Discussion Intel current and future Lakes & Rapids thread

[511]

How does it make sense that they won't be competitive in performance on the low end, but will be on the higher end?

View attachment 134748
Only explanation I have is that per-core perf might be much stronger with Zen 6 than DMR, and in lower end skus where the boosts are also typically higher, you can really push that advantage, unlike in higher core count skus where you get limited by power-per-core much faster.

Most likely the per core performance on low core count SKU cause for dense SKU no one will be targeting 5Ghz clockspeed but for Low End AMD will be shipping 5Ghz SKUs easily and than SMT will be more useful at lower core count.

A little bit worryingly though, those lower end skus are also typically where the AI headnode skus are coming from. You don't need 192 or 256 cores for an AI headnode.

Oh Nvidia SKU is separate from these ones and than there are custom Xeons as well.

 

[511]

As far as Vera Rubin based boxes, don't know if it was announced which x86 CPU Intel will use, but most likely GNR. It would be too risky to bet on DMR.

There is a DMR SKU with Rubin For those who want x86 on Nvidia platform don't forget Nvidia/Intel deal was work in a while (since pat was the CEO).

 

[regen1]

How does it make sense that they won't be competitive in performance on the low end, but will be on the higher end?
View attachment 134748
Only explanation I have is that per-core perf might be much stronger with Zen 6 than DMR, and in lower end skus where the boosts are also typically higher, you can really push that advantage, unlike in higher core count skus where you get limited by power-per-core much faster.

A little bit worryingly though, those lower end skus are also typically where the AI headnode skus are coming from. You don't need 192 or 256 cores for an AI headnode.

Don't know how much and what to make of some of Intel's PR responses(spins?) this year. If they are not that competitive with more cores, more memory B/W, 2 Gen(RWC->LNC->PNC-X) worth of core upgrades, better optimized fabric, node gains then what GNR-SP is going to do in front of the competition ?
Then there's double whammy of foundry under-utlisaztion by canceling DMR-SP. They are already hinting that foundry might not breakeven in 2027 although citing 14A.
They must be really hoping to sell a lot of custom Xeons for Nvidia and hope for a general uptrend in Server CPU sales.

BTW, as far as GNR, I don't think NVidia is using it currently in their reference designs. But 3rd parties can.

GNR is in DGX B300(uses Xeon 6776P(GNR-SP, 8-Channel))

NVIDIA DGX B300 Datasheet

resources.nvidia.com

As for custom Xeons with NVLink for Nvidia they likely won't do a completely custom design(at least initially), just something like switching internal I/O tile for NVLink one.

 

[511]

As for custom Xeons with NVLink for Nvidia they likely won't do a completely custom design(at least initially), just something like switching internal I/O tile for NVLink one.

well that's what you would do tbh for Custom Xeon the IO Die would have NVLink that's it you don't need tape out different CPU Tile to support it.

Don't know how much and what to make of some of Intel's PR responses(spins?) this year. If they are not that competitive with more cores, more memory B/W, 2 Gen(RWC->LNC->PNC-X) worth of core upgrades, better optimized fabric, node gains then what GNR-SP is going to do in front of the competition ?
Then there's double whammy of foundry under-utlisaztion by canceling DMR-SP. They are already hinting that foundry might not breakeven in 2027 although citing 14A.

i think 2X general performance improvement is not a tall order with 1.5 Node worth of performance per watt improvement and 2 uArch along with everything you mentioned.

I am curios about the 384C DMR-HD

 

[regen1]

i think 2X general performance improvement is not a tall order with 1.5 Node worth of performance per watt improvement and 2 uArch along with everything you mentioned.

Let's see how far they have come up with fabric, scaling and other factors.
From GNR-AP to DMR-AP it should be a big upgrade. Wrt volume availability there's a decent gap vs the Venice launch and they have abandoned DMR-SP.
Another worrying thing(though too far in future) for Intel is that Coral Rapids might not be that distant from Zen 7 server launch.

I am curios about the 384C DMR-HD

I hope it doesn't get cancelled with some stupid excuse of it not having SMT or not enough margin.

 

[Geddagod]

Another worrying thing(though too far in future) for Intel is that Coral Rapids might not be that distant from Zen 7 server launch.

I personally expect Coral Rapids to, at best, come out half a year earlier than Zen 7 epyc. Based on the ~2 years launch cycle AMD seems to have.

 

[511]

I hope it doesn't get cancelled with some stupid excuse of it not having SMT or not enough margin.

It would be beast of a CPU but LBT is truly an idiot if he kills something just cause not enough margins to meet 50% also with 384 Physical cores would you need SMT

 

[OneEng2]

It would be beast of a CPU but LBT is truly an idiot if he kills something just cause not enough margins to meet 50% also with 384 Physical cores would you need SMT

If Intel can achieve the same MT SMT effectiveness as AMD, it's just about how efficient those transistors are. 5% more core space gets you 40% performance improvement.

I can't think of any better uses of those transistors than that.

 

[Thunder 57]

If Intel can achieve the same MT SMT effectiveness as AMD, it's just about how efficient those transistors are. 5% more core space gets you 40% performance improvement.

I can't think of any better uses of those transistors than that.

You know it improves by a variable amount depending on code. 40% would be seem to be an extreme outlier.

 

[511]

If Intel can achieve the same MT SMT effectiveness as AMD, it's just about how efficient those transistors are. 5% more core space gets you 40% performance improvement.

I can't think of any better uses of those transistors than that.

Also you need to rethink HT introduces Side Channel issues as well some Server Config explicitly disables SMT the biggest advantage is in Virtualization for VMs but at 384C you have to think does it makes sense to have it or not it is useful at lower config

 

[Geddagod]

You know it improves by a variable amount depending on code. 40% would be seem to be an extreme outlier.

RPC got 18% in specint at iso power

 

[511]

RPC got 18% in specint at iso power

RPC?

 

[Joe NYC]

GNR is in DGX B300(uses Xeon 6776P(GNR-SP, 8-Channel))

NVIDIA DGX B300 Datasheet

resources.nvidia.com

Thanks, I missed that. When I was looking for Blackwell DGX somehow only the B200 came up...

 

[Joe NYC]

I personally expect Coral Rapids to, at best, come out half a year earlier than Zen 7 epyc. Based on the ~2 years launch cycle AMD seems to have.

If Zen 7 is to be A14 and Coral Rapids 14A, then a lot depend on which one will be out first.

 

[OneEng2]

You know it improves by a variable amount depending on code. 40% would be seem to be an extreme outlier.

SMT Remains Very Advantageous For 5th Gen AMD EPYC Performance Review - Phoronix

www.phoronix.com

If you take out the HPC oriented benchmarks (where others are correct that sometimes these even work better with SMT off), its pretty high.

Even with all the crappy SMT performance metrics in the summary it is 15% overall.

 

[jdubs03]

RPC?

Raptor Cove I presume.

 

[Thunder 57]

SMT Remains Very Advantageous For 5th Gen AMD EPYC Performance Review - Phoronix

www.phoronix.com

If you take out the HPC oriented benchmarks (where others are correct that sometimes these even work better with SMT off), its pretty high.

Even with all the crappy SMT performance metrics in the summary it is 15% overall.

15-20% is more realistic and still worth it it IMHO. But stating "40%" is just misinformation.

 

[Doug S]

RPC got 18% in specint at iso power

SPECint is the classic example of an "embarrassingly parallel" benchmark, since 'rate' runs are just multiple copies of the same benchmark with zero communication or dependence between them. Pretty much the best case for SMT. If your workload is like that great, but if it has much inter-thread dependency you don't want to be looking at SPECrate for guidance (or Cinebench)

 

[fastandfurious6]

why is specint so mystical? it requires such specialized installation yet it's still problematic

 

[511]

why is specint so mystical? it requires such specialized installation yet it's still problematic

the license cost $1000 as well and it's a compiler benchmark you have to use the correct compiler and flags across different CPU For a fair comparison most of the tech tubers don't have this skill

 

[Geddagod]

SPECint is the classic example of an "embarrassingly parallel" benchmark, since 'rate' runs are just multiple copies of the same benchmark with zero communication or dependence between them. Pretty much the best case for SMT.

Hardly the best case for SMT, considering many other benchmarks see much greater gains from SMT than specint.

If your workload is like that great, but if it has much inter-thread dependency you don't want to be looking at SPECrate for guidance (or Cinebench)

True, but I think every major company in server advertises with specint anyway.

why is specint so mystical? it requires such specialized installation yet it's still problematic

Who knows. Sounds like it's a bit of a meme even in conferences as well, regardless though it's also the one benchmark that every company seems to use when talking about perf.

 

[511]

Who knows. Sounds like it's a bit of a meme even in conferences as well, regardless though it's also the one benchmark that every company seems to use when talking about perf.

it's a proper benchmark and stresses all parts of the Core and the memory subsystem GB tries to be SPEC but fails

 

[regen1]

Someone shared it, decent interview.

 

[regen1]

Intel Takes Major Step in Plan to Acquire Chip Startup SambaNova

Intel has signed a term sheet to acquire the AI chip startup SambaNova Systems, two sources with direct knowledge of

dnyuz.com

Let's see how and where this goes

 

[DrMrLordX]

Another AI acquisition? Oh my.