AMD ran their benchmarks using a hidden administrator account that bypassed a bunch of Windows security features. 24H2 was supposed to bring those performance benefits to regular accounts.
Last edited:
The Zen Thread
- Thread starter w00key
- Start date
There was a thing early on about having to run the benchmarks in administrator mode, and I've been kind of wondering if maybe there was a disconnect between the people doing the benchmarking and the people selling the chips. The benchmarkers came up with the fastest number they physically could, and failed to notice that the old chips got just as much of a boost out of Admin mode. Then the execs ran with the bogus numbers.
It still ended up being a lie, and AMD should still be held culpable, but I wonder if maybe it wasn't done with actual intent. That's the sort of foulup I've come to expect from their software teams.
I still think they're pretty good chips: the same performance in half the power, plus a very good AVX-512 implementation. Considering they weren't upgrading process nodes, that's some sharp engineering. It was just oversold.
edit: and then of course the 9800X3D just melts face. That is a damn good chip.
I've seen a number of reviews about this over the weekend... to be honest, it looks fairly impressive. It's not going to be a high end gamer box but I bet it can compile the heck out of some code, serve up some stuff, and the like...
Made in Hurry
Ars Praefectus
In the same line of thought, what do you think about those Aliexpress boards with engineering sample CPUs soldered on them, some of them are quite reasonable priced and could make for a decent machine. Level1 has featured some of them earlier.
That I'd never, ever buy one, because I wouldn't know what I was actually getting, and wouldn't ever get the same thing twice.
Engineering samples are problematic, because they are not officially supported. Future firmware updates could brick them, or important bugfix updates could fail to apply. So if you ever have to rely on an engineering sample machine, you should try to narrow down its purpose and lock down its configuration as much as you can.
So I'm not sure what is going on with my computer:
All settings are default (mostly auto in everything?). EXPO is on.
I've tried two different memory kits (both on QVL list) from different companies (Lexar/G.Skill), same problem. Slots are A2/B2.
- When I cold boot it, it works fine.
- When I restart after having been on for a while (not sure of the trigger), it... doesn't complete and I get a yellow light on the motherboard (Memory not detected or fail). I have to power it off.
- If I restart immediately after a cold boot, it does, but the UEFI screen appears three times before it finally boots into Windows.
All settings are default (mostly auto in everything?). EXPO is on.
I've tried two different memory kits (both on QVL list) from different companies (Lexar/G.Skill), same problem. Slots are A2/B2.
If it's AMD, I had a 2600X and 5800X that did that and otherwise worked fine (eventually, after many BIOS updates). As long as you aren't having other stability issues I would just put it down to AMD's crappy firmware and work around it.
I had an Intel that did something similar (i7-6700k) and it ended up being a MB software issue that was eventually fixed by Gigabyte.
I had odd memory issues on a X570 board when using XMP (the memory didn't have EXPO) and I had to dial in 1.4V on VDIMM manually, never had problems again. This was for DDR4 at 3600MHz. If XMP carried the voltage setting, it wasn't applied (as diagnostic tools showed.)
Maybe I'm confused, but I thought EXPO was ASUS' method of reading XMP data and translating it to AMD settings? They can't call it XMP because of copyright, I gather.
Does any RAM have EXPO built in?
EXPO is an AMD thing actually, not an Asus thing.
https://www.amd.com/en/products/processors/technologies/expo.html
Newegg even has a filter for it.
https://www.newegg.com/p/pl?storeNa...=Desktop+Memory&Order=3&N=100007611+601426524
Yeah, it's AMD, as I mentioned EXPO and this is the Zen thread. MSI X870E and 9800x3D. No other issues that I can tell.
That's weird, my Gigabyte X470 board that I bought in 2018 grabbed the XMP settings from two different sets of DDR4-3200 RAM, and it was rock solid stable at that speed once I upgraded to a 3700X which actually supported DDR4-3200 (and then a 5800X two years later as prices started to drop). With a 2600X, it wasn't happy at anything higher than 2933.
As for my current build, it's a MSI X670E Gaming Plus with a 9900X and some 2 x 16 GB T-Force DDR5-6000. It runs perfectly fine all day with a full load of Folding@Home at EXPO speeds, and I am very happy that I spent 15 or 20 minutes going through the motherboard's QVL list; it meant a lot less frustration, especially since MSI lists which RAM is actually certified to hit those EXPO speeds!
I did the same thing, that memory was on the QVL. But nevertheless, the 1.35V it was fed by default was not enough, I had to set it to 1.4 manually.
Somewhat related, I've been playing around on a week off with RAM settings. As it happens, 1.35V is good for the Expo setting at 6000. And not a micron more. So I was lucky that everything Just Worked, because at 1.35V I can't even set it to 6200 without a bazillion RAM errors. On the other hand, at 1.4V I can do 6400 and drop the CAS by 2. CPUs may not thrive on Moar Volts! but RAM still seems to.
The big APU, "Strix Halo" has been reviewed by NotebookCheck. No shocking surprises: CPU and GPU performance are both respectable to excellent.
IMHO if this chip is a breakthrough of any kind, then in terms of form factor more than anything. AMD does make a good stride towards closing the gap to Apple Silicon, but doesn't achieve enough to really change the status quo.
Price and availability are still pretty much unknowns, though.
IMHO if this chip is a breakthrough of any kind, then in terms of form factor more than anything. AMD does make a good stride towards closing the gap to Apple Silicon, but doesn't achieve enough to really change the status quo.
Price and availability are still pretty much unknowns, though.
I'm about 2/3rds of the way through that review. It looks excellent. Also glad I did not wait for it given my laptop needs are more thin and light than ultimate performance, but that in no way diminishes my opinion that Strix Halo looks excellent.
Unless AMD has a big surprise up their sleeve, I read it as "looks expensive". There's probably some room in the market for a competitor to the M lineup, but they're not going to suddenly start selling well at MBP prices overnight, even if the chips are great. I don't actually know what the strategy should be; possibly sign up a Tier 1 manufacturer to make a bespoke, classy laptop out of it, rather than hoping their partners make shit stick?
I expect it will be a super-premium part. It's probably using a lot of silicon, and nothing else in the x86 space can do what it does, so I imagine AMD will price appropriately.
They surprised me some with the 9800X3D, though, so maybe we'll get lucky. If they keep the price down, they could move quite a lot of them, I would think. That strikes me as a super appealing chip for a mini-PC. Not as much for a laptop, because it's so hot... 80 watts in a laptop is pretty dismal, both for comfort and battery life, but that's easily handled in a NUC-alike.
As someone pointed out in the front-page article comments, the splittable VRAM might be very interesting for running AI models. It won't likely be as good as Apple Silicon, and not as good as a GPU with the same amount of RAM, but if you dedicate, say, 64 or even 96G of system RAM to the GPU, that could load a very large model, and possibly run it at usable speed.
They surprised me some with the 9800X3D, though, so maybe we'll get lucky. If they keep the price down, they could move quite a lot of them, I would think. That strikes me as a super appealing chip for a mini-PC. Not as much for a laptop, because it's so hot... 80 watts in a laptop is pretty dismal, both for comfort and battery life, but that's easily handled in a NUC-alike.
As someone pointed out in the front-page article comments, the splittable VRAM might be very interesting for running AI models. It won't likely be as good as Apple Silicon, and not as good as a GPU with the same amount of RAM, but if you dedicate, say, 64 or even 96G of system RAM to the GPU, that could load a very large model, and possibly run it at usable speed.
I'd happily take a laptop based on that processor as a work machine (with 64 or 128GB RAM), or for my teams next machine given they are typically offered Dell workstations with crappy dGPUs. Would be even more interested if AMD sorts out ROCm for the GPU and it gets good support with Tensorflow, pyTorch etc.
AFAIK that's on it's still improving and getting more and more usable.
https://www.phoronix.com/review/radeon-cuda-zluda
https://www.phoronix.com/review/amd-instinct-mi300x-rocm6
https://www.phoronix.com/review/linux-opencl-eoy2024
Strix halo should be very good at running large AI models due to its unified memory, yes. Keep in mind that it’s only a ~4060 in performance though. I run the 32B qwen distill of deepseek R1 on my mac M4 max with 64GB RAM and get around 17 tokens/sec, which is usable but not exactly great, particularly for a thinking model where you need to wait for all those CoT tokens to generate for every question.
If the AI models are run at int8, then you've also got the AMD NPU available for '50TOP' + whatever the GPU brings to the table.
50 TOPs isn't much compared to any GPU, those laptop NPUs are primarily for energy savings. The desktop RTX4060, which strix halo should approach in gaming performance at least, offers 240 TOPs. That won't directly transfer to AMD of course, just an example of GPU performance compared to NPUs.
Also misspoke earlier and for some reason the forums won't allow me to edit it, my laptop is a M3 Max.
Also misspoke earlier and for some reason the forums won't allow me to edit it, my laptop is a M3 Max.
Whelp:
https://wccftech.com/amd-ryzen-7-9800x3d-now-experiencing-widespread-cpu-failure-incidents/
Apparently the 9800X3D is under the weather. AMD has acknowledged the problem, but no official statement.
Now I have to question if Avowed is crashing because of the CPU, or because of the game. Guess I'll go back to the 12600K/B580 for a bit.
https://wccftech.com/amd-ryzen-7-9800x3d-now-experiencing-widespread-cpu-failure-incidents/
Apparently the 9800X3D is under the weather. AMD has acknowledged the problem, but no official statement.
Now I have to question if Avowed is crashing because of the CPU, or because of the game. Guess I'll go back to the 12600K/B580 for a bit.
I don’t think it is time, but uptime, that is the issue, and my 9800X3D hasn’t been used too much. Been using the 12600k and 12900k a bit 
They also mention 50ish examples, so it doesn't yet appear to be a huge plague or anything. But, dammit, AMD. Can't you ever get your freaking launches right?
Fifty-something of how many total in use? Is there any reliable data on such product failure ratios at all?
As bad as it is, but are we tallking about ten times the normal falure rate, or a hundred times? Or one times the normal failure rate? Do we have anything more solid than a screaming headline?
Phoronix just did a roundup of 9000 vs Intel Core Ultra Series 2 benchmarks with all the recent kernel & BIOS updates.
9950X just DOMINATES in overall performance, wow. Not too big a surprising, especially for the multi-threaded benchmarks.
9950X just DOMINATES in overall performance, wow. Not too big a surprising, especially for the multi-threaded benchmarks.
I can confirm that a 9950X under Linux is a powerhouse even in "105W" ECO mode. Especially if you can leverage AVX-512.
You can keep trying to make Linux on the desktop happen, but the only ones having any success at that are Valve with SteamOS. Games don't care about the second CCD.
Speaking of single CCD...
View: https://youtu.be/RYLUBvISa7I
Thermal Grizzly is now offering delidded 9800X3D CPUs with a 2yr warranty. Inside the EU, 699 EUR. Outside the EU, they remove the VAT, so around 585 EUR. They are not doing overclock testing and binning.
Delidding at moderate scale is a two step process. The first step involves putting all the CPUs into TG's delidder jig, baking to 170C, then applying one cycle of the delidder jig. After that, they remove the CPUs from the delidder jig and place them in a custom tray. upside down. Back into the oven for a bit until the IHSes fall off into the tray.
After they cool down, they test the CPUs on a B650E motherboard with a direct-die block and liquid metal. He notes that they don't polish the dies after the testing phase as it doesn't matter - you're going to stain the dies with liquid metal on your first application anyway. They include a USB stick with a picture of the CPU, IHS, and the serial number, and a picture of the Cinebench/hwinfo results during a load test.
If you were planning to delid and watercool, might be a good way to go.
Speaking of single CCD...
View: https://youtu.be/RYLUBvISa7I
Thermal Grizzly is now offering delidded 9800X3D CPUs with a 2yr warranty. Inside the EU, 699 EUR. Outside the EU, they remove the VAT, so around 585 EUR. They are not doing overclock testing and binning.
Delidding at moderate scale is a two step process. The first step involves putting all the CPUs into TG's delidder jig, baking to 170C, then applying one cycle of the delidder jig. After that, they remove the CPUs from the delidder jig and place them in a custom tray. upside down. Back into the oven for a bit until the IHSes fall off into the tray.
After they cool down, they test the CPUs on a B650E motherboard with a direct-die block and liquid metal. He notes that they don't polish the dies after the testing phase as it doesn't matter - you're going to stain the dies with liquid metal on your first application anyway. They include a USB stick with a picture of the CPU, IHS, and the serial number, and a picture of the Cinebench/hwinfo results during a load test.
If you were planning to delid and watercool, might be a good way to go.
Last edited:
Linux on the desktop has already happened. It's done. It's not super popular, but it exists, it's pretty decent in most cases, and a noticeable fraction of us use it. Enough, in fact, that we can actually have a meaningful conversation about it, all on our lonestomes, without needing to worry about the opinions of those who don't.
The mass market may not care, but since when has the Ars audience ever been mass market?
Mostly not, which is much of why I'm often so dismissive of big core counts, but at least a few games do. Someone here, possibly @hobold, uses a 16-core chip for big Civ 5 games, as an example.
Last edited:
Whatever helps you sleep at night.
From my perspective, I see a small but noisy group arguing over which bizarrely-named distribution is "best".
Yeah, and there's a whole forum just for that.
Which is not relevant in this case, because they're talking, in the Zen Thread, about how Zen CPUs compare with Intel CPUs running under Linux.
You jumping in and ranting about Linux, unasked-for, is the off-topic bit. Mine too, for feeding the troll.
(edit: when I originally replied, IceStorm's comment was just the single paragraph complaining about Linux. The video and other stuff was added later.)
Last edited:
AM4 continues to not be dead:
https://videocardz.com/newz/amd-lau...5g-apus-am4-socket-and-vega-graphics-lives-on
There must be a mountain of Cezanne dies lying around, as AMD has now introduced no less than SIX new AM4 CPUs, all with Vega graphics. There are 3 pairs - 8C/16T/Vega 8, 6C/12T/Vega 7, and 4C/8T/Vega 6. Each pair has a 65W(G) and 35W(GE) variant.
https://videocardz.com/newz/amd-lau...5g-apus-am4-socket-and-vega-graphics-lives-on
There must be a mountain of Cezanne dies lying around, as AMD has now introduced no less than SIX new AM4 CPUs, all with Vega graphics. There are 3 pairs - 8C/16T/Vega 8, 6C/12T/Vega 7, and 4C/8T/Vega 6. Each pair has a 65W(G) and 35W(GE) variant.
Made in Hurry
Ars Praefectus
Linux has been perfectly fine and i like using it, it is not the scary monster it once was, but i agree, i wish the community would actually pool together and decide on a distribution or two instead of making a new distro every week. The fragmentation is what kills it and will probably amount to nothing in a decade, like the previous one.