Discussion Intel current and future Lakes & Rapids thread

[H433x0n]

Oh but this gen is the closest we'll ever get to a semi-fair fight.

Then what’s the problem?

Some do it better than the others.
As I've said, only a few quarters of waiting left!

It’s hard to tell what to expect. You’ve said RPL wasn’t even in the same stratosphere as Zen 4, whereas I view them both as performing similarly with Zen 4 being much more efficient (while being on a superior process).

For all I know they end up performing similarly again while the internet calls it a blow out (similar to today).

 

[Joe NYC]

Then what’s the problem?


It’s hard to tell what to expect. You’ve said RPL wasn’t even in the same stratosphere as Zen 4, whereas I view them both as performing similarly with Zen 4 being much more efficient (while being on a superior process).

For all I know they end up performing similarly again while the internet calls it a blow out (similar to today).

That efficiency advantage really kicks in the server, which is where (and why) Zen 4 is in a different category then Sapphire Rapids, which is closer to Zen 3 than Zen 4.

Intel has some ground to make up on per core basis to Zen 4 level. And then, it has to make up Zen 4 to Zen 5 improvement.

Bigger part of Zen 4 leads vs SPR was core count, 96 vs. 60 for standard cores.

Intel going from ~64 to ~128 standard cores, while Genoa to Turin is 96 to 128 is a bigger increment for Intel. But that also means that power efficiency improvement, per core, has to be so much higher vs. what Zen 5 needs to achieve.

 

[Joe NYC]

Wasn't talking ARL at all.
There are far, far more important and interesting products launching in 2024.

MTL is a wee bit earlier but yes.
GNR vs Turin will be a fight to ugh, remember.
~same core count and wattage and and all.
Very fair! (Not really and you'll see why).

Is Intel still sharing the L3 cache and burning power to access L3 from different die?

Intel will also need greater gains in power efficiency, doubling the cores while also delivering big increases in per core performance to catch up with Zen 5.

So, Intel would probably need the biggest core improvement in its history to achieve both of these goals.

 

[adroc_thurston]

Then what’s the problem?

You will understand.

You’ve said RPL wasn’t even in the same stratosphere as Zen 4

It's not, PPA is still a thing.

 

[JoeRambo]

Is Intel still sharing the L3 cache and burning power to access L3 from different die?

Isn't that a feature? The problem is L3 size and mesh performance overall, the fact that compute die is full SoC with I/O etc. I think if Intel can add respectable sized L3, things would get better immediately. Taking 50ns hit on checking 100MB L3 cache is beyond stupid.
IF they produce compute chiplets with Cores + L3, have mem and IO on separate die AND still share L3 i'd be quite happy for plenty of workloads we have. Remember on AMD, once you go beyond 8 core sized workload, you communicate through memory.

 

[SiliconFly]

Because in its whole history AMD never made such derogatory statements, Intel on the other hand routinely practice such baseless statements, among others they once made a slide to spread FUD and saying that AMD was using defective dies for the Phenom X3...

Not really. AMD has said many absurd, filthy and derogatory things about Intel in the past.

And Intel saying that AMD is in the rear view mirror (in client) isn't that big an Insult. It's just a fact.

 

[igor_kavinski]

It's just a fact.

In terms of sheer volume, yes.

What's good for the world, no.

 

[Saylick]

Then what’s the problem?


It’s hard to tell what to expect. You’ve said RPL wasn’t even in the same stratosphere as Zen 4, whereas I view them both as performing similarly with Zen 4 being much more efficient (while being on a superior process).

For all I know they end up performing similarly again while the internet calls it a blow out (similar to today).

I think he's saying that in terms of fairness, Zen 5 Turin vs. Granite Rapids (and possibly Zen 5 DT vs ARL too) will be comparable in terms of similar node, similar core area, and possibly similar number of transistors thrown at it. Therefore, one can say that this is as level of a playing field as you can get in a long time. Previously, Intel had a node advantage, then AMD had a node advantage, which muddied the waters.

Given similar node characteristics, it then boils down to which company designs a better core, for the one with the better core will likely have better PPA. I think he's saying that AMD's Zen 5 team will do a better job than Intel's Israeli Design Center.

 

[Henry swagger]

In terms of sheer volume, yes.

What's good for the world, no.

Intel is good for the world.. they invented the cpu they've been blessed

 

[igor_kavinski]

Given similar node characteristics

Maybe similar nodes but Intel usually leads in transistor density. Is that going to be the case for Intel 3 this time?

 

[igor_kavinski]

they invented the cpu they've been blessed

Wright brothers invented the plane. Where is the Wright Bros Corporation today?

Continuous innovation is required to maintain an edge over the competition. Nokia and Blackberry, both innovators, lost to Google and Apple, companies that weren't even their direct competitors and not even in the telecom sector.

 

[Saylick]

Intel is good for the world.. they invented the cpu they've been blessed

Intel is also the company that brought us 5% gen over gen gains for almost a decade. Gotta love it when the incumbent market leader charges us $300 for the same 4 core CPUs every 2 years, right? /s

Yeah, I have no desire for Intel to return to its "former glory". They do not deserve it, or rather no company deserves anything. A return to the former status quo is simply not good for the industry.

 

[Joe NYC]

Isn't that a feature? The problem is L3 size and mesh performance overall, the fact that compute die is full SoC with I/O etc. I think if Intel can add respectable sized L3, things would get better immediately. Taking 50ns hit on checking 100MB L3 cache is beyond stupid.
IF they produce compute chiplets with Cores + L3, have mem and IO on separate die AND still share L3 i'd be quite happy for plenty of workloads we have.

I see, so bigger L3 would make it more worth while to take the hit (to check), because there would more likely be a cache hit rather than miss.

If MLID leak is close to correct, and GNR has 3 large die with ~44 cores each, I don't think there is going to be room for a lot more L3. Maybe some more, but nothing like a gigabyte.

Remember on AMD, once you go beyond 8 core sized workload, you communicate through memory.

We had a peak at future AMD designs, and it seems that's what AMD is going with (going forward). Turbo charged with MALL (Memory Attached Last Level Cache). So seemingly sending the request to memory, but having a cache there to serve some requests from cache.

So, it seems (for now), AMD has given up on any direct sharing of L3. I wonder if Venice changes it in any way, or if AMD just strengthens the MALL.

It may be possible to have some cache coherency algorithms with homogeneous chiplets, but likely impossible with heterogeneous package, sharing CPU, GPU and maybe some AI or other dedicated chiplets.

 

[Henry swagger]

Intel is also the company that brought us 5% gen over gen gains for almost a decade. Gotta love it when the incumbent market leader charges us $300 for the same 4 core CPUs every 2 years, right? /s

Yeah, I have no desire for Intel to return to its "former glory". They do not deserve it, or rather no company deserves anything. A return to the former status quo is simply not good for the industry.

they will return and amd is no better lol

 

[SiliconFly]

Wright brothers invented the plane. Where is the Wright Bros Corporation today?

Continuous innovation is required to maintain an edge over the competition. Nokia and Blackberry, both innovators, lost to Google and Apple, companies that weren't even their direct competitors and not even in the telecom sector.

Assuming Intel isn't innovating now is just plain wrong. All the upcoming new nodes, new architectures & new techs is more than what we've ever seen before. When it comes to innovation, they've set they own tail on fire just to stay competitive. They're just getting started with MTL.

Intel will prevail. Thats for sure. The big question is, will AMD? After all, AMD managed to grab only a tiny piece of the market when Intel fell flat on the ground. When Intel gets up & starts pushing forward, what happens to AMD?

 

[SiliconFly]

Intel is also the company that brought us 5% gen over gen gains for almost a decade. Gotta love it when the incumbent market leader charges us $300 for the same 4 core CPUs every 2 years, right? /s

Yeah, I have no desire for Intel to return to its "former glory". They do not deserve it, or rather no company deserves anything. A return to the former status quo is simply not good for the industry.

Yep. They shouldn't return to the "former glory". Not good for any of us. Otherwise, we'll be getting the same old thing again & again with a 5% performance bump every year for the next 10 years! Competition is essential for consumers like us.

I'd be very happy if both the companies are financially healthy, has decent market share & innovates like there's no tomorrow. That would be a dream come true!

 

[ondma]

Intel is also the company that brought us 5% gen over gen gains for almost a decade. Gotta love it when the incumbent market leader charges us $300 for the same 4 core CPUs every 2 years, right? /s

Yeah, I have no desire for Intel to return to its "former glory". They do not deserve it, or rather no company deserves anything. A return to the former status quo is simply not good for the industry.

Well, if you didnt like Intel's pricing, you were always free to buy AMD (oh, wait, even with the 5% increases year/year, Intel was far superior to AMD). I just dont get this good guy/bad guy metric. They are both companies and will charge as much as they can get away with for their products.

 

[SiliconFly]

Well, if you didnt like Intel's pricing, you were always free to buy AMD (oh, wait, even with the 5% increases year/year, Intel was far superior to AMD). I just dont get this good guy/bad guy metric. They are both companies and will charge as much as they can get away with for their products.

Not really. Previously, if we didn't like Intel's price or performance, there wasn't a good alternative. Thankfully, we have one now. Because of the competition, we now have multi-core cpus with high ipc, zillion threads, terrific performance-per-watt and a decent price. Thank god for that.

Otherwise, Intel will still be pushing out the stupid old quad-core cpus on Intel 14nm++++++++++ even now. And rocket lake would have been their biggest achievement till date. A crown jewel we can say! 👑

 

[A///]

thought this was an intel node and cpu discussion thread, not the let's slobber on Intel like they're the third coming of christ thread.

 

[igor_kavinski]

they will return and amd is no better lol

I hope you are alive when they "return"...

 

[igor_kavinski]

Assuming Intel isn't innovating now is just plain wrong.

Intel's problem is that they are innovating in isolation, without caring about what others are doing better. They suffer from the "not invented here" syndrome and their cabal-like mentality prevents them from accepting good advice from outsiders. It will be interesting to see how they market MTL. So far, it all seems to be about AI which means the core performance and battery life is not worth boasting about relative to the competition.

 

[H433x0n]

Intel's problem is that they are innovating in isolation, without caring about what others are doing better. They suffer from the "not invented here" syndrome and their cabal-like mentality prevents them from accepting good advice from outsiders.

There's a case to be made for this mindset. It's a 'winners' strategy but I acknowledge it is a dual edged sword. For example, you can credit a large chunk of SPR's sales to their accelerators that were developed as a result of this mindset.

I think it's something that Radeon needs desperately, they're always responding to Nvidia's latest thing, which is a losing strategy.

It will be interesting to see how they market MTL. So far, it all seems to be about AI which means the core performance and battery life is not worth boasting about relative to the competition.

Battery life will be great with MTL, it's a generational leap in that aspect.

 

[SiliconFly]

Battery life will be great with MTL, it's a generational leap in that aspect.

Definitely a generational leap. All the info leaks fit like a jigsaw puzzle. MTL should be a hyper-power-efficient chip. And ARL might take it to the next level next year.

Only a few more days to go for MTL...

 

[ondma]

There's a case to be made for this mindset. It's a 'winners' strategy but I acknowledge it is a dual edged sword. For example, you can credit a large chunk of SPR's sales to their accelerators that were developed as a result of this mindset.

I think it's something that Radeon needs desperately, they're always responding to Nvidia's latest thing, which is a losing strategy.


Battery life will be great with MTL, it's a generational leap in that aspect.

A "generational leap" from Intel's previous offerings, or a generational leap that makes them more efficient than AMD? I dont really follow the mobile market, but I would assume AMD is far ahead in power consumption in mobile as well, so a "generational leap" from intel might only reach parity with AMD's current offerings. We will see when (if?) Intel reaches process lead or parity, but I think they need a new core design to get power usage down to the level they need.

 

[igor_kavinski]

Definitely a generational leap.

Sounds like they will be maximizing the LP core usage in the SoC tile. I don't see how else they can reduce power, unless their Thread Director coupled with new as-yet-unreleased scheduler improvements in Win11, has become 99% accurate in predicting which core(s) will complete the task most efficiently, all the while not letting the user sense the slowed down frequency on battery power.