lot of hours into Factorio. (I will buy the expansion the instant they open sales.)

In general, NVidia cards benefit from a really strong CPU, because their drivers do more of the work there, rather than on dedicated hardware.

Starfield is another game that likes beefy CPUs, but I haven't and probably won't buy that, so I can't use it as an excuse.

Oh, and I'm real fond of emulation in general, and that takes all the CPU grunt you can scare up. I can't quite emulate a P200MMX with 86Box, for example, and it would be very nice to get that working.
 
  • Like
Reactions: CPX
Bit off topic, but I've had my eye on that, really good?
I haven't played it since February, and they've added quite a bit of stuff since, but I was pretty unimpressed with that early version. It had potential, but I didn't stick with it long at the time.

This is a sufficiently outdated take, however, that I wouldn't give it too much credence.
 
Mystical (author of Y-Cruncher and numberworld.org) posted this in the Overclock thread:

Too bad that the change doesn't have an on/off toggle to be able to compare between workloads.
Yeah, that kind of implies that doing actual work might have been faster the old way. We may be paying too much attention to benchmarks.

edit: although, high latency seems like it would mostly be bad, no?
 
He recommends the 770X for gaming because you get 95% of the performance (averaged) for half the power consumption or less. It's not an unreasonable position.
Particularly since so many games are bottlenecked on GPU anyway.

Doubling power consumption for 5% performance would be an absolute nonstarter here. This room gets too hot as it is.
 
I am considering just getting a separate machine for idle/low power, 24/7 workload.
The 13ish W I see consumed on my 5800X have a runtime cost of about $1.60 a month, at 17 cents per KWh. ( ((13 * 24 * 30) / 1000) * 0.17). If your replacement PC is $150, it would take about 7.8 years to pay for it, and that's ignoring the fact that the mini PC CPU itself uses some power and has a run cost.

If power costs more where you are, it might make more sense, but that's a slow ROI. Mini-PCs are built cheap enough that it might not last 7.8 years.

edit: the total system power will be higher than 13W, but I'm using that figure because that's about the max wattage you would potentially save. I'm assuming the rest of the components in both systems will pull about the same amount of power.
 
  • Like
Reactions: fitten
Yeah, up to three or four doublings might still be in the cards before CPUs based on silicon hit "the final fabrication node" or become entirely unaffordable. But these doublings will arrive later and later. I am currently not sure if I will get to see two or three more doublings before I run out of lifetime.
With the way transistors are becoming more expensive now, you might not see any more doublings, or maybe only one more, simply because the chips will cost too much to make.

Well, unless we can get something else working besides silicon, anyway.
 
I can't exactly go back and hunt bargain bins 20 years ago to verify, but that certainly wasn't a normal price for that SFF, and unless it's used, they are losing money on it, so an Apples to Oranges comparisons.

Where are you that you a Mac Mini starts at $850 and you get deal like that SFF PC?
Intel is stuck with a ton of chips they can't sell, so it's entirely possible that $190 price for a 12400F is the new normal.

That's an amazing little machine for under $200.

There's talk that they may update the Mini next month. If so, it might be pretty appealing, but while even the present Mini is probably faster than the 12400F, it's definitely not multiples as fast.
 
Status
You're currently viewing only Mental Gear Reduction's posts. Click here to go back to viewing the entire thread.