Discussion Intel current and future Lakes & Rapids thread

[moinmoin]

Rather than have 56 GC-Server cores in mesh, would it not be better to have 56 Gracemont-Server quad-core clusters in mesh?
Then, Intel won't be sniped by 192C N2 & 96C V1 as well as 64C/96C/128C Zens.

Considering in many typical workloads ST still matters and the huge amount of cores is not so much for MT but rather for consolidating a lot of ST heavy servers, no, going with a massive amount of GM cores wouldn't help Intel much.

 

[lobz]

Sorry comet lake not tiger lake, my bad

Well since 10nm has almost 2x the density vs 14nm it could be made.

Almost 2 times sounds rather unimpressive in comparison with 2.7 times

 

[Zucker2k]

Intel 12th Gen AMA on Reddit:

https://www.reddit.com/r/intel/comments/qh6xs2

 

[NostaSeronx]

Considering in many typical workloads ST still matters and the huge amount of cores is not so much for MT but rather for consolidating a lot of ST heavy servers, no, going with a massive amount of GM cores wouldn't help Intel much.

Across multiple loads Gracemont's 17 available execution ports are superior to Golden Cove's measly 12.
So, ST/MT heavy repeated/looped workloads would prefer Gracemont-server over Goldencove-server.

Not all workloads in server are Vec-width or ALU-amount junkies. Some workloads prefer 1t to 1c and discrete ports for units.

 

[nicalandia]

Rather than have 56 GC-Server cores in mesh, would it not be better to have 56 Gracemont-Server quad-core clusters in mesh?
Then, Intel won't be sniped by 192C N2 & 96C V1 as well as 64C/96C/128C Zens

Skylake Single Thread like performance can't cut it any more, it's embarrassing how bad they are getting beat by Zen3 EPYC and with newer ever more powerful ARM based Servers. Gracemont/Skylake level of performance is unacceptable anymore

 

[dmens]

One of the most deceptive marketing claims that the Atoms are now equivalent to fairly recent big cores. The big cores spend area to extract parallelism and reduce latency. The atoms scale area back and sacrifice some of that capability. In some workloads, the atoms do just as well as the big cores because the parallelism is easily extracted with fewer resources and their caches are big enough. On other workloads they will fall flat. There is no free lunch. Just go look at existing benchmarks for prior Atoms and you will see exactly where things land with various workloads.

 

[dullard]

Just go look at existing benchmarks for prior Atoms and you will see exactly where things land with various workloads.

I'm having a bit of trouble with this task. What benchmarks should we use, then how much performance do I add for the extra cache, AVX2 instruction capability, larger buffer size, and double execution ports in Gracemont compared to older Atoms?

 

[dmens]

I'm having a bit of trouble with this task. What benchmarks should we use, then how much performance do I add for the extra cache, AVX2 instruction capability, larger buffer size, and double execution ports in Gracemont compared to older Atoms?

LOL, I don't know what workloads you care about. Go look up whatever you want. The differentiation is fairly obvious.

BTW, dispatch port count is a red herring. If you can't feed the CPU backend, having all those ports does not help anything. Notice the ROB/LQ/SQ sizes in Gracemont haven't really changed that much compared to Tremont, and the rename width is still the same. That along with your branch predictor and prefetching is what matters. I don't doubt the last two were improved but those are contextual improvements and highly workload dependent.

 

[inf64]

Don't know if this was posted before:

https://twitter.com/x/status/1453537803011117056

12900K@6.8Ghz GB4 ST Score 11669
11900K@6.819Ghz GB4 ST Score 10807
5950X@6.025Ghz GB4 ST Score 9661

Normalized for clock, IPC for each in GB4 (singlethread) is:
Zen 3 = 1
Cypress Cove = 0.988
Golden Cove= 1.07

edit:

GB 5 scores
12900K@6.8Ghz GB5 ST Score 2740
11900K@6.7Ghz GB5 ST Score 2309
5950X@6.0Ghz GB5 ST Score 2126

Zen 3 = 1
Cypress Cove = 0.972
Golden Cove= 1.137

 

[dullard]

LOL, I don't know what workloads you care about. Go look up whatever you want. The differentiation is fairly obvious.

I care about your workloads. What should I be looking at? How does Gracemont compare in your workloads? You want me to look it up, so I'm here to do so. I want to do exactly what you asked. Just what should I be looking at if it is obvious to you that Gracemont won't be comparable to Skylake?

 

[dmens]

I care about your workloads. What should I be looking at? How does Gracemont compare in your workloads? You want me to look it up, so I'm here to do so. I want to do exactly what you asked. Just what should I be looking at if it is obvious to you that Gracemont won't be comparable to Skylake?

It is slow for compilation and development, which is mainly what I care about. The other thing I care about is video encoding and I haven't even tried running that on an Atom, but I suspect it wouldn't be so stellar at that workload.

 

[Exist50]

I know it's hard to accept for some, but Gracemont really does have Skylake level IPC. There is no "gotcha". One would hope they now see sense putting Atom into a proper manycore Graviton et al. competitor.

 

[Dayman1225]

I know it's hard to accept for some, but Gracemont really does have Skylake level IPC. There is no "gotcha". One would hope they now see sense putting Atom into a proper manycore Graviton et al. competitor.

That happens with Next-Mont. AKA Sierra Forrest

 

[jpiniero]

I know it's hard to accept for some, but Gracemont really does have Skylake level IPC. There is no "gotcha". One would hope they now see sense putting Atom into a proper manycore Graviton et al. competitor.

You might want to wait for benchmarks to say that. I'd like to see how well Gracemont does in gaming. Sadly it seems that you can't run Alder Lake without any P cores but it'd be interesting to see how comparable a 1+8 is to a ~3.8 9700K.

 

[nicalandia]

I know it's hard to accept for some, but Gracemont really does have Skylake level IPC. There is no "gotcha". One would hope they now see sense putting Atom into a proper manycore Graviton et al. competitor.

Skylake level IPC is UNACCEPTABLE When Zen3 EPYC are baby seal clubbing Skylake based Xeons...

 

[jpiniero]

Skylake level IPC is UNACCEPTABLE When Zen3 EPYC are baby seal clubbing Skylake based Xeons...

Not sure why you are comparing Alder Lake to Epyc.

 

[nicalandia]

Not sure why you are comparing Alder Lake to Epyc.

I am not.

I am replying to the guys insisting on having an "All e Core" Xeon type of CPUs... Skylake IPC was acceptable way back on 2016, not anymore

 

[Exist50]

I am not.

I am replying to the guys insisting on having an "All e Core" Xeon type of CPUs

There are plenty of markets that would gladly tolerate Skylake level single core (or worse) for significantly higher multicore performance.

 

[gdansk]

Gracemont is such an unusual design it is hard to guess how it will perform. There is what Intel says but I'll wait for 3rd party benchmarks before being impressed/disappointed.

 

[Zucker2k]

Not to mention Gracemont is only going up, ipc-wise. With node shrinks, Intel engineers will beef it up and pack on more cores. I don't know why this person thinks Intel is done with Gracemont development.

 

[lightmanek]

I know it's hard to accept for some, but Gracemont really does have Skylake level IPC. There is no "gotcha". One would hope they now see sense putting Atom into a proper manycore Graviton et al. competitor.

I would also wait for benchmarks as I do wonder how 2MB shared L2 across 4 cores will affect performance when all cores are loaded.

 

[eek2121]

Not where money matter as in Xeons, you wont find e-cores on 10nm and 7nm Xeons because they need the absolute best performance to compete with ZEN4/ZEN5 EPYC. They can't do that with Skylake type performance Xeons(they are already getting mauled by Zen3)

Well of course, for Single threaded workloads, Gracemont brings around Skylake levels of performance at a fraction of the power consumption of Skylake. Next gen *mont will likely be even faster. For multicore performance, it is a bit trickier. If 4 atoms take the space of one GC core You could squeeze 224 cores in the space of a single 56 core SPR chip. Those 224 cores would likely be much faster than the 56 GC cores, assuming no odd scaling issues and such. I'm actually surprised Intel isn't releasing a special Xeon SKU with 224-256 Gracemont cores.

 

[eek2121]

Intel 12th Gen AMA on Reddit:

https://www.reddit.com/r/intel/comments/qh6xs2

For fun, someone should ask them why NewEgg shipped ADL-S chips out early, and then started harassing customers about it, or why the review embargo date is the same day as launch date.

 

[Hulk]

There are plenty of markets that would gladly tolerate Skylake level single core (or worse) for significantly higher multicore performance.

I'm still rockin' a Haswell for my main rig. Skylake level performance would be an upgrade for me.

 

[Hulk]

Turbo 3.0 question.

Ian wrote that the 12900K will all-core turbo to 5.1GHz. The preferred cores will go to 5.2GHz. If thermal head room is there does that mean 6 cores at 5.1 and 2 at 5.2? Or is the 5.2 only when the app is less heavily threaded and the other cores are running <5.1GHz?