Discussion Intel current and future Lakes & Rapids thread

[lobz]

Seems to me it is only talking about the HEDT (X) platform. Not sure how you read it that no new desktop products at all are coming out.

Click refresh in your browser so the picture can reload, you might have only seen the top half of it.

 

[uzzi38]

Where do you get this? 15% per generation is very respectable, and Intel cores are much larger which explains the better performance.

I never said Genoa wasn't impressive. I know full well it is. Think you misread my point there man.

 

[mikk]

I'm just a liiiitle bit hesitant on this one still because the physics score is so poor - a sign of poor stability. It falls behind your average 1065G7 score by about 1k points in CPU testing.

Not neccessarily poor, the physics score is 14% higher than Intels I7-1065G7 preview sample at 15W running with LPDDR4x. It could make sense if this sample is running at 15W and possibly DDR4. The reported base should be valid only for 28W because it's aparently the new default TDP. 15W is a TDP down option.

 

[JoeRambo]

The problem with "believing" Intel roadmaps is that they had zero credibility for like 4-5 years now. Even when things are supposedly released, they might be impossible to buy ( like 10980XE ). We had 3-4 generations of Lakes inserted in desktop, and in servers If someone bought a bunch of servers with say 6148 during late 2017, they are still golden.

Except time to replace said servers is approaching fast and Intel is not offering anything new. In fact the way things are in the market now, the price of equivalent 20C from Intel might be higher than in 2017.
So roadmaps mean nothing, HotChips presentations mean nothing. What is important for average Joe like me, is when can i buy gear from vendors with CPU that offers enough advancement per core and also increase in core counts without loss of peak clock.

 

[piokos]

See, unlike in PCs where for vast majority the upgrades are for pleasure, next generation server parts have a compelling reason to upgrade, such as lower TCO due to greatly improved perf/watt and immediately being able to replace large amount of servers with smaller amounts.

The idea of upgrading CPUs (putting a new one into an existing motherboard) is almost non-existent in server space (and pretty rare in workstations). So I hope that's not what you meant.

Absolutely, when you buy a new server, it's best to get the most modern generation.
But as I said before: the process of buying a server usually takes a while. You're opening a tender, collect offers, choose something, get it through the upper management, order, ship.
So the shiny new rack that arrives to your company is built from parts that were available few months earlier. And of course there's still a long way to go for actually starting them in production.

But of course there's the matter of price - previous gen servers can be had cheaper (OEMs and distributors cleaning inventory). It could easily beat the TCO gains you've mentioned above.

 

[IntelUser2000]

The iGPU performance is suprisingly lower than expected - slightly below the 4800U. Don't know if that's down to memory configuration or if @IntelUser2000's speculation about a G9 SKU is correct.

The perils of samples.

I don't think the leaks are showing us everything. Sharkbay alluded UP4 being only for Y, but Intel clearly showed a U chip in a UP4 config.

Knowing Intel, they'll make even more configurations than Icelake.

We don't know what TDP the leaked chips are running at either. The official move to 28W doesn't really matter in practice since majority of the devices are already running Icelake at 25W. What if the test samples are at 15W?

The idea of upgrading CPUs (putting a new one into an existing motherboard) is almost non-existent in server space (and pretty rare in workstations).

Of course not. This is Intel we're talking about.

It doesn't even apply to desktops. Most enthusiasts I reckon upgrade in 3-5 years, no normal person does a yearly refresh. Personally I see 50%+ gains in ST and 2-3x gains in MT to be worth it. Anything else is just wasting money.

But of course there's the matter of price - previous gen servers can be had cheaper (OEMs and distributors cleaning inventory). It could easily beat the TCO gains you've mentioned above.

They were pretty damn good at clearing inventory if 100% replacement happened in a year.

Yes it was the Broadwell-Skylake generation where that seemed to have changed but we don't have confirmation. It should still be very high, nothing like client.

 

[mikk]

Not neccessarily poor, the physics score is 14% higher than Intels I7-1065G7 preview sample at 15W running with LPDDR4x. It could make sense if this sample is running at 15W and possibly DDR4. The reported base should be valid only for 28W because it's aparently the new default TDP. 15W is a TDP down option.

I have to correct me more than 14%:

https://twitter.com/x/status/1268414277288955905

15W I7-1065G7 preview sample physics: 6475
graphics: 3677

24% higher physics score if it's running at 15W.

Intel Core i7-1065G7 Laptop Processor (Ice Lake)

Benchmarks, information, and specifications for the Intel Core i7-1065G7

www.notebookcheck.net

 

[IntelUser2000]

24% higher physics score if it's running at 15W.

Even if that's true, Icelake 25W being 24% better than Tigerlake 15W sucks. That's pretty much zero gain at same TDP.

 

[mikk]

Even if that's true, Icelake 25W being 24% better than Tigerlake 15W sucks. That's pretty much zero gain at same TDP.

24% gain at the same TDP assuming this sample runs at 15W.

 

[IntelUser2000]

24% gain at the same TDP assuming this sample runs at 15W.

At the worst, it should be equal, Icelake not being 20%+ faster.

 

[mikk]

At the worst, it should be equal, Icelake not being 20%+ faster.

Not sure what you are saying.

I7-1065G7 15W ~6500
I7-1065G7 25W ~10000

I7-1165G7 8000

It only makes sense if this I7-1165G7 was running with TDP down 15W because Tigerlake won't be slower at the same TDP, and then it's ~24% faster at the same TDP.

 

[IntelUser2000]

Fair enough.

Though that's overstating the differences.

Real differences between the applications are 25-30%. https://www.pcworld.com/article/3412071/intel-10th-gen-ice-lake-performance-benchmarks.html

Rocketlake's graphics is also underperforming. What's up with that? Are the drivers that far behind? Normally I wouldn't come to that conclusion but Xe uses software scheduling, and well if you can't make such a critical feature work correctly, then such serious impact to performance will happen.

 

[IntelUser2000]

So if Rocketlake is on 14nm, and has only 512KB cache, then its Sunny Cove.

Because according to Intel, Willow Cove's two main features are "New transistor optimizations" and "Cache redesign".

 

[lobz]

Where do you get this? 15% per generation is very respectable, and Intel cores are much larger which explains the better performance.

It's not like Intel's doing much better. In fact they need to do extra per year to make up for the fact that Willow Cove is only 5-7%.

And the roadmap is either flat out wrong or refers to a lower end segment. Example is Cascade Lake and Cooper Lake.



Not sure if its even reporting correctly. Every tweet for that chip the clocks go higher!

without the performance following it at all

 

[lobz]

The perils of samples.

I don't think the leaks are showing us everything. Sharkbay alluded UP4 being only for Y, but Intel clearly showed a U chip in a UP4 config.

Knowing Intel, they'll make even more configurations than Icelake.

We don't know what TDP the leaked chips are running at either. The official move to 28W doesn't really matter in practice since majority of the devices are already running Icelake at 25W. What if the test samples are at 15W?



Of course not. This is Intel we're talking about.

It doesn't even apply to desktops. Most enthusiasts I reckon upgrade in 3-5 years, no normal person does a yearly refresh. Personally I see 50%+ gains in ST and 2-3x gains in MT to be worth it. Anything else is just wasting money.



They were pretty damn good at clearing inventory if 100% replacement happened in a year.

Yes it was the Broadwell-Skylake generation where that seemed to have changed but we don't have confirmation. It should still be very high, nothing like client.

It didn't until a couple of years ago. Even if it's not the average, a lot of people upgraded to Zen 2 without changing the motherboard and I'm pretty sure they were pretty happy to be able to do so. If it was never a real desire / demand in the first place, nobody would have ever cared about AMD announcing dropping support for Zen 3 on 400 series mobos.

As for servers, well in that area I have no idea about this specific matter, so I'm inclined to believe you

 

[IntelUser2000]

@lobz I'd have upgraded from Bulldozer derivatives, but certainly not from original Zen.

It's not that 640KB isn't enough for everybody. It's just our wants have no ceiling.

 

[piokos]

Of course not. This is Intel we're talking about.

Being Intel or not doesn't matter. Servers aren't upgraded (and they're not DIY in general).

It doesn't even apply to desktops. Most enthusiasts I reckon upgrade in 3-5 years, no normal person does a yearly refresh. Personally I see 50%+ gains in ST and 2-3x gains in MT to be worth it. Anything else is just wasting money.

Well, you're the one who said that.
Apparently, a lot of AMD fans replace a CPU yearly (or even more often: https://forums.anandtech.com/threads/my-new-ryzen-5-3600-hits-4-5-ghz-1-28v.2581268/).
Just look around on forums like this one.

 

[IntelUser2000]

Being Intel or not doesn't matter. Servers aren't upgraded (and they're not DIY in general).

You understand what I mean. Upgrades are upgrades whether you throw out the old thing or replace parts.

Apparently, a lot of AMD fans replace a CPU yearly (or even more often: https://forums.anandtech.com/threads/my-new-ryzen-5-3600-hits-4-5-ghz-1-28v.2581268/).

Yes and that no doubt represents 100% of AMD owners. Not even one forgets a generation.

1 million people is a lot, but tiny when you consider the population of the world.

 

[lobz]

@lobz I'd have upgraded from Bulldozer derivatives, but certainly not from original Zen.

It's not that 640KB isn't enough for everybody. It's just our wants have no ceiling.

Most people I've seen were upgrading from a 1600 or 2600-like to 3700X or 3900X, so jumping a category too. But if you're gaming, even a 3600 is a huge upgrade over a 1600.
Even Techjesus Steve himself calls the 3600 AMD's Sandy Bridge.

I'm not saying that it's the market standard for all consumers planning on upgrading on the same motherboard, but even before Zen I've always heard the silent murmuring of people being pissed off about always having to upgrade to a new intel chipset. It's only become louder since they can't ignore the fact that Zen owners don't always have to.

 

[jpiniero]

I could see Icelake Server being Willow Cove actually, and Sapphire Rapids being Golden Cove.

 

[dullard]

It's not that 640KB isn't enough for everybody. It's just our wants have no ceiling.

You do realize that the 640 KB was the proposed MINIMUM limit? When you are dealing with a minimum, why do you then talk about ceilings?

 

[IntelUser2000]

Even Techjesus Steve himself calls the 3600 AMD's Sandy Bridge.

I'd call it their Athlon 64. And the original Ryzen chips the original Athlon. I do not see Alderlake changing the situation. Intel's comeback is probably towards 2023, not when 7nm is available(7nm is just a good foundation, but you also need to build the house). Sandy Bridge did not make this big of an impact.

But I see parallels potentially even further back. To 1990's when we had multiple vendors competing everywhere. Three GPU vendors, alternative server competition, rise of ARM, all this hasn't happened in decades! In that case a hypothetical Conroe won't crush AMD because they won't blunder like they did with Phenoms, and will instead continue to be a big threat.

You do realize that the 640 KB was the proposed MINIMUM limit? When you are dealing with a minimum, why do you then talk about ceilings?

I don't really care. It was partly a joke, if you did not get it.

 

[mikk]

Fair enough.

Though that's overstating the differences.

Real differences between the applications are 25-30%. https://www.pcworld.com/article/3412071/intel-10th-gen-ice-lake-performance-benchmarks.html

Rocketlake's graphics is also underperforming. What's up with that? Are the drivers that far behind? Normally I wouldn't come to that conclusion but Xe uses software scheduling, and well if you can't make such a critical feature work correctly, then such serious impact to performance will happen.

The RKL leak is a 0000 sample, it's not really meaningful at the moment. It's pretty obvious from the physics scores alone, slower than 9900K. And do you have the GPU clock speed from this sample? I guess this info is missing as well.

 

[jpiniero]

I wouldn't expect much from Rocket Lake's IGP given only 32 EUs.

 

[lobz]

Being Intel or not doesn't matter. Servers aren't upgraded (and they're not DIY in general).

Well, you're the one who said that.
Apparently, a lot of AMD fans replace a CPU yearly (or even more often: https://forums.anandtech.com/threads/my-new-ryzen-5-3600-hits-4-5-ghz-1-28v.2581268/).
Just look around on forums like this one.

Well you just had to say AMD fans in this sentence, right? It's OK though, it clarifies a lot about your posts here.