Discussion Intel Nova Lake in H2-2026: Discussion Threads

[Thunder 57]

That's right, R&D leads to the future

Certainly a better way to keep the business moving forward than spending tons of money on stock buybacks.

 

[poke01]

Now that Zen6 thread has clock bets, let’s start nova lake one.

I estimate 6.1GHz on the Ultra 9 SKU.

 

[511]

6.2-6.4 GHz I am expecting similarish clock between Zen6/NVL also finally someone will be dethroning 14900KS for the highest clock.

 

[DavidC1]

6.2-6.4 GHz I am expecting similarish clock between Zen6/NVL also finally someone will be dethroning 14900KS for the highest clock.

So we're expecting the P core in Novalake to be perhaps even less than the 10% per clock gain?

I think the clocks are optimistic for both AMD and Intel. Based on earlier rumors they were talking about 8-9% for both the cores in Pantherlake and in Novalake. This new rumor(assuming same Hz) would suggest that there's no core gain in either Pantherlake nor Novalake, or the gains on each are a measily 4-5%.

The more optimistic point of view is that it's 8-9% on both Panther and Nova and Novalake clocks at 5.3GHz.

 

[Thunder 57]

6.2-6.4 GHz I am expecting similarish clock between Zen6/NVL also finally someone will be dethroning 14900KS for the highest clock.

Isn't Nova Lake using N2 and 18A? What do you think clocks will be on either?

 

[511]

So we're expecting the P core in Novalake to be perhaps even less than the 10% per clock gain?

I don't know tbh they shipped the early silicon it's entirely possible that either the clock are to low remember Intel takes quite a time to determine the clock speed it is possible that 11% increase came from IPC alone and the clock regressed in early samples only time will tell.
The P core is not in good state they have lost quite the talent in their Haifa design team.

I think the clocks are optimistic for both AMD and Intel. Based on earlier rumors they were talking about 8-9% for both the cores in Pantherlake and in Novalake. This new rumor(assuming same Hz) would suggest that there's no core gain in either Pantherlake nor Novalake, or the gains on each are a measily 4-5%.

The more optimistic point of view is that it's 8-9% on both Panther and Nova and Novalake clocks at 5.3GHz.

Possibly.

 

[511]

Isn't Nova Lake using N2 and 18A? What do you think clocks will be on either?

N2/18AP it's difficult to guess but I will guess N2 has better clocks than 18AP by few% due to not having to deal with BSPDN.

 

[DavidC1]

Can't we all stop for a second and correlate MHz predictions with Intel's estimated performance claims?

>1.1x ST gain

I don't know tbh they shipped the early silicon it's entirely possible that either the clock are to low remember Intel takes quite a time to determine the clock speed it is possible that 11% increase came from IPC alone and the clock regressed in early samples only time will tell.

You know this was the copium around Arrowlake leaks right? They don't take significantly underperforming samples and make a powerpoint presentation around them. >1.1x ST means they are expecting >1.1x in ST. Assuming they are somewhat sane and round down at 4 and round up at 5, >1.1x suggests 14% or less gain.

This means at 6.2GHz, it's 5-6% faster per clock compared to Lion Cove and at 6.4GHz, just 2-3% faster for a product coming 2 years later with a new process and supposedly significantly improved uncore. This is akin to getting Tejas after Prescott, or Bulldozer after Athlon X2.

Seems about right but nT uplift is very underwhelming for 2x the cores, maybe a bunch of poorly scaling benchmarks or actually keeping power to a similar PL2?

The latter.

I assume the 60% gain in MT is split between three areas, 1) Process gains 2) straight up doubled cores allowing lower power per core and better perf/W in MT 3) Uarch gains

This suggests significant gains coming on the Arctic Wolf side. If it's 25% faster than Skymont, and keeps the 4.6GHz speed, the 6.4GHz P core would be just 12% faster. I guess the silver lining for me is that it would prove E core hybrid is useless guys wrong, but in a very counterproductive way.

Personally I'm expecting 30% in Integer with further gains in FP with AVX2 code, meaning the P core would be single digit % faster in absolute performance.

 

[511]

Can't we all stop for a second and correlate MHz predictions with Intel's estimated performance claims?

>1.1x ST gain

You know this was the copium around Arrowlake leaks right? They don't take significantly underperforming samples and make a powerpoint presentation around them. >1.1x ST means they are expecting >1.1x in ST. Assuming they are somewhat sane and round down at 4 and round up at 5, >1.1x suggests 14% or less gain.

1.1x> ST means 1.2X as well last time it was exactly 1.05x.

This means at 6.2GHz, it's 5-6% faster per clock compared to Lion Cove and at 6.4GHz, just 2-3% faster for a product coming 2 years later with a new process and supposedly significantly improved uncore. This is akin to getting Tejas after Prescott, or Bulldozer after Athlon X2.

Which is ridiculous at this point the P core team should be embarrassed Lion Cove fine you were migrating to Industry standard Tooling but not this time.

The latter.

I assume the 60% gain in MT is split between three areas, 1) Process gains 2) straight up doubled cores allowing lower power per core and better perf/W in MT 3) Uarch gains

This suggests significant gains coming on the Arctic Wolf side. If it's 25% faster than Skymont, and keeps the 4.6GHz speed, the 6.4GHz P core would be just 12% faster. I guess the silver lining for me is that it would prove E core hybrid is useless guys wrong, but in a very counterproductive way.

Personally I'm expecting 30% in Integer with further gains in FP with AVX2 code, meaning the P core would be single digit % faster in absolute performance.

If Arctic Wolf is such a big upgrade u would expect 1.8X-1.9X I would tune my requirements to give E core more budget.

 

[DavidC1]

If Arctic Wolf is such a big upgrade u would expect 1.8X-1.9X I would tune my requirements to give E core more budget.

No you won't. Power isn't free. Inverse square law says if you straight up increase everything, twice the perf/clock needs four times the power. So if you get 30% increase in performance with 30% extra power, you are doing extremely well.

My rough guess:
-15% from Arctic Wolf and faster P core
-20% from doubling cores at same power*
-15% from process

*Slightly off topic but the whole cube law regarding power and frequency doesn't work in modern processors for peak frequencies. At 0.5-0.7V, frequencies increase drastically while keeping voltage the same. You could increase voltage by 1.1x but increase frequency by 3x. At 0.7-0.9V, you get your typical cube law frequency gain. Above 1V, you get diminishing returns of frequency compared to voltage.

Cube law regarding peak frequency applies to Raspberry Pi snail performance computers and PC chips from 25 years ago, not the ones in 2025. The old timers in computers might want to update their memory banks.

 

[511]

No you won't. Power isn't free. Inverse square law says if you straight up increase everything, twice the perf/clock needs four times the power. So if you get 30% increase in performance with 30% extra power, you are doing extremely well.

It isn't but sharing is caring 😛

My rough guess:
-15% from Arctic Wolf and faster P core
-20% from doubling cores at same power*
-15% from process

From process it's approx 1.15*(1.18/1.12) approx 1.21.

 

[Det0x]

Now that Zen6 thread has clock bets, let’s start nova lake one.

I estimate 6.1GHz on the Ultra 9 SKU.

5.7ghz for top SKU

 

[coercitiv]

I estimate 6.1GHz on the Ultra 9 SKU.

5.7ghz for top SKU

5.8Ghz for "top" SKU followed in 6 months by the 6Ghz edition.

 

[511]

Shame on Intel if they can't even beat Raptor Lake with 3 process shrinks and 2-3 architecture improvements.

 

[poke01]

5.8Ghz for "top" SKU followed in 6 months by the 6Ghz edition.

5.7ghz for top SKU

wow, thats a basically a dead core. 4 years of clock stagnation. If Intel doesn't use Nanoflex to its advantage they have bad designers, it should be easy to surpass 6GHz day 1 of NVL-S release. No need for a special sku, this isn't 2022 anymore.

 

[adroc_thurston]

If Intel doesn't use Nanoflex to its advantage they have bad designers, it should be easy to surpass 6GHz day 1 of NVL-S release

Kind of. Remember, Intel cores are fat.
They clock appropriately, it's just PPC is not up to par.

 

[511]

Kind of. Remember, Intel cores are fat.
They clock appropriately, it's just PPC is not up to par.

And here I thought they would improve their PPC/PPA with Panther Cove but man the P core sucks

 

[DavidC1]

Why does it need to clock at a certain frequency?

Since we know ST gains are 1.1x, it would be beneficial for Intel if they can achieve that at noticeably below 5.7GHz.

wow, thats a basically a dead core. 4 years of clock stagnation.

No one cared about clock stagnation from 2006 to 2011.

It isn't but sharing is caring 😛

From process it's approx 1.15*(1.18/1.12) approx 1.21.

You can't predict that exactly. It's beyond silly when Intel/AMD claims exact numbers. 17% really? Why not 16%? Just say 15-20% like in the old days and be done with it. Or, undersell and call it 15%.

 

[msj10]

I assume the 60% gain in MT is split between three areas, 1) Process gains 2) straight up doubled cores allowing lower power per core and better perf/W in MT 3) Uarch gains

This suggests significant gains coming on the Arctic Wolf side. If it's 25% faster than Skymont, and keeps the 4.6GHz speed, the 6.4GHz P core would be just 12% faster. I guess the silver lining for me is that it would prove E core hybrid is useless guys wrong, but in a very counterproductive way.

Personally I'm expecting 30% in Integer with further gains in FP with AVX2 code, meaning the P core would be single digit % faster in absolute performance.

60% MT is a terrible look for arctic wolf. ARL already gets like 80% of its full performance at 120W so doubling core count gets you 1.6X MT at similar power. this suggest there are no gains from the new node and cores which I find hard to believe. I know performance doesn't scale linearly and mem bandwidth plays a role but still it should be higher.

 

[511]

Also please tell me Geekbench was not part of MT Average

 

[eek2121]

60% MT is a terrible look for arctic wolf. ARL already gets like 80% of its full performance at 120W so doubling core count gets you 1.6X MT at similar power. this suggest there are no gains from the new node and cores which I find hard to believe. I know performance doesn't scale linearly and mem bandwidth plays a role but still it should be higher.

Cores need power. The more cores you have, the less power you can allocate per core. Doubling cores at a fixed power means (less than) half the power per core.

 

[MS_AT]

Cores need power. The more cores you have, the less power you can allocate per core. Doubling cores at a fixed power means (less than) half the power per core.

Yes, but the power / frequency curve is not linear, so the top of the curve costs more power than it brings frequency. Therefore, lowering the frequency to peak efficiency point should provide enough power headroom to insert more cores, also operating at near peak efficiency point, what it theory should lead to effective performance increase, within the same power envelope. The problem comes when doing this slows down your critical path (Cinebench users should not worry), but they have P cores for that.

Still, it's theory, the execution remains to be seen

 

[msj10]

Cores need power. The more cores you have, the less power you can allocate per core. Doubling cores at a fixed power means (less than) half the power per core.

you got me wrong I meant 80% of the performance at 125W and 160% of the performance with double the cores at 250W which is ARL default PL2 so same power per core.

 

[poke01]

No one cared about clock stagnation from 2006 to 2011

Intel was the undisputed champion back then. They had an almost complete monopoly on CPUs from servers to laptops.

Now not so. They need a good core to bounce back and clocks are a part of that. The best scenario would be to get gains from architecture and clocks and process node.

 

[AMDK11]

On this slide it is >1.1x for ST which can be interpreted that single thread performance will increase above 10% or will be at least 10% higher.