The Intel Execution in [2024]

What and how this 'clock tree circuit' and vMin? Is the silicon of this circuit the one that's degrading, or is the whole chip degrading?
Clock tree is another term for the integrated on-die network which distributes clock tics to the entire chip. The "trunk" of the tree starts at the clock generator, which is onboard the CPU itself. The "limbs" and "branches" of the tree speak to how those clock signals are distributed around the die. It's more a tree in logical representation than physical layout, of course.

They weren't very specific about how the clock tree plays into the vMin conversation. One interpretation might be the clock tree network itself is somehow damaged by voltage (it's made of silicon transistors like everything else in the die) and somehow this physical damage results in missing or otherwise skewed / corrupted clock signals being transmitted. If the discrete components of a CPU get out of lock-step with eachother, everything immediately goes to hell. Think of latch and atomic operations where there are mandatory predecessor / successor steps for something to work, and then imagine when those get processed out of order. The user's perception of such a thing would be a hard lock, but it might come along with data corruption depending on what instructions were being processed when it stopped.
 
As per past processors (AMD Epyc 9684X), Phoronix released Linux HPC benchmarks for Granite Rapids. Specifically, the Xeon 6980P processors paired with MRDIMM 8800MT/s memory in an Intel AvenueCity reference platform. The 6980P is quite the heavyweight. Sporting 128 P-cores, a 2.0 GHz base clock, a 3.2 all-core turbo cock, 504 MB L3 cache, and a 500 W TDP, Intel has brought a howitzer to gunfight. AI and GPU performance notwithstanding, as indicated by the HPC benchmarks courtesy of Phoronix, Intel is back in the game in a big way.

The following results are some snapshots of popular HPC benchmarks. Consult the Phoronix article for a full account of all the results and test setups.
 
Last edited by a moderator:
One interpretation might be the clock tree network itself is somehow damaged by voltage (it's made of silicon transistors like everything else in the die) and somehow this physical damage results in missing or otherwise skewed / corrupted clock signals being transmitted...
That would be my gut feeling based on the reported errors being unpredictable, diverse, more consistent over time and leading to total chip failure. But why would these bits be susceptible? Did Intel make them different? I guess they won't go into details, unless someone sues.
 
Yup, we're just left to guess WTF at this point. It's interesting they mentioned the clock tree at all, because it obviously brings up all the questions you just described. How did they isolate the cause specifically to the clock tree network? What makes that network of transistors more susceptible to this voltage issue than any other part might have been? Asking the same question in reverse: why isn't this vMin problem causing damage to other circuitry?

So many questions we'll likely never know the answers to. :(
 
Not surprised that Intel 3 is genuinely good. I'm just annoyed that they didn't release anything for consumers using it. Desktops have spent the last three years still on 10nm, and Intel has faced multiple levels of scrutiny as a result. For one, they're widely known as 'inefficient' compared to Ryzen competitors. Two, all the negative press over Intel's 14th gen refresh offering basically nothing for consumers. Three, spending an extra year releasing products that are pushed so hard they're creating widespread reliability issues.

I mean, I get it, desktop is not their primary money making market, but it's still annoying all the same. And is still hurting them in terms of PR and reputation. And now they're jumping to TSMC N3B for desktop parts, which is likely not good for their financials and probably means that they cant release anything else in the consumer space on an Intel node at all til 18A is in proper mass production with scale.

I do think Intel is getting back on track, but man there's a lot of pain in getting there. Hopefully they start doing better financially cuz we kind of really need them to.
 
Last edited:
My guess is they were physically examining/scanning failed CPUs and noticed that commonality?
I would assume that would narrow down the potential failure points but there is only so much information you can get from scans.
I think nanoprobing would give them the confirmation.
https://www.thermofisher.com/blog/m...-analysis-technique-for-todays-tech-industry/
https://en.wikipedia.org/wiki/Nanoprobing

Smallest I've dealt with in legacy probe is 5µm, so not even in the same ballpark as what they are using now.
 
Last edited:
I actually don't mind the direction Intel seemed to go with this cpu if they really did bring down power consumption significantly. The cpus still show a max power of 250W, which is way too much, so I'm hoping you can lower the limit without affecting gaming performance, since gaming is not supposed to hit anywhere near that limit. I really do not care about doing cinebench runs to max out cpu power. There are a couple of things I'm curious/have concerns about.

One is the sort of "1% lows" advantage that they seemed to have, and whether they'll still maintain that advantage now that they've gone to a chiplet architecture. I kind of liked the idea of moving to intel just for an overall smoother experience. It's going to be really interesting to see detailed performance numbers.

Hyper-threading is one of those things that I'm not sure is necessary anymore, especially when you have tons of e-cores, but some games do benefit from hyper-threading. I honestly don't know if the windows scheduler really handles things well yet. I like the idea of having big P-cores for my game threads and e-cores for all of the windows threads so cache isn't being thrashed and my game threads aren't being interrupted. I just don't know how that actually plays out.

The lack of AVX512 is kind of ... not sure. It never took off, but now AMD cpus have AVX512, which means the next consoles will probably have AVX512. So whenever the next console gen launches you'll have games that might use it, which means PC versions of games may naturally support it. Will be interesting to see if there's a bit of a drop off for Arrow Lake in three to four years, or whatever it is.

Going to be super interesting. I haven't kept up on when reviews should start hitting. Does anyone know?
 
The lack of AVX512 is kind of ... not sure. It never took off, but now AMD cpus have AVX512, which means the next consoles will probably have AVX512. So whenever the next console gen launches you'll have games that might use it, which means PC versions of games may naturally support it. Will be interesting to see if there's a bit of a drop off for Arrow Lake in three to four years, or whatever it is.
This one particular point is interesting. I have friends that work in the industry and sometimes they request some very specific instructions, like bit rotations etc, that assist with encryption, or compression, or whatever niche use case it is. I think some prominent industry engines/tools also have a say in these instruction sets, so I'm a bit surprised to see at times if they get dropped. But given the power envelope they are looking at, perhaps this is the reason. It runs counter intuitive to what they are trying to achieve here. 512 eats power.
 
Hyper-threading is one of those things that I'm not sure is necessary anymore, especially when you have tons of e-cores, but some games do benefit from hyper-threading. I honestly don't know if the windows scheduler really handles things well yet. I like the idea of having big P-cores for my game threads and e-cores for all of the windows threads so cache isn't being thrashed and my game threads aren't being interrupted. I just don't know how that actually plays out.

The lack of AVX512 is kind of ... not sure. It never took off, but now AMD cpus have AVX512, which means the next consoles will probably have AVX512. So whenever the next console gen launches you'll have games that might use it, which means PC versions of games may naturally support it. Will be interesting to see if there's a bit of a drop off for Arrow Lake in three to four years, or whatever it is.

Going to be super interesting. I haven't kept up on when reviews should start hitting. Does anyone know?
There's just as many if not more games that lose performance when hyper-threading is enabled, especially older games. I don't think gaming is going to lose anything with hyper-threading gone. I also think software developers will just target the shared subset between Intel's new AVX10 and AMD's AVX-512 implementation.
 
There's just as many if not more games that lose performance when hyper-threading is enabled, especially older games. I don't think gaming is going to lose anything with hyper-threading gone. I also think software developers will just target the shared subset between Intel's new AVX10 and AMD's AVX-512 implementation.

I guess intel ISPC is supposed to handle differences in simd width? I guess they kind of expect that if you're using their software development libraries/tools you don't really have to worry about those kinds of differences.

I've sort of toyed with disabling smt, but I wonder if an 8-core cpu is enough for that to be ideal. Looking up benches it's very hard to get a sense of whether it's worth it. Varies so much per game. On these new intel cpus with as much as 20 hardware threads, it really doesn't seem like hyper-threading would be useful.
 
Games don’t make use of the E cores unless I’m mistaken so you will not have 20.
That's not at all true. Certainly, some games run better with the e-cores disabled entirely, but an awful lot of modern games that are well multithreaded will absolutely use the e-cores too.

Even on something like an 12th gen CPU with a wee little single 4-E-Core cluster, they're definitely doing their part, as in the attached screenshot in Cyberpunk 2077 with a 12600k.

Given that some of Intel's Raptor Lake-U laptop SKUs are 2x P-Core + 8x E-Core, or even 1x P-Core + 4x E-core, they pretty much have to be in play if the CPU load is such that the P-cores can't keep up.

1728681905075.png
 
Intel scrathed AVX after they realized heir cores had to downclock as much as they did when doing thsoe

Well they scrapped avx512, but not avx2 which is 256bit width registers. A modern core without SIMD would be straight trash-bin worthy. I just wondered if they supported AVX10, because at least they'd be fixing some issues with early AVX standards. 512 bit width seems to be optional for AVX10.

zen5 power and clock don't seem to be affected by avx512, which means a next gen console could include it. Gives me at least a little bit of pause about investing in a high-end cpu that doesn't support it now. Just wouldn't want a generation of games coming out that suddenly don't run optimally. Obviously can't predict the future though.
 
Back
Top