News Intel details 14A performance and new 'Turbo Cells' that unlock maximum CPU and GPU frequency

[Admin]

Intel shared performance metrics for its upcoming 14A process node and teased its new Turbo Cell technology, a customizable design approach aimed at providing maximum CPU frequency and boosting performance for critical speed paths in GPUs.

Intel details 14A performance and new 'Turbo Cells' that unlock maximum CPU and GPU frequency : Read more

 

[Findecanor]

Remember that "14A" is a trademark, not a measurement.

 

[TerryLaze]

Remember that "14A" is a trademark, not a measurement.

It is a marketing thing sure, but it is there to show that some progress was made, they also showed actual numbers. Or at least numbers, we will see when they come out if they are actual.
30% density 15-20% per/W ,looks pretty good.

 

[rluker5]

Turbo cells seems like something they already should have had. Along with Power Direct to minimize EM interference.

Getting stronger on the basics.

 

[Rob1C]

So, they plan on leaving the 6.2 GHz i9-14900KS in the dust?

 

[tennis2]

Chip design is amazing. How did humanity even get to this point at all?

 

[TheSecondPower]

Turbo cells seems like something they already should have had. Along with Power Direct to minimize EM interference.

Getting stronger on the basics.

Power Direct is the second iteration of Intel's backside power delivery, and by this schedule will come to market about the same time as TSMC's first iteration of backside power delivery.

 

[bit_user]

Turbo cells seems like something they already should have had.

They already do things like this, with cells designed for different power/density tradeoffs. For a long time, IIUC.

Another technique they've at least used in the past is to place multiple transistors in parallel, when necessary.

https://www.anandtech.com/show/4818/counting-transistors-why-116b-and-995m-are-both-correct
​

 

[bit_user]

Chip design is amazing. How did humanity even get to this point at all?

Engineering involves a lot of iteration and progressive refinement. You figure out what are the limiting factors, focus on addressing those, and then re-analyze and probably find new limitations that need new solutions. All along the way, the process is fed by thousands of brilliant minds and many $B of investment. That's why some country can't just replicate all of that progress overnight, no matter what resources they throw at it.

There's also something of a virtuous circle, where the machines produced by current technology enable us to design better machines tomorrow. Even with all the knowledge about how current chip manufacturing works, you couldn't just start turning out the same chips decades ago. Not only is there innovation along the entire supply chain, but also the sheer compute power needed to create photomasks for modern chips simply didn't exist very long ago. Not to mention place & route.

 

[bit_user]

So, they plan on leaving the 6.2 GHz i9-14900KS in the dust?

FWIW, the numbers quoted for performance & efficiency increases are usually based on a simple reference design. I think TSMC used an Arm Cortex-A7 core, for a long time. It's hard to predict exactly how they will translate into metrics of actual products made on these nodes.

In the case of Arrow Lake, it seems like a big part of what held it back was the SoC architecture. So, even though the process node and microarchitecture were both substantially improved, those got kneecapped by a SoC design that prioritized other things above performance.

 

[JamesJones44]

Chip design is amazing. How did humanity even get to this point at all?

Time and investment. It's been almost 80 years since the transistor was invented and it's probably one of the most heavily invested in areas over that time. Time and money can drive a lot of refinement in a product. Even cars and airplanes are orders of magnitude more complex than they were 80 years ago thanks to heavy amounts of investment over that time.

 

[S58_is_the_goat]

Since when did intel turn into a vaporware pr machine? They used to make cpus before...

 

[TerryLaze]

Since when did intel turn into a vaporware pr machine? They used to make cpus before...

Since 1991 when they came up with "intel inside" .
By the way, little known fact, they still make CPUs...

 

[S58_is_the_goat]

Since 1991 when they came up with "intel inside" .
By the way, little known fact, they still make CPUs...

They seem to be churning out more press releases then cpus though.

 

[TerryLaze]

They seem to be churning out more press releases then cpus though.

Doubtful, how many press releases did you count?!
Also they are preparing for a new gen of cpus and a whole new business (foundry) it would be a bad thing if they didn't have anything to say/advertise.

 

[acadia11]

Turbo cells seems like something they already should have had. Along with Power Direct to minimize EM interference.

Getting stronger on the basics.

Kind of like Rocky going back and running a few steps so he can take a proper whipping!

 

[FunSurfer]

What about some Turbo 3D cache for gaming, Intel?

 

[bit_user]

They seem to be churning out more press releases then cpus though.

Besides making its own CPUs, Intel is trying to operate a public foundry business, where external parties can contract them to fabricate chips.

A foundry has to make periodic announcements about the state of their manufacturing technology. That communication is how they attract new customers. It's also important for investor relations.

TSMC and Samsung (among others) all do the same thing. They periodically release process roadmaps and status reports.

 

[thestryker]

In the case of Arrow Lake, it seems like a big part of what held it back was the SoC architecture. So, even though the process node and microarchitecture were both substantially improved, those got kneecapped by a SoC design that prioritized other things above performance.

There's certainly design limitations, but 285K losing 300MHz across the board compared to the 14900K for P-cores is starting ~5% behind. While that isn't going to make up for low ring clock or the cross tile latency it's lost performance.

I have a hard time believing the lower clock speed is due to anything other than N3B. ARL runs into very similar clock limits as Zen 4/5 when overclocking (talking stable whole CPU). It's bad enough that an overclocked 285K isn't likely to match a stock 14900K clock speed wise.

This is something I'm really curious about with Intel 3 and 18A, but I'm not sure we'll ever find out. It seems unlikely that we'll see an unlocked CPU on Intel 3 (maybe GNR Xeon-W if that's a thing) and 18A isn't necessarily guaranteed either. Ian Cutress mentioned that there are already two products that have taped out on 18A-P and it doesn't seem unreasonable that one could be NVL.

 

[thestryker]

Since when did intel turn into a vaporware pr machine? They used to make cpus before...

They seem to be churning out more press releases then cpus though.

This event was entirely about foundry services so if that wasn't their focus they'd have failed. While the articles here on Tom's didn't cover all of it there was a significant amount regarding what's in place to support the new process nodes. They didn't announce major customers, but they slipped in mention of various tape outs. There was also a massive emphasis on utilization of and compatibility with industry standard tools.

Now if Intel was doing a product event and didn't talk about actual products that would be an issue.

 

[TheSecondPower]

There's certainly design limitations, but 285K losing 300MHz across the board compared to the 14900K for P-cores is starting ~5% behind. While that isn't going to make up for low ring clock or the cross tile latency it's lost performance.

I have a hard time believing the lower clock speed is due to anything other than N3B. ARL runs into very similar clock limits as Zen 4/5 when overclocking (talking stable whole CPU). It's bad enough that an overclocked 285K isn't likely to match a stock 14900K clock speed wise.

This is something I'm really curious about with Intel 3 and 18A, but I'm not sure we'll ever find out. It seems unlikely that we'll see an unlocked CPU on Intel 3 (maybe GNR Xeon-W if that's a thing) and 18A isn't necessarily guaranteed either. Ian Cutress mentioned that there are already two products that have taped out on 18A-P and it doesn't seem unreasonable that one could be NVL.

Also Arrow Lake mobile only recently came to market, but its clock speeds are more similar to Raptor Lake's than the desktop models were. Which is what we should expect if N3B is limiting the clock speed rather than Lion Cove just being less frequency-optimized.

 

[usertests]

30% density 15-20% per/W ,looks pretty good.

The density increase is low, but in line with TSMC's similarly low density increase: from N2 to A14, 10-15% perf, 25-30% power reduction, 1.23x logic density. Better numbers expected later when they introduce the Super Power Rail variation of A14.

It will be interesting to see chips start to approach 7 GHz, but the power reduction is the most useful improvement for most people, great for phones, tablets/laptops, and handhelds.

Chip design is amazing. How did humanity even get to this point at all?

Lots of money spent iterating on what worked before, and decades of basic research to produce the next big thing (such as EUV or GAAFETs). Computing has been the most practical application of nanotechnology, and is full of cash. As bit_user said, having computers helping design better computers is a game changer.

Hopefully we are not nearing a brick wall, and 3D fabrication is going to unlock orders of magnitude more performance.

 

[JRStern]

Since when did intel turn into a vaporware pr machine? They used to make cpus before...

Exactly.
At this rate there won't be an Intel in 2027.

 

[bit_user]

I have a hard time believing the lower clock speed is due to anything other than N3B. ARL runs into very similar clock limits as Zen 4/5 when overclocking (talking stable whole CPU). It's bad enough that an overclocked 285K isn't likely to match a stock 14900K clock speed wise.

Part of it is the node and part of it is how you design for it.

We know Intel originally planned for Arrow Lake to launch on 20A. Perhaps that version of at least the Lion Cove cores was more frequency-optimized. With Arrow Lake, and the time frame in which it was delivered on N3B, I think they probably just reused the layout they did for Lunar Lake, which would be more efficiency-optimized than frequency-optimized. If true, it doesn't mean more wasn't possible on N3B, just that they didn't have enough time to get there.

As for Zen 5, don't forget that it's on a N4-class node, not N3B. Zen 5C is on N3B, but their C-cores trade high frequencies for improved density and efficiency. And actually, I think even their full-frequency cores are more server-optimized than desktop-optimized. Just a quick look at the highest-clocking variant of Turin should drive home this point that server CPUs don't come anywhere close to the kinds of frequencies Zen easily hits on the desktop.

https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9575f.html
​

Yeah, that's a 4.5 GHz all-core boost, with max single-core boost at only 5 GHz. And for that, what you give up is fully half of the total cores you could have. And yet it still burns up to 400 W.

So, further frequency-optimizing Zen 5 would've benefited only desktop and not their main market - servers. Probably not a good thing for the high-end laptop CPUs that use them, either.

 

[bit_user]

Computing has been the most practical application of nanotechnology,

Particle physics, even?

Hopefully we are not nearing a brick wall, and 3D fabrication is going to unlock orders of magnitude more performance.

I don't know about orders of magnitude, but I'm no expert. It seems to me like power-density and fabrication costs are going to be very big challenges.

New materials could be a game changer, though. That's where I'm pinning most of my hopes. Although, eh... I don't know if I really care that much whether computers get any faster or not.