Discussion Intel current and future Lakes & Rapids thread

Page 736 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Hitman928

Diamond Member
Apr 15, 2012
4,375
5,766
136
Do any of you know what's going in here? Even with the OpenVIMO AMX boost(They are real, you can click on the link and see them) the 8490H is getting beat pretty soundly by two 9374F Genoa CPUs(64C total, so it's a 64C vs 60C)

View attachment 74420

AMD Genoa 2S 9374F vs 1P Intel Xeon SPR 8490H
Their Epyc system spans the 64 cores over two CPUs, so the all core boost clocks should be much higher for the Epyc CPUs. Power consumption should be higher for the Epyc system as a result as well.
 

Exist50

Golden Member
Aug 18, 2016
1,694
1,778
136
That was more Xeon E. You can still buy those but who knows if Intel will refresh it with anything beyond Rocket Lake.

One chiplet Epyc probally makes more sense.
By "file server", I really meant a storage rack. Basically any use case where you just need the IO bandwidth, but the CPU is more or less irrelevant.
 

nicalandia

Platinum Member
Jan 10, 2019
2,955
4,557
136
I don’t know enough about this software or how the accelerator works to really say. It does look strange though.
That's the Only App in there that is taking advantage of Intel AMX Intructions... It's weird that 1P is doing so poorly. If it's not a Bug and The AMX scales exponentially with Sockets...Well the performance of the 4P parts will be mind-blowing.
 

Hulk

Diamond Member
Oct 9, 1999
3,680
1,275
136
Looks like many of them are tightly coupled to memory, so probably not possible. Maybe the encryption stuff though?

Training, largely true, but inference is surprisingly dominated by CPUs. The GPU advantage diminishes until tight latency constraints or small batch sizes, and CPUs are obviously more flexible if you're not going to be doing inference all the time.

For training, GPUs are great if your model fits in VRAM (or is amenable to being streamed in), but for really, really large models, they sometimes have to fall back to CPUs.

Jim Keller remarked on this in talk once. I skipped to the relevant part, but the whole thing is worth a watch. Timestamp at 42:47, if the media embed messes it up.


TLDW: He estimated at the time (about 3 years ago) that AI was something like 80% CPU, 20% GPU, 0% other, and says that if things moved quickly, it would be something like 1/3 each in 5 years, but things probably wouldn't move that quickly.
This is a quite amazing video for people like us that are into this stuff. If you haven't watched it I suggest you do.
Thanks for sharing.
 

DrMrLordX

Lifer
Apr 27, 2000
20,487
9,563
136
Impressive outshipping the competition with ease
That's a lot of people getting stuck with Cascade Lake-SP because nothing else was available. And Cascade Lake-SP was selling at a significant discount. From a revenue and technology point-of-view, that's not really great for Intel.

A slight clock penalty isn't meaningful if it only happens when you're running significantly more ops/cycle.
It is if you're hosting VMs on that system and someone starts up an AVX512 workload on their VM, slowing down the entire host CPU. At least that was a problem with Skylake/Cascade Lake generation CPUs.
 
  • Like
Reactions: ftt and moinmoin

Exist50

Golden Member
Aug 18, 2016
1,694
1,778
136
It is if you're hosting VMs on that system and someone starts up an AVX512 workload on their VM, slowing down the entire host CPU. At least that was a problem with Skylake/Cascade Lake generation CPUs.
The clocks are controlled on a per-core granularity. So only the core running heavy AVX512 or AMX code will have a frequency penalty.
 
  • Like
Reactions: Carfax83

IntelUser2000

Elite Member
Oct 14, 2003
8,578
3,634
136
Usage of Intel's AVX-512 was known to reduce operating frequency due to higher power consumption, which doesn't happen with AMD's implementation of AVX-512 where AVX-512 is actually using slightly less power than AVX2.
Icelake-SPs AVX-512 is far better than on Skylake-SP, for obvious reasons like being on the 10nm process.

The big reason AMD's implementation results in less downgrade is because it uses 2 cycles to execute an AVX-512 instruction. Vector instructions take more than 60% of the power use of a core, it makes sense if you increase that, then you need to sacrifice. But if you are getting 2x as much per clock, lowering clock by few % still pays off.
 
  • Like
Reactions: Carfax83

Geddagod

Member
Dec 28, 2021
171
152
76
But.. but.. MLID saying Lion Cove was the ground up design?!?!?!
Btw Harukaze quotes that tweet and said that there prob won't be much new, exciting info on the 26th about Lunar Lake, so idk about getting my hopes up too much for any exciting info.
Kind of like how we were supposed to get "more info" about MTL at hotchips, but we didn't get much really at all...
 

Exist50

Golden Member
Aug 18, 2016
1,694
1,778
136
But.. but.. MLID saying Lion Cove was the ground up design?!?!?!
Btw Harukaze quotes that tweet and said that there prob won't be much new, exciting info on the 26th about Lunar Lake, so idk about getting my hopes up too much for any exciting info.
Kind of like how we were supposed to get "more info" about MTL at hotchips, but we didn't get much really at all...
Lunar Lake uses Lion Cove. But it's still not Royal.

But agreed. They're not going to seriously discuss LNL till after MTL launches.
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,578
3,634
136
The PC Roadmap seminar that went live today talked about nothing significant. I was hoping for a mobile roadmap leak.

Why do they do this? Just don't talk about it at all. It's like a kid crying and screaming "Please PLEASE PAY ATTENTION TO ME!!"

The real juicy stuff is always hidden behind a veil. Who'd have known without going to Twitter or Anandtech forums?

MJ will probably talk about one more thing in addition to what we already know about Lunar Lake(almost nothing), and we'd have to wait few weeks or even couple of months by a leaked slide to get something real. Since Lunar Lake is few generations away, that leaked slide could be couple of years away.
 

DrMrLordX

Lifer
Apr 27, 2000
20,487
9,563
136
The PC Roadmap seminar that went live today talked about nothing significant. I was hoping for a mobile roadmap leak.

Why do they do this? Just don't talk about it at all. It's like a kid crying and screaming "Please PLEASE PAY ATTENTION TO ME!!"
Odd. It seemed urgent when first announced.
 

moinmoin

Diamond Member
Jun 1, 2017
4,176
6,253
136

You must add personal commentary when dropping links/tweets.

Daveybrat
AT Moderator


Edit: Further details about MTL/ARL/LNL from Intel Client head Michelle Johnston Holthaus.
"important features (...) like silencing dog barking or lawnmowers"

I'm not sure what to think about this being the only use case mentioned for the upgrades of their chips...
 
  • Haha
Reactions: lightmanek

Saylick

Platinum Member
Sep 10, 2012
2,282
4,177
136
MTL will be released in 2023, while Lunar Lake will be production ready in 2024 on 18A. Intel's roadmap indicates that 18A is ready for HVM in 2H 2024, so we'll likely see some form of Lunar Lake SKU come out during the 2024 holiday season so that Intel can say they met their timeline goals, even if it's in limited quantities.
 

ASK THE COMMUNITY