News 13th and 14th Gen Intel CPU instability also hits servers — W680 boards with Core i9 K-series chips are crashing

[jeremyj_83]

And what is the goal of that?

See if Raptor Cove is really anything other than a rebranded Golden Cove (which it really isn't).

 

[TheHerald]

See if Raptor Cove is really anything other than a rebranded Golden Cove (which it really isn't).

Okay, let's then just run them both at 5.5 ghz and compare. Oh wait...

 

[jeremyj_83]

Okay, let's then just run them both at 5.5 ghz and compare. Oh wait...

Well whenever they have been at the same frequency you see performance is equal except where the Raptor based CPUs are able to use so many e-cores. That is what happens when IPC is the same. Just because one can or cannot get to X frequency is meaningless.

 

[DeathRabit]

The data from Level1Techs YT channel ALSO shows error logs at Oodle game telemetry data. Intel's 13th and 14th Gen CPUs represent a major portion of the error logs.

Intel accounted for 1,431 decompression errors (out of 1584 over 90 days), while AMD, only had four such errors, which is significantly lower than Intel.

Breakdown shows that more than 70% of Intel's CPUs were prone to errors compared to 30% of AMD.

This data show distribution between brands in crash reports, not that AMD is 2x les prone to errors.

Also according data from server farm to host multiplayer games that compared between i9 13900/14900 vs Ryzen 9 7950x
Also intel based system failure rate reach near 100% over time.
AMD system generate 100x fewer server crashes
Source: https://alderongames.com/intel-crashes

 

[TheHerald]

Well whenever they have been at the same frequency you see performance is equal except where the Raptor based CPUs are able to use so many e-cores. That is what happens when IPC is the same. Just because one can or cannot get to X frequency is meaningless.

The 14900k has a 16 % higher clockspeed on the Pcores alone. That's as much as the rumored uplift is on the new zen 5 cpus. Add the larger cache, decoupled cache clockspeed and the extra amount and clockspeeds on the ecores, it's delusional to say they are in any way similar. But you do you

 

[thestryker]

They called it Raptor Cove lets be honest here it is just Golden Cove with a few minor uArc changes.

You calling it Golden Cove is dishonest and false period.

They called it Raptor Cove lets be honest here it is just Golden Cove with a few minor uArc changes. If it were really that different you would see IPC increases beyond 2% or so. I remember seeing benchmarks of the 3 generations with them all set to the same clock speed and in ST applications they all had identical performance. The only reason for MT performance increase was due to more e-cores on the 13th and 14th gen. Therefore you can say that Raptor Cove doesn't even equal a Tick as there really wasn't any IPC increase.

They also had the ring bus set to 3ghz for no apparent reason so I wouldn't put any stock into that analysis. Nobody has done, to my knowledge, a proper test comparing IPC on GC/RC but all estimates are around 4% which mirrors a lot of what we saw gen on gen in the last decade.

The major changes on RC were to allow clockspeeds to go higher and have more efficient perf/watt. Several of the advantages also improve performance with E-cores active (like higher ringbus clocks) which is why the 13700K is faster than the 12900KS.

 

[bit_user]

The highest clocked EMR SKU has a 4.2ghz boost on an 8 core part so the voltage required for that, even if all 8 cores are running 4.2, is relatively low.

I think you're probably right, but if the server version of the cores isn't designed to clock as high as client, then it could still be running this with higher voltage than it would take a client Raptor Cove to hit the same speed.

 

[bit_user]

They called it Raptor Cove lets be honest here it is just Golden Cove with a few minor uArc changes.

As far as I've read, the changes only went as deep into the core as enlarging the L2. I don't really consider that part of the micro-architecture, but maybe some would.

If it were really that different you would see IPC increases beyond 2% or so. I remember seeing benchmarks of the 3 generations with them all set to the same clock speed and in ST applications they all had identical performance.

Depends on the workload. There should be a few things that would benefit from the larger L2. According to this, Golden Cove's L2 bandwidth is about double that of L3 and latency is about 1/5th. So, if the larger L2 meaningfully increases the cache hit rate of your workload, then you should see a measurable IPC uptick, though probably still small.

Source: https://chipsandcheese.com/2021/12/02/popping-the-hood-on-golden-cove/
​

The only reason for MT performance increase was due to more e-cores on the 13th and 14th gen.

That's not true. Did you miss where I quoted the slide citing their improved manufacturing process yielding a better VF curve? Also, they tweaked with the ring bus, from what I've heard.

 

[thestryker]

I think you're probably right, but if the server version of the cores isn't designed to clock as high as client, then it could still be running this with higher voltage than it would take a client Raptor Cove to hit the same speed.

There are two 8 core parts that boost fairly high one 4.2 and one 4.1 and the difference in TDP is 195W vs 165W, but the interesting part is that for that 30W increase the base clock goes from 3.2 to 3.9. That is a pretty dramatic clockspeed increase considering it's less than 4W/core which indicates to me that they're not using a lot of voltage even on these lower core count parts. When you consider that plus the 60 additional PCIe lanes and 8 channel memory controller it seems unlikely the voltages would be higher.

 

[jeremyj_83]

The 14900k has a 16 % higher clockspeed on the Pcores alone. That's as much as the rumored uplift is on the new zen 5 cpus. Add the larger cache, decoupled cache clockspeed and the extra amount and clockspeeds on the ecores, it's delusional to say they are in any way similar. But you do you

Raptor Cove is 99% Golden Cove. Therefore they are VERY similar. But you do you.

 

[bit_user]

Raptor Cove is 99% Golden Cove. Therefore they are VERY similar. But you do you.

I don't know why this had to turn into some sort of pitched battle. I look at Raptor Lake as a proper refresh of Alder Lake. Yes, the core microarchitectures are the same (as far as I'm concerned), but there are enough other changes that make Raptor Lake much more than just an overclocked Alder Lake.

 

[jeremyj_83]

I don't know why this had to turn into some sort of pitched battle. I look at Raptor Lake as a proper refresh of Alder Lake. Yes, the core microarchitectures are the same (as far as I'm concerned), but there are enough other changes that make Raptor Lake much more than just an overclocked Alder Lake.

IMO Raptor Cove isn't even a Zen+ level refresh as IPC didn't increase. It is a refresh though and Intel was able to increase clocks and have better boost profiles to stay at higher clocks.

 

[bit_user]

IMO Raptor Cove isn't even a Zen+ level refresh as IPC didn't increase.

Zen+ didn't change the microarchitecture, either! However, unlike Raptor Lake, it also didn't change the macro-architecture. All they did was port from 14 nm -> 12 nm (no layout changes required) and tweak the cache management algorithm (i.e. not even changing the size)!

Comparing the two, I think Zen+ probably gained more from the node change, but otherwise had only microcode-level changes. Overall, I'd say Raptor Lake was probably the bigger change and certainly yielded the bigger improvement.

P.S. Why do you insist on dragging this out? I'd just let it go.

 

[TheHerald]

~35-40% increase in MT performance within a year at same power but yeah, RPL was no biggie. Probably the biggest increase we've ever seen within a year the last 10 to 15 years.

 

[bit_user]

~35-40% increase in MT performance within a year at same power but yeah, RPL was no biggie. Probably the biggest increase we've ever seen within a year the last 10 to 15 years.

No, the Rocket Lake -> Alder Lake improvement was bigger.

 

[TheHerald]

No, the Rocket Lake -> Alder Lake improvement was bigger.

True but rocket was a downgrade so whatever, let's hide that one under the rug.

 

[bit_user]

True but rocket was a downgrade so whatever, let's hide that one under the rug.

It boosted single-threaded performance, but there were a few regressions on multi-threaded workloads (i.e. due to having only 8 cores instead of Comet Lake's 10).

Overall, it was not considered a regression, except on the efficiency front. Here's what Anandtech concluded (since I'm having trouble finding Tom's review):

"In our testing, we saw the following:

Single thread floating point: +22%
Multi-thread floating point: +16.2%
​

Sounds great, right?

Single thread integer: +18.5%
Multi-thread integer: +5.8%
​

Oh. While Intel’s claim of +19% is technically correct, it only seems to apply to math-heavy workloads or single thread integer workloads. The benefits of non math-based throughput are still better than average, but only at 5.8% for multithreaded. Very rarely do Intel’s big claims come with an easily identifiable asterisk.

When we look at our real-world data, in almost every benchmark the 11700K either matches or beats the 10700K, and showcases the IPC gain in tests like Dolphin, Blender, POV-Ray, Agisoft, Handbrake, web tests, and obviously SPECfp."
​

Note that I quoted their i7-11700K review, since they got their hands on an i7 almost a month before the official review embargo. So, that's when they did their main performance analysis. Their i9-11900K review basically reiterates those points.

 

[TheHerald]

It boosted single-threaded performance, but there were a few regressions on multi-threaded workloads (i.e. due to having only 8 cores instead of Comet Lake's 10).

Overall, it was not considered a regression, except on the efficiency front. Here's what Anandtech concluded (since I'm having trouble finding Tom's review):

​

"In our testing, we saw the following:​

​

Single thread floating point: +22%​

Multi-thread floating point: +16.2%​

​

​

Sounds great, right?​

​

Single thread integer: +18.5%​

Multi-thread integer: +5.8%​

​

​

Oh. While Intel’s claim of +19% is technically correct, it only seems to apply to math-heavy workloads or single thread integer workloads. The benefits of non math-based throughput are still better than average, but only at 5.8% for multithreaded. Very rarely do Intel’s big claims come with an easily identifiable asterisk.​

​

When we look at our real-world data, in almost every benchmark the 11700K either matches or beats the 10700K, and showcases the IPC gain in tests like Dolphin, Blender, POV-Ray, Agisoft, Handbrake, web tests, and obviously SPECfp."​

​

Note that I quoted their i7-11700K review, since they got their hands on an i7 almost a month before the official review embargo. So, that's when they did their main performance analysis. Their i9-11900K review basically reiterates those points.

The 11700k was kinda decent when compared t the 10700k (since same core count) but the 11900k sucked. Also they were insanely power hungry. Yes unlike alderlake and raptorlake, they weren't just pushed to high power draws, they were not performing at lower ones. Also the IMC was hitting a wall at around 3400-3600 for most chips and 3800 if you were lucky on an 11900k. Cometlake was casually hitting 4400.

 

[Lorien Silmaril]

Intel fanboys spinning like a top is the quality content I come here for. 🙃

 

[ingtar33]

First of all, 13th gen has extra cache, lower latency on the cache, decoupled cache frequency from ecores, and also higher clocks at the same voltage. Why would you run all 3 chips at the same speed? That's the whole point, that you CANT run the 12900k at 5.5ghz that the 13900k can achieve.

to see the IPC change generation over generation, IPC means instructions per clock, clock being "clock speed" so in order to compair ipc, you need to clock the chips to the same ghz, else you won't be measuring ipc... the reason there is confusion over this fact is due to intel's marketing department confusing ipc for improved performance, and often quote gains in performance as an ipc increase, for example say you can get 100fps in game 1 with 12th gen, and 115fps with a 13th gen, intel marketting will cite that as a 15% ipc gain. but the 13th gen chips will be running at a higher clock speed then 12th gen. so it's not really ipc, its a performance gain. despite the minor changes from 12 to 13th gen, the ipc gains were basically 0. same for from 13th to 14th gen. the gains in intel performance from 12th to 14th gen are entirely due to higher clock speeds.

so while there were some minor changes from 12th to 13th gen, there was no functional difference in the chips.

 

[TheHerald]

to see the IPC change generation over generation, IPC means instructions per clock, clock being "clock speed" so in order to compair ipc, you need to clock the chips to the same ghz, else you won't be measuring ipc... the reason there is confusion over this fact is due to intel's marketing department confusing ipc for improved performance, and often quote gains in performance as an ipc increase, for example say you can get 100fps in game 1 with 12th gen, and 115fps with a 13th gen, intel marketting will cite that as a 15% ipc gain. but the 13th gen chips will be running at a higher clock speed then 12th gen. so it's not really ipc, its a performance gain. despite the minor changes from 12 to 13th gen, the ipc gains were basically 0. same for from 13th to 14th gen. the gains in intel performance from 12th to 14th gen are entirely due to higher clock speeds.

so while there were some minor changes from 12th to 13th gen, there was no functional difference in the chips.

Why do I care about the IPC? I care about the actual performance, IPC is useless.

 

[dalek1234]

It's an architectural issue. Intel can't fix it, and that' s why they aren't saying what the problem is.

 

[bit_user]

It's an architectural issue. Intel can't fix it, and that' s why they aren't saying what the problem is.

Except it doesn't seem to affect all CPU models that are based on the same die. That's why I think the reason they're not saying is that it would cost too much performance.

 

[Peksha]

Technically, we have one GLC core for three client generations (ADL/RPL/RPL-R) and two server generations (SPR/EDR). But these cores are configured differently: ADL uses GLC 1.25MB L2, SPR/RPL/RPL-R/EDR uses 2MB L2. There are also some changes to the L2 predictor/branch predictor. We also know that RPL/RPL-R has DLVR turned off. This could be a problem! Is it enabled for SPR/EDR? We know that power management on ADL and RPL/RPL-R works differently, even to the point that ADL supported UV, but RPL/RPL-R did not! Remember this.

The same with the GMT core - it was originally designed for two L2 variants - 2(ADL)/4MB (RPL/RPL-R/SF). We also know that RPL/RPL-R support separate GLC/GMT clocking, but overall strange power management of cores and L2/3 cache (I didn’t have ADL, and now I don’t have RPL - switched to AMD after all this mess)

 

[bit_user]

Technically, we have one GLC core for three client generations (ADL/RPL/RPL-R) and two server generations (SPR/EDR). But these cores are configured differently: ADL uses GLC 1.25MB L2, SPR/RPL/RPL-R/EDR uses 2MB L2. There are also some changes to the L2 predictor/branch predictor.

The server version of Golden Cove is actually a fair bit different from client. Yes, both it and Raptor Cove have 2 MB L2 caches, but the similarities probably end there.

​

Sadly, I can't find a similar image of Raptor Cove, so we can confirm. However, I'm pretty sure Intel wouldn't burn so much silicon by adding (disabled) AMX and a second AVX-512 FMA to all the Raptor Lake P-cores.