Discussion Intel current and future Lakes & Rapids thread

[DavidC1]

Considering how poor Lion Cove's perf/watt is in Lunarlake, one could imagine how much better their future server chips could be if the much more dynamic and performant E core team takes the spot for the leader architecture.

Granite Rapids is quite decent if you consider that into account.

This is probably their most important product in years. Question is how fast they ramp it up. We will know via earnings call for sure. Either way customer is the winner if Intel is competitive.

From Apple launches we know a new product launch will supress sales for that particular segment for a quarter. And it'll increase it the next quarter.

Full impact of GNR will not be shown at least until Q1 2025 earnings.

Hmm, it seems despite all odds there's some positive light being shone on Intel.

 

[naukkis]

hardly, you have a fetish for something that doesn't matter , generally it costs lots of performance and gives very little to no benefit for the vast majority of software running on high core count servers.

AMD have got free pass to implement their cheap chiplet solution because Intel mesh solution have been total dud. It ain't anymore - look those benchmark results.

 

[naukkis]

Considering how poor Lion Cove's perf/watt is in Lunarlake, one could imagine how much better their future server chips could be if the much more dynamic and performant E core team takes the spot for the leader architecture.

Granite Rapids is quite decent if you consider that into account.

Lion Cove perf/watt is pretty much identical to Hx370 Zen5 at similar 4-cpu stop ring. Big part of that inefficiency might come just from that ringbus. But either way Intel E-core seems to be order of magnitude better from perf/watt view than both LC and Zen5. Intel is fine with their E-core team but AMD should really hurry up to implement something that can actually be competitive against something else than Intel P-cores.

 

[Joe NYC]

Lion Cove perf/watt is pretty much identical to Hx370 Zen5 at similar 4-cpu stop ring. Big part of that inefficiency might come just from that ringbus. But either way Intel E-core seems to be order of magnitude better from perf/watt view than both LC and Zen5. Intel is fine with their E-core team but AMD should really hurry up to implement something that can actually be competitive against something else than Intel P-cores.

Turin Dense will has HT and will be on N3. So, it will be fine against Intel Sierra Forrest.

Against Clearwater Forrest? Intel has a chance to take performance lead there, if it ships on time.

 

[cebri1]

GNR has 8P solutions with 6700P so they can get 86*8 cores ina 8 socket system AMD Doesn't have anything more than 2 socket also the accelerators no one is counting those will destroy everything and free up core resources for other task something no one tests but it is on the silicon

Most costumers do not need 8S

Turin Dense will has HT and will be on N3. So, it will be fine against Intel Sierra Forrest.

Against Clearwater Forrest? Intel has a chance to take performance lead there, if it ships on time.

Skymont is so close to Strix’s Zen5c. It will be very competitive. Intel should reevaluate who is the team in charge of their main arch.

 

[DavidC1]

Lion Cove can barely match Zen 5 while being on N3B. Even if N3B's advancements are small, the density improvements are going to be big, and it's still a better process than the one in Zen 5.

Lion Cove is substantially larger despite this. Zen 5 also introduces new ideas which means it has much more headroom to grow in a more efficient manner.

Skymont is the one that is majorly responsible for greatly improved power efficiency in Lunarlake. In Arrowlake it'll be responsible for having great multi-threaded performance. Lion Cove P core is the one that is dragging the whole thing you could say.

 

[jpiniero]

GNR has 8P solutions with 6700P so they can get 86*8 cores ina 8 socket system AMD Doesn't have anything more than 2 socket also the accelerators no one is counting those will destroy everything and free up core resources for other task something no one tests but it is on the silicon

All of the Granite Models launched are 2S.

 

[Hulk]

I think there is some nuance in understanding Lunar Lake performance. It's easy to simply compare the MT performance in "ridiculously" multithreaded applications/benches like CB and Blender and conclude, "it's not good." Many reviewers have done that and they don't get it. The ARL mobile parts will compete in the segment of the market that requires massive MT capability.

Lunar Lake with 4 really strong cores and 4 ~Raptor cores is aimed at thin and light performance laptops running business applications while still being able to work well with graphics, video, and audio editing.

As we all know, to date, there still aren't a ton of applications outside of rendering and video transcoding that harness all cores. Most applications still rely heavily on 1 or 2 cores with the others chipping in with the "app on top" and/or taking care of background processes.

I'm still looking to buy and had the Asus with the 258v in my cart and almost checked out, but I really don't want to regret not finding one with the 288v processor.

 

[511]

All of the Granite Models launched are 2S.

I said 6700 series the currently launched one is 6900 which is 2S

 

[511]

@gdansk we can continue 786 TOPS int 8 GNR Performance here i have time to waste 🤣

 

[lightisgood]

This is it. The end of EPYC Era.
GNR & SRF have good competitiveness.

AMD EPYC 9005 Turin Turns Transcendent Performance with 768 Threads Per Server

With up to 768 threads in a dual socket server, the AMD EPYC 9005 "Turin" generation offers transcendent performance for servers

www.servethehome.com

 

[Markfw]

This is it. The end of EPYC Era.
GNR & SRF have good competitiveness.

AMD EPYC 9005 Turin Turns Transcendent Performance with 768 Threads Per Server

With up to 768 threads in a dual socket server, the AMD EPYC 9005 "Turin" generation offers transcendent performance for servers

www.servethehome.com

Until benchmarks come out to prove this, I doubt it. Right now EPYC has no competition.

 

[Hitman928]

Until benchmarks come out to prove this, I doubt it. Right now EPYC has no competition.

Phoronix has a large suite of benchmark results. In their tests, it's closer than prior gens for sure, but it's still a pretty lopsided win for 5th gen Epyc.

 

[DavidC1]

Until benchmarks come out to prove this, I doubt it. Right now EPYC has no competition.

He did link a benchmark.

Yup, AMD is quite far ahead. Granite Rapids is showing strange results in few tests though. It's almost dead last in few benchmarks, meaning there's a problem(less likely) or is immature(probably). That'll drag the averages down a fair bit.

Granite Rapids and Sierra Forest is also not scaling well in some tests like the Linux Compilation and Timed Node JS benchmark, where they are beating all Turin or very close to it in 1P but losing significantly in 2P. It doesn't even scale well even compared to Emerald Rapids, where it's showing normal behavior.

In fact, scaling sucks for most of the benchmarks in first page of benchmarks.

 

[Markfw]

He did link a benchmark.

Yup, AMD is quite far ahead. Granite Rapids is showing strange results in few tests though. It's almost dead last in few benchmarks, meaning there's a problem(less likely) or is immature(probably). That'll drag the averages down a fair bit.

Granite Rapids and Sierra Forest is also not scaling well in some tests like the Linux Compilation and Timed Node JS benchmark, where they are beating all Turin or very close to it in 1P but losing significantly in 2P. It doesn't even scale well even compared to Emerald Rapids, where it's showing normal behavior.

In fact, scaling sucks for most of the benchmarks in first page of benchmarks.

What I am saying is that those benchmarks only show the strength of Turin. No where in those does it even imply that "The end of EPYC Era." or
that "GNR & SRF have good competitiveness."

 

[DavidC1]

Oh wow, it's simply not scaling as well. GNR shows 25% improvement Geomean from 1P to 2P while Turin shows 34% for Dense and 40% for regular. Having similar scaling would have reduced Turin's lead from 40% to 20%.

Reduces gain against EMR from 47% to 27% in Phoronix's Geomean.

I have to suspect aside from platform immaturity for GNR, there's something inhibiting scaling for the Birch Stream platform. Could be the immaturity of their tile implementation as well, because there are multiple levels of latency now.

 

[H433x0n]

He did link a benchmark.

Yup, AMD is quite far ahead. Granite Rapids is showing strange results in few tests though. It's almost dead last in few benchmarks, meaning there's a problem(less likely) or is immature(probably). That'll drag the averages down a fair bit.

Granite Rapids and Sierra Forest is also not scaling well in some tests like the Linux Compilation and Timed Node JS benchmark, where they are beating all Turin or very close to it in 1P but losing significantly in 2P. It doesn't even scale well even compared to Emerald Rapids, where it's showing normal behavior.

In fact, scaling sucks for most of the benchmarks in first page of benchmarks.

The 2P scaling is broken on GNR, it's a measly 1.2X. There's also quite a few individual benchmarks like NAMD that flat out don't work and are included in the final geomean calculation. If you compare 1P vs 1P configurations, at the moment Turin has 18% advantage, I'd say that's probably closer to reality than the 2P score that has it at like 40-50% ahead.

Hopefully Phoronix tests this later on once everything is fixed/patched.

 

[DavidC1]

The 2P scaling is broken on GNR, it's a measly 1.2X. There's also quite a few individual benchmarks like NAMD that flat out don't work and are included in the final geomean calculation. If you compare 1P vs 1P configurations, at the moment Turin has 18% advantage, I'd say that's probably closer to reality than the 2P score that has it at like 40-50% ahead.

Hopefully Phoronix tests this later on once everything is fixed/patched.

Aside from the individual benchmarks that run horribly on GNR, Sierra Forest exhibits the same issue, meaning Birth Stream, or their tile implementation, or a third option which is platform immaturity, which based on the differences mean very immature.

 

[H433x0n]

Aside from the individual benchmarks that run horribly on GNR, Sierra Forest exhibits the same issue, meaning Birth Stream, or their tile implementation, or a third option which is platform immaturity, which based on the differences mean very immature.

For a few of these tests (NAMD in particular), they do produce expected results if compiled with oneAPI. I’m not sure how to interpret that info. I don’t know if that points to an issue with the tile implementation or Birch Stream.

 

[AcrosTinus]

After reading Serverthehome and Phoronix, I just can't understand how AMD got this done with less resources and personnel.

Might jump on the HEDT version...

 

[DavidC1]

For a few of these tests (NAMD in particular), they do produce expected results if compiled with oneAPI. I’m not sure how to interpret that info. I don’t know if that points to an issue with the tile implementation or Birch Stream.

Sierra Forest also scales worse in the Linux compile, and the JS benchmark. In fact most of the first page.

After reading Serverthehome and Phoronix, I just can't understand how AMD got this done with less resources and personnel.

More personnel means more layers are needed, meaning more delays, more bureaucracy, etc, etc. There's a propagation delay for any message meaning time consuming meetings, just like huge caches are slow. Also if you have too much money, you tend to waste more.

If you work in a mom and pop shop with 5 employees, you can basically solve all communication problems in 5 mins or less. Try doing that with 100,000.

 

[jpiniero]

After reading Serverthehome and Phoronix, I just can't understand how AMD got this done with less resources and personnel.

TSMC

 

[511]

After reading Serverthehome and Phoronix, I just can't understand how AMD got this done with less resources and personnel.

Cause AMDs P core Arch in Zen 5 is superior than P core arch in GNR one reason

 

[DavidC1]

@H433x0n It seems while GNR has more bugs, definitely with NAMD and others where 2P performs at half the rate while other platforms improve by 15%, Sierra Forest in code compilation isn't scaling as well, so even if they remove the obvious ones I just mentioned, I have to think there's a more fundamental issue since Sierra has been out for quite some time now.

 

[511]

Yes even scrolling through level1 tech some are bugs they are slower than EMR 🤣

AMD EPYC Turin Is OUT! Zen5 9000 Benchmarking w/ COMSOL & Phoronix Test Suite

Official Press Deck Slides COMSOL FGMRES Compute Pardiso Compute CFD Only 10GB EM Only 260GB Phoronix Test Suite AMD EPYC Turin Benchmarks.rar (896.4 KB) Test Systems Summary Timed Compilations...

forum.level1techs.com