News Intel Engineer Outs Panther Lake Architecture on LinkedIn

[SiliconFly]

In a desktop environment you do not want to be thermally throttling. If you are that means your cooling solution is not good enough for your usage. Not to mention that having your CPU hit the thermal ceiling all the time can lead to premature aging/death of the CPU.


Arrow Lake isn't coming out until end of 2024/beginning of 2025 at the earliest. That said even if Intel's node is superior that doesn't mean anything. Intel still cannot get their power draw under control on their P cores. Do you really think Intel would have gone to a BIG.little design if they could have gotten 16 P cores to clock at a decent level all while using lower power? Of course not they would have stayed with nothing but P cores. However, they knew how much power their P cores used and needed to reign in the absolute power requirements so the went hybrid. AMD on the other hand has been able to put 16 big cores all on the same die while keeping power under control. This is because Zen is more efficient that Core. We see this in benchmarks that limit the power (which I referenced earlier). Intel needs double the power draw to equal AMDs performance.


Again Intel needed to throw power limits out the window in able to beat Zen 4. As I had said earlier Core does scale better with additional power than Zen. Once Zen got to 105W the scaling beyond that was minimal. Core on the other hand kept scaling pretty well all the way up to 253W. While that isn't a big deal on the desktop, that can be an issue in laptops. The Intel laptops need more power to keep up with Zen but that has the other problem of reducing battery life. Since Zen 2 was released, on similarily configured laptops AMD has had the battery life advantage over Intel. We also knew that Zen 4 3D was going to be released long before Zen 4 was released. AMD released roadmaps with Zen 4 3D on them over a year ago. One other thing don't forget that Intel made the 13900KS in an attempt to beat Zen4 3D in gaming.

Client computing isn't where companies like Intel and AMD make most of their money. Most of their money is made on data center products. Intel not doing well in data center products is a big reason why they posted a net loss of more than $600M in Q4 2022. Intel needs to strengthen their data center products to make the company financially stable. They cannot rely on client computing to make a profit. Without profits they cannot afford the R&D on new process nodes or CPU designs. Right now Intel is at least 2-3 years behind AMD in the data center market. If AMD keeps executing like they have, Intel will stay 2-3 years behind.

I want Intel to do well same as I want AMD to do well. Competition is the market is the best thing for everyone. It leads to innovation, lower prices, and better performance. The fact that you seem to want AMD to fail outright is a scary thought. Had Zen 1 not been a success, they bet the entire company on the Zen line and it was a bet that paid off, AMD would probably have gone bankrupt. Had that happened do you really think the i3's would be 4c/8t and the i7's 8c/16t? Of course not. We would have been stuck with i7's still being 4c/8t, maybe 6c/12t with little need to Intel to innovate beyond Skylake.

When, Nvidia jumped from Ampere to Lovelace they went from Samsung N8 straight to TSMC N4X! Thats a 1.5 node jump (pretty close to 2 if you ask me). And they managed to more than double performance. Same applied to 20A. With Arrow Lake they're taking a massive jump from 4 to 2. Thats a massive1.5 node jump like Nvidia. It ain't trivial.

 

[jeremyj_83]

When, Nvidia jumped from Ampere to Lovelace they went from Samsung N8 straight to TSMC N4X! Thats a 1.5 node jump (pretty close to 2 if you ask me). And they managed to more than double performance. Same applied to 20A. With Arrow Lake they're taking a massive jump from 4 to 2. Thats a massive1.5 node jump like Nvidia. It ain't trivial.

You are comparing apples to oranges so your point is bogus.

 

[bit_user]

If you have issues with my predictions, you should learn to ignore.

That's not how it works. You can't just post whatever you want and then attack anyone who disagrees with you.

You're free to post your opinions and supporting information, but it crosses a line when you start insisting they either agree or shut up.

 

[bit_user]

When, Nvidia jumped from Ampere to Lovelace they went from Samsung N8 straight to TSMC N4X! Thats a 1.5 node jump (pretty close to 2 if you ask me). And they managed to more than double performance.

Performance depends on more than just the node. For one thing, RTX 4000 also increased L2 cache by about 10x, basically following what AMD did with Infinity Cache.

Another point to consider is that GPUs can scale performance, as a function of density, much more easily than CPUs.

Same applied to 20A. With Arrow Lake they're taking a massive jump from 4 to 2. Thats a massive1.5 node jump like Nvidia. It ain't trivial.

In addition to what I said above, you also can't translate node jumps like that. If you compare Nvidia's generational density increase with moving from Intel 4 to Intel 20A, I'm sure it's much greater. Again, 20A is just a name. Plus, there are more factors at play than just density.

 

[TerryLaze]

Actually... Intel's latest results had $6.6B in Client Computing Group and $4.3B in Data Center & AI:

https://www.tomshardware.com/news/intel-posts-largest-loss-in-years-as-sales-of-pc-and-server-cpus-nosedive

And another 2.1bil in network and edge, which are datacenter and server products as well.
Just saying.

I slows transistor aging and electromigration. These compromise the maximum boost clocks of newer CPUs, and outright stability of older ones.

That would mean that amd/intel are lying about the safety margins and that would open them up to lawsuits without end.
Within the operating temps stated on their pages there should be zero aging going on.

If they didn't throttle until the thermal trip point, then the result would be CPUs wearing out prematurely - even before the warranty period expires! The trip point is just there to avoid instantaneous destruction of the CPU, in situations like: "Oh noes!!!111 my heatsink fell off!"

Yes.
And the throttle point is much lower and set there because there is no real danger in reaching it, at least that's what common sense would suggest.

 

[SiliconFly]

That's not how it works. You can't just post whatever you want and then attack anyone who disagrees with you.

You're free to post your opinions and supporting information, but it crosses a line when you start insisting they either agree or shut up.

Well, you're the first to attack me by saying bizarre. Honestly, if you ask me, which company who likes is their own preference. No one has the right to say otherwise! Ok. I have more important things to do. So, I'm moving on.

 

[jeremyj_83]

I'd put Sapphire Rapids only 1 generation behind. It fares well against EPYC Milan (Zen 3).

Agreed that it is only 1 generation behind AMD. However, server products tend to be on a 2ish year release schedule which is why I said Intel is 2-3 years behind AMD in data center products. For reference Genoa was released Nov 10, 2022 vs Jan 10, 2023 for Sapphire Rapids.

Actually... Intel's latest results had $6.6B in Client Computing Group and $4.3B in Data Center & AI:

I should have looked at the financial statements thank you.

Intel's server cpus suck so hard!

You do realize that Sapphire Rapids uses Golden Cove cores which are the same P cores as in Alder Lake. By saying their server CPUs suck so hard you are saying that the P core design is bad and that is the same P core as in the consumer products. Now there are difference to be used in consumer vs server products, however, the underlying execution units, etc... is still the same.

Sapphire Rapids itself isn't a bad CPU. Its biggest issue is it was release 1.5 years late. Had it been released on time it would have been compared against Epyc Milan. Those two CPUs compete quite well with each other. However, instead it was released after Epyc Genoa which is Zen 4 based and brings about more performance.

 

[bit_user]

Well, you're the first to attack me by saying bizarre.

I did not say you were bizarre. I just meant that I cannot understand why you're being so insistent about Intel's forthcoming dominance. Please feel free to enlighten us.

Note that I did not even say your prediction is wrong! I just think it's too early to be so sure.

 

[SiliconFly]

Performance depends on more than just the node. For one thing, RTX 4000 also increased L2 cache by about 10x, basically following what AMD did with Infinity Cache.

Another point to consider is that GPUs can scale performance, as a function of density, much more easily than CPUs.


In addition to what I said above, you also can't translate node jumps like that. If you compare Nvidia's generational density increase with moving from Intel 4 to Intel 20A, I'm sure it's much greater. Again, 20A is just a name. Plus, there are more factors at play than just density.

Well, I didn't mean Intel is getting a 2X performance leap. All I'm saying is that they're getting a massive node advantage. It's huge!

 

[bit_user]

That would mean that amd/intel are lying about the safety margins and that would open them up to lawsuits without end.
Within the operating temps stated on their pages there should be zero aging going on.

They don't guarantee boost/turbo speeds, and that's what suffers (within the warranty period).

Beyond the warranty period, it could start to affect product reliability & longevity.

 

[SiliconFly]

Performance depends on more than just the node. For one thing, RTX 4000 also increased L2 cache by about 10x, basically following what AMD did with Infinity Cache.

Another point to consider is that GPUs can scale performance, as a function of density, much more easily than CPUs.


In addition to what I said above, you also can't translate node jumps like that. If you compare Nvidia's generational density increase with moving from Intel 4 to Intel 20A, I'm sure it's much greater. Again, 20A is just a name. Plus, there are more factors at play than just density.

Well, I didn't mean In

Agreed that it is only 1 generation behind AMD. However, server products tend to be on a 2ish year release schedule which is why I said Intel is 2-3 years behind AMD in data center products. For reference Genoa was released Nov 10, 2022 vs Jan 10, 2023 for Sapphire Rapids.


I should have looked at the financial statements thank you.


You do realize that Sapphire Rapids uses Golden Cove cores which are the same P cores as in Alder Lake. By saying their server CPUs suck so hard you are saying that the P core design is bad and that is the same P core as in the consumer products. Now there are difference to be used in consumer vs server products, however, the underlying execution units, etc... is still the same.

Sapphire Rapids itself isn't a bad CPU. Its biggest issue is it was release 1.5 years late. Had it been released on time it would have been compared against Epyc Milan. Those two CPUs compete quite well with each other. However, instead it was released after Epyc Genoa which is Zen 4 based and brings about more performance.

Golden Cove is excellent. But in server space, if given a fresh/fair choice, a user would surely opt for a 96C cpu with slightly less performance per core rather than a 56C cpu with higher performance (assuming we aren't talking about other factors). As of now, sapphire rapids doesn't look like a strong enough competitor to EPYC.

One more thing. Golden Cove is actually very good as it scales well with frequency. But it's way too fat & Intel inspite of cutting down AVX, still couldn't make it slimmer. Hope they retire their fat cores sooner and replace it more advanced lean & mean cores.

 

[SiliconFly]

I did not say you were bizarre. I just meant that I cannot understand why you're being so insistent about Intel's forthcoming dominance. Please feel free to enlighten us.

Note that I did not even say your prediction is wrong! I just think it's too early to be so sure.

That i agree. It's way to early to be sure. But numbers don't lie. If they manage to do what they say, things are going to be a lot better for them for sure. And the client group still hasn't missed a step since Q4 2020.

Looking back, golden cove performed well. So does raptor cove. After all, it did manage to keep up mostly with Zen 4. Even assuming (just assume for now) that their Redwood Cove (meteor lake) & Lion Cove (arrow lake) don't have architectural IPC increases, Lion Cove is still going to get a lot of free IPC increase compare to Raptor Cove due to the massive 1,5X node jump from Intel 4 to 20A (assuming same power draw). And due to the massive density increase due to the 1.5X jump, they'll be able to increase cache sizes too to a certain level which in turn improves IPC again. Now add two generation of architectural improvements to IPC too (not much but some, i'll explain later). Arrow Lake isn't trivial. It's a beast. Meteor Lake not so much. A Zen 4 refresh might easily & completely decimate meteor lake actually even if comes out with 6P + 16E. 2023 is AMD's year. 2024 is Intel's.

The only question is, whether Intel will be able to pull off Arrow Lake on 20A on time? And many reports & press releases keep mentioning (aggressively) that they're on track. And unlike their server group & AXG, their client group hasn't slipped in the last couple of years. That says something.

There are always some ifs and buts. But no huge show stoppers I believe. Thats why I think it's actually going to happen.

 

[jeremyj_83]

Well, I didn't mean In


Golden Cove is excellent. But in server space, if given a fresh/fair choice, a user would surely opt for a 96C cpu with slightly less performance per core rather than a 56C cpu with higher performance (assuming we aren't talking about other factors). As of now, sapphire rapids doesn't look like a strong enough competitor to EPYC.

One more thing. Golden Cove is actually very good as it scales well with frequency. But it's way too fat & Intel inspite of cutting down AVX, still couldn't make it slimmer. Hope they retire their fat cores sooner and replace it more advanced lean & mean cores.

Don't forget that SPR was scheduled to go against Epyc Milan but kept getting delayed. Same as Ice Lake was supposed to be against Epyc Rome. However, due to those delays SPR is against Genoa and Ice Lake was against Milan. In other words they are a generation behind in the server world. In benchmarks SPR does compete well with Milan and same as Ice Lake benchmarked competitively with Rome. Basically they win one for every one they lose.

As a VMware Administrator I would love to get some dual 96c/192t Genoa's with 6TB RAM/server. However, licensing costs for the CPUs is a pain. VMware licenses per 32 physical core/socket. That means a 24c CPU needs 1 license but a 96c CPU needs 3 licenses. At $5000/license the costs add up VERY quickly. Often times in a virtual environment you are more constrained by RAM than CPU anyways. My servers are all dual socket Epyc Rome 32c 7502's with 1TB RAM/host. On one host I have 37 powered on VMs and my CPU utilization sits around 10%, however, my RAM usage is at 67% as I have VMs with more than 100GB RAM. Basically CPU over provisioning (handing out more virtual CPUs that you have physical CPUs) is easy and you can usually go at lest 25% over if not more (with dual 32c/64t you can probably easily do 50-75% as resource contention isn't very high). However, over provisioning on RAM (handing out more RAM than is physically on the host) hits performance issues VERY quickly. I've noticed VMs performance issues at a 10% RAM over provision and had applications crash by 20%. Basically what I am saying is that sure those halo parts are nice, but you will see more companies buying the 32c variants.

 

[SiliconFly]

Don't forget that SPR was scheduled to go against Epyc Milan but kept getting delayed. Same as Ice Lake was supposed to be against Epyc Rome. However, due to those delays SPR is against Genoa and Ice Lake was against Milan. In other words they are a generation behind in the server world. In benchmarks SPR does compete well with Milan and same as Ice Lake benchmarked competitively with Rome. Basically they win one for every one they lose.

As a VMware Administrator I would love to get some dual 96c/192t Genoa's with 6TB RAM/server. However, licensing costs for the CPUs is a pain. VMware licenses per 32 physical core/socket. That means a 24c CPU needs 1 license but a 96c CPU needs 3 licenses. At $5000/license the costs add up VERY quickly. Often times in a virtual environment you are more constrained by RAM than CPU anyways. My servers are all dual socket Epyc Rome 32c 7502's with 1TB RAM/host. On one host I have 37 powered on VMs and my CPU utilization sits around 10%, however, my RAM usage is at 67% as I have VMs with more than 100GB RAM. Basically CPU over provisioning (handing out more virtual CPUs that you have physical CPUs) is easy and you can usually go at lest 25% over if not more (with dual 32c/64t you can probably easily do 50-75% as resource contention isn't very high). However, over provisioning on RAM (handing out more RAM than is physically on the host) hits performance issues VERY quickly. I've noticed VMs performance issues at a 10% RAM over provision and had applications crash by 20%. Basically what I am saying is that sure those halo parts are nice, but you will see more companies buying the 32c variants.

Oh. Didn't know that! Thats a lot of valuable insight! Thank you.

 

[The Historical Fidelity]

That would mean that amd/intel are lying about the safety margins and that would open them up to lawsuits without end.
Within the operating temps stated on their pages there should be zero aging going on.

Actually, simply having electrons flowing causes aging, the greater the current flow, the greater the friction in the wire, the greater the friction, the more heat is released, the more heat that is released, the greater the resistance of the wire, the greater the resistance, the greater the energy is transformed into heat. Also the greater the resistance, the greater the voltage is needed to bring the same power to the component at the end of the wire. If you are not careful, these positive feedback loops will easily get out of hand allowing electromigration and other issues to accelerate.
All of these variables will eventually degrade the CPU, but if you keep everything in check it shouldn’t happen for a very long time.
AMD, Intel, and Nvidia utilize formulas that predict failure rates for different inputs of parameters. So what they say is safe is simply what they could get away with spec wise that maintains a predicted less than 0.1% failure rate within the warranty period.

 

[SiliconFly]

That i agree. It's way to early to be sure. But numbers don't lie. If they manage to do what they say, things are going to be a lot better for them for sure. And the client group still hasn't missed a step since Q4 2020.

Looking back, golden cove performed well. So does raptor cove. After all, it did manage to keep up mostly with Zen 4. Even assuming (just assume for now) that their Redwood Cove (meteor lake) & Lion Cove (arrow lake) don't have architectural IPC increases, Lion Cove is still going to get a lot of free IPC increase compare to Raptor Cove due to the massive 1,5X node jump from Intel 4 to 20A (assuming same power draw). And due to the massive density increase due to the 1.5X jump, they'll be able to increase cache sizes too to a certain level which in turn improves IPC again. Now add two generation of architectural improvements to IPC too (not much but some, i'll explain later). Arrow Lake isn't trivial. It's a beast. Meteor Lake not so much. A Zen 4 refresh might easily & completely decimate meteor lake actually even if comes out with 6P + 16E. 2023 is AMD's year. 2024 is Intel's.

The only question is, whether Intel will be able to pull off Arrow Lake on 20A on time? And many reports & press releases keep mentioning (aggressively) that they're on track. And unlike their server group & AXG, their client group hasn't slipped in the last couple of years. That says something.

There are always some ifs and buts. But no huge show stoppers I believe. Thats why I think it's actually going to happen.

To wrap up all the Coves:

Golden Cove was a big ramp up.
Raptor Cove is just a minor refresh
Redwood Cove (meteor lake) taped out last year itself. Meaning, that too possibly is just a simple refresh.

Not much is known about Lion Cove. But we can make an educated guess. Considering 20A is a brand new node with a host of new technologies, I think its wise to say Intel may not choose to implement a much improved/modified architecture with Arrow Lake. It's very risky and they may iterate too much and lose a LOT of time if things don't work out well. So, the best guess is Lion Cove will bring in some architectural IPC improvements in lower single digit is my guess (not counting IPC gains due to node jump).

 

[jeremyj_83]

(not counting IPC gains due to node jump).

Going to a different process node does not increase IPC. The smaller node can allow for higher clock speeds at the same power level which increases performance or same clock speeds and thus performance but at lower power consumption. If you want to increase IPC you need to make changes to the uarch, increase cache, etc..

 

[TerryLaze]

Actually, simply having electrons flowing causes aging, the greater the current flow, the greater the friction in the wire, the greater the friction, the more heat is released, the more heat that is released, the greater the resistance of the wire, the greater the resistance, the greater the energy is transformed into heat. Also the greater the resistance, the greater the voltage is needed to bring the same power to the component at the end of the wire. If you are not careful, these positive feedback loops will easily get out of hand allowing electromigration and other issues to accelerate.
All of these variables will eventually degrade the CPU, but if you keep everything in check it shouldn’t happen for a very long time.
AMD, Intel, and Nvidia utilize formulas that predict failure rates for different inputs of parameters. So what they say is safe is simply what they could get away with spec wise that maintains a predicted less than 0.1% failure rate within the warranty period.

That's why all of these (amps, volts, and watts) have limits.
They can't get into a feedback loop and rise into infinity, unless you are overclocking, and all of these fears do come from extreme overclocking.
If you stay within the allowed vcore and the allowed watts then hitting the throttle temps shouldn't create any added aging at all because with the temps at the limit one or all of the three values will drop or just stay where they are.

 

[SiliconFly]

Going to a different process node does not increase IPC. The smaller node can allow for higher clock speeds at the same power level which increases performance or same clock speeds and thus performance but at lower power consumption. If you want to increase IPC you need to make changes to the uarch, increase cache, etc..

I meant performance increase at similar power levels.

 

[The Historical Fidelity]

That's why all of these (amps, volts, and watts) have limits.
They can't get into a feedback loop and rise into infinity, unless you are overclocking, and all of these fears do come from extreme overclocking.
If you stay within the allowed vcore and the allowed watts then hitting the throttle temps shouldn't create any added aging at all because with the temps at the limit one or all of the three values will drop or just stay where they are.

Right but I was refuting your “zero aging” comment which is incorrect, simply having electrons moving through the CPU ages it. Your “added aging” comment is correct, if you keep your cpu stock then aging will be slow enough to where less than 0.1% of CPU’s will require warranty replacement

 

[SiliconFly]

Right but I was refuting your “zero aging” comment which is incorrect, simply having electrons moving through the CPU ages it. Your “added aging” comment is correct, if you keep your cpu stock then aging will be slow enough to where less than 0.1% of CPU’s will require warranty replacement

Zero aging is impossible even in theory! Even the lifetime of the our universe is finite. Meaning, anything and everything will cease to operate and/or exist over time.

I believe when he mentioned zero aging, I think he would have implied something like many years. Or maybe even a decade or two. Just thinking out of the box.

 

[The Historical Fidelity]

Intel's Server & AXG groups suck! No doubts there. But their client CPU group is currently rocking. Alder Lake was revolutionary & on time. Raptor Lake beat Zen 4 in performance and was released 1 month in advance. It took Zen 4 3D VCache models to beat Raptor Lake 3 months later. Meteor lake is perfectly on track for a Q3 release. Arrow Lake is progressing very very nicely without any hiccups. There's no stopping Intel on the client side.

AMD's road map is now in shambles as they're stuck with inferior TSMC nodes compared to far superior nodes that Intel has starting 2024. AMD had an unfair node advantage over Intel all these (5+) years. Intel now has that same unfair node advantage starting next year. Zen 5 next year will be dead on arrival like Zen 4. AMD future is now in question. AMD might die sooner than you think. My condolences.

Well I just got updates from sources in Intel and they say that meteorlake desktop has been cancelled and some ML laptop dies will fill the lower ranks of the raptor lake refresh that they will be rushing to market because arrow lake on Intel 3 (it will not be on 20A as you say) is being delayed to 2024 due to yield/defect issues. Definitely not a rocking client side either.

But I do agree with you on AMD, if their completely new ground up core design for Zen5 isn’t impressive, AMD will need to abandon their high price points.

 

[bit_user]

Lion Cove is still going to get a lot of free IPC increase compare to Raptor Cove due to the massive 1,5X node jump from Intel 4 to 20A

That's not how it works. Moving to a smaller node makes CPUs faster in two main ways:

1. Gives you additional transistor budget that you can spend on architectural improvements or things like more cache.
2. Enables higher frequencies, as shrinking distances within the same design naturally results in tighter timing.

The new wrinkle that's recently come to light is that SRAM scaling has broken down. So, increased density applies mostly just to logic.

TSMC's 3nm Node: No SRAM Scaling Implies More Expensive CPUs and GPUs

Big problems from tiny memory cells.

www.tomshardware.com

Say goodbye to free cache enlargements. This is one huge reason AMD's strategy of using chiplets for cache (on both CPUs and GPUs) is so key. Intel will have to follow. In recent generations, Intel has been increasing cache quite a lot. That can't continue, without significantly affecting price.

That's why I'm saying node improvements ain't what they used to be.

their client group hasn't slipped in the last couple of years.

They're rumored to have cancelled Meteor Lake for the desktop. If so, that must be counted as a miss. If the desktop spends a 3rd year on Golden Cove, a lot of people are going to start having that Skylake de ja vu.

 

[bit_user]

Basically what I am saying is that sure those halo parts are nice, but you will see more companies buying the 32c variants.

What about cloud operators and hyperscalers? What hypervisors do companies like Facebook, Amazon, MS, and Google use?

Also, VMware doesn't seem to be the only game in town. Linux has KVM, for instance. And containers are now being used in a lot of cases that were previously handled by VMs.

 

[jeremyj_83]

What about cloud operators and hyperscalers? What hypervisors do companies like Facebook, Amazon, MS, and Google use?

Also, VMware doesn't seem to be the only game in town. Linux has KVM, for instance. And containers are now being used in a lot of cases that were previously handled by VMs.

Your large cloud providers will get the halo products. The reason Epyc Bergamo is being made is so that cloud providers can have even higher density. Google and AWS use KVM I believe and MS uses Hyper-V.

I know there are other hypervisors around. I was just speaking from my experience as a VMware Admin. I have never talked with another admin at conferences whos company uses CPUs with more than 32 cores. That said when it comes to non cloud providers VMware is still the biggest player. Newer versions of VMware allow for use of containers at the same time. Granted now that Broadcom bought VMware I don't know how long VMware will stay the top dog. Their pricing structure is already a lot different and they are making it more expensive. Pretty soon Broadcom will have priced out small companies from VMware.