Discussion Intel current and future Lakes & Rapids thread

[cebri1]

More benchmarks https://www.phoronix.com/review/intel-xeon-6980p-performance

 

[igor_kavinski]

Short lived victory until Turin eats their lunch. Good luck, Intel

Better ship out those servers ASAP!

 

[gdansk]

It's good news for Intel. Should help their DC margin.
But I'd like to see a comparison of the 6952P to the 9654.

 

[naukkis]

OK, Intel did get competitive Xeon out. If that has performant unified 500MB L3 it also will pretty much stomp everything AMD has to offer with their chiplet implementation - AMD pretty much also has to change their direction towards more advanced packaging.

 

[igor_kavinski]

Full line-up: https://ark.intel.com/content/www/us/en/ark/products/series/240357/intel-xeon-6.html

Hotchips presentation: https://www.intel.com/content/www/u...te-rapids-xeon-architecture-presentation.html

In case anyone's interested in those.

 

[trivik12]

Which explains why Intel barely mentioned anything about the successor to Granite Rapids, since it's 2 years away according to that info.

You sure about the client names? You'd think Panther is Panther and Nova is Cougar.

Raichu had posted about Panther Lake using Cougar Cove almost 2 years ago. Anyway we have to wait until close to release to see what these Coves are. Tick or Tock.

 

[Khato]

Short lived victory until Turin eats their lunch. Good luck, Intel

Better ship out those servers ASAP!

Or could be a long-term victory if Turin ends up with as little gain as Zen5 has shown us thus far.

Have to remember that with EMR->GNR Intel has gone from a deficit to at least parity if not advantage on core count, process technology, and memory bandwidth. Turin will primary get AMD back to a core count advantage temporarily, along with a slight node advantage probably.

Intel's approach is certainly interesting thus far - GNR for best performance and mediocre efficiency, SRF for acceptable performance and best efficiency. Certainly looking forward to CWF bringing the best of both.

 

[gdansk]

Turin ends up with as little gain as Zen5 has shown us thus far.

Was Zen 5 "little gain" in these workloads? I seem to recall it did OK in Phoronix. Moreover Turin follows Intel and increases TDP and core count; unlike Granite Ridge.

But yes, overall it's a return to competition which should mean pricing is key.

 

[controlflow]

Short lived victory until Turin eats their lunch. Good luck, Intel

Better ship out those servers ASAP!

The days of AMD destroying Intel in DC are likely over.

Turin vs GNR will be much more closely matched than the one sided rounds vs SPR and EMR.

Intel is no longer competing with a massive core count disadvantage and process node disadvantage. They have also fixed the deficiencies in memory/cache that previous generations had while also retaining things like AMX accelerators and the SW integration for that. This looks like a pretty solid offering.

 

[Markfw]

It's good news for Intel. Should help their DC margin.
But I'd like to see a comparison of the 6952P to the 9754.

Also, its interesting that Phoronix generally shows the actual wattage used, but not this time. The 2 things Zen 5 has shown, is that they have the full avx-512 implementation, and lower wattage usage then Zen 4. If Turin has the same, it may kill this CPU totally. We will have to wait until ~Oct 10th to see.

 

[Nothingness]

Was Zen 5 "little gain" in these workloads? I seem to recall it did OK in Phoronix. Moreover Turin follows Intel and increases TDP and core count; unlike Granite Ridge.

Yes, Phoronix had +18% over a set of benchmarks for 9950x vs 7950x. But one would have to look at individual benchmarks to make a comparison with their 6980P benchmarks.

Better wait for Turin release before drawing any premature conclusion.

Ref: https://www.phoronix.com/review/amd-ryzen-9950x-9900x/15

 

[gdansk]

Yes, Phoronix had +11% over a set of benchmarks for 9950x vs 7950x. But one would have to look at individual benchmarks to make a comparison with their 6980P benchmarks.

Spoiler

11%? I don't want to derail the Intel thread any more but

When taking the geometric mean of all the raw performance results, the Ryzen 9 9950X came out to being 17.8% faster than the Ryzen 9 7950X

with a lower PPT of 200W instead of 232W. I hazard against disregarding Turin simply because gamers are disappointed with Granite Ridge.

 

[Nothingness]

@gdansk Sorry for the typo, you are right.

 

[lightisgood]

Welcome back Intel, the king of HPC.

 

[Joe NYC]

Welcome back Intel, the king of HPC.

For 3 weeks?

Turin launch is in first half of October. IIRC, on 10th.

 

[fastandfurious6]

The days of AMD destroying Intel in DC are likely over.

Turin vs GNR will be much more closely matched than the one sided rounds vs SPR and EMR.

The days of Intel destroying AMD in DC are definitely over

GNR vs Turin will be 100% one-sided in Turin's favor

 

[cebri1]

No independent benchmarks yet from Wendall but good recap. By the way he mentions that Sierra Forest is selling very well. Intel is having issues meeting demand.

Edit: the intel’s threads are great honey spots for *****s. Love the ignore button.

 

[lightisgood]

Turin launch is in first half of October. IIRC, on 10th.

Turin ? WHAT IS IT? lol.

 

[yuri69]

OK, Intel did get competitive Xeon out. If that has performant unified 500MB L3 it also will pretty much stomp everything AMD has to offer with their chiplet implementation - AMD pretty much also has to change their direction towards more advanced packaging.

AMD always aims for the most cheapo solution. Intel has been way more adventurous and it seems it starts to pay off.

 

[511]

Margins on these will be better and it will definitely sell better also they said Amazon Google are receiving their custom versions

 

[DavidC1]

Turin will primary get AMD back to a core count advantage temporarily, along with a slight node advantage probably.

There's no core count advantage for Turin. It's 128 core vs 128 core for the P segment. It's the Turin Dense that gets 192 cores.

Short lived victory until Turin eats their lunch. Good luck, Intel

Better ship out those servers ASAP!

It should be quite close. The gains are greater than 2x over EMR in many workloads.

Genoa had a 50% core count advantage. That goes to zero this generation.

 

[trivik12]

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.

Gaudi 3 is meh considering its still using HBM2e. There are rumors they are getting out of training game and that would be a mistake and they need to iterate faster. Bionic was saying they will have yearly cadence on DC Training product from next year onwards. Only time will tell if there are any change in plans.

Granite needs to hold for long until Diamond Rapids release in 2026. I am not sure how many cores they can add to one chip at this point. There would be TDP limitations for these chips. Rumor is next gen will hit 144 cores.

 

[itsmydamnation]

OK, Intel did get competitive Xeon out. If that has performant unified 500MB L3 it also will pretty much stomp everything AMD has to offer with their chiplet implementation - AMD pretty much also has to change their direction towards more advanced packaging.

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.

 

[igor_kavinski]

AMD always aims for the most cheapo solution. Intel has been way more adventurous and it seems it starts to pay off.

I think the word you are looking for is cost effective. Intel's adventurous undertakings come at the cost of yields and higher fabrication expense, driving down their profits.

 

[511]

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