Discussion Intel current and future Lakes & Rapids thread

[Hulk]

My understanding is that Zen3 and even Zen3D should still be cheaper to manufacture than a large monolithic die such as ADL-S when looking at the flagship SKUs, alas production volume vs. demand will complicate the equation quite a bit.

I don't have much insight into this and would love to know more? At the top of the stack I assume the two AMD chiplets are smaller in total area than the 8+8 ADL die? Also when there is a defect in ADL you could lose the whole die, which is less of a loss than one chiplet. But then it seems as though there might be more assembly time/cost in the chiplet approach? I know, I'm jumping the gun here but I'm curious about the economics of these approaches specifically in the case of 7nm 5950X vs 10nm 12900K production?

 

[Schmide]

The key thing here is that with Zen3D rumored to be a bit more expensive piece of tech to manufacture, how much more is AMD going to tag onto the Zen 3 tax, and are enthusiasts going to respond favorably to it? As it is, with the 12600kf rumored to be 5800x killer and listed at a paltry $261 price, things are not looking good for Zen3D pricing.

ADL-S and next line of releases are going to be a real problem for AMD. ADL-S looking this strong this early, and with RPL with 8 raptor cove cores and 16 e-cores, and all of a sudden, it's AMD that's under pressure in the "moar cores" war.

What is a Zen3 tax?

Is there a governing body receiving these funds?

Or is this just loaded language?

I don't think you understand taxes and what they are used for.

 

[Hulk]

I've been thinking a bit about the leaked 30,000+ ADL Cinebench R23 score. Assuming this is true it seems like Golden Cove would have to be scoring about 2,500, which seems astronomical.

Here are my assumptions. Coves running at 5GHz, Monts running at 4GHz.
Perfect scaling. If Golden Coves are scoring 2500, then Gracemonts would be scoring 1,250 at 4GHz, which would put them theoretically at 1,562.5 if running 5GHz.

Since Gracemont is said to be "about" Skylake, then 1562.5 at 5GHz would be "about" Skylake performance.

My problem is that a Golden Cove score of 2,500 seems astronomical because it would imply a 40+% throughput increase over Rocket Lake?

Where am I going wrong here? Other than assuming the 30,000 R23 score is legit?

 

[LightningZ71]

Perhaps, with Golden Cove being wider, it gains a lot more from SMT than previous cores?

 

[Accord99]

My problem is that a Golden Cove score of 2,500 seems astronomical because it would imply a 40+% throughput increase over Rocket Lake?

Where am I going wrong here? Other than assuming the 30,000 R23 score is legit?

The 2500 score would be from a single GC cove running two threads, taking advantage of SMT. I don't think you can configure Cinebench to run two threads on a specific core but if you could, a Rocket Lake core could probably do 1900-2000.

A single thread in GC would be probably be in the range 2000-2100, a still impressive score but not out of the realm of possibility give a 5.3GHz clock speed and other advertised improvements over the previous Cove cores.

 

[Hulk]

The 2500 score would be from a single GC cove running two threads, taking advantage of SMT. I don't think you can configure Cinebench to run two threads on a specific core but if you could, a Rocket Lake core could probably do 1900-2000.

A single thread in GC would be probably be in the range 2000-2100, a still impressive score but not out of the realm of possibility give a 5.3GHz clock speed and other advertised improvements over the previous Cove cores.

Yes I understand. But my point is that assuming the 30,000 R23 score is legit and the Coves are running 5GHz all core and the Monts 4GHz all core, then you need Golden Cove to score 2500 in order for Gracemont to score in the Skylake range.

2500x8=20,000 for the Golden Coves at 5GHz

1250x8=10,000 for the Gracemonts at 4GHz

Total R23 MT score = 30,000

Gracemonts running theoretical 5GHz would mean R23 ST score of (5/4) x 1250 = 1562.5, which is right in the Skylake range.

If Golden Cove scores lower than 2500 then Gracemont IPC would be higher than Skylake, which is possible, Ian was figuring +8% for Gracemont over Skylake now that I think about it. Even if this is the case it would still put Golden Cove at +40% IPC in R23 over Rocket Lake, which seems nuts?

Either that 30,000 score is BS or those Cove (and Monts) are doing the do in Cinebench R23.

 

[Joe NYC]

I'd say if leaks are correct and 12600KF is really ~$250, that will force AMD to stop milking customers with "proper" 8C and come up with either better product in the form of Z3D or to adjust pricing accordingly. IF 8+8 match 16 AMD cores in MT then 6+4 will definately beat 8C in MT while also providing better ST performance as well. $400 5800X will need adjustments and AMD 6C will be hurt even more so.

Good that AMD still has ton of lots for SKUs to release at whatever prices they want. What about 5700X with 100mhz lower boost clock on paper for 330$ ?

5800x is already on sale for $349, at Micro Center. So this may just become the new MSRP, while 5800 X3D could take place at the original price of $449 (as one possibility)

 

[Bouowmx]

I've been thinking a bit about the leaked 30,000+ ADL Cinebench R23 score. Assuming this is true it seems like Golden Cove would have to be scoring about 2,500, which seems astronomical.

Here are my assumptions. Coves running at 5GHz, Monts running at 4GHz.
Perfect scaling. If Golden Coves are scoring 2500, then Gracemonts would be scoring 1,250 at 4GHz, which would put them theoretically at 1,562.5 if running 5GHz.

Since Gracemont is said to be "about" Skylake, then 1562.5 at 5GHz would be "about" Skylake performance.

My problem is that a Golden Cove score of 2,500 seems astronomical because it would imply a 40+% throughput increase over Rocket Lake?

Where am I going wrong here? Other than assuming the 30,000 R23 score is legit?

Going by the Cinebench R20 (different version) scoring model of Alder Lake by OneRaichu from a bit back, the i9-12900K gets (in R20) 72% score with P cores, and 28% with E cores. Translating to R23, 30500 total score means 22000 by P cores, and 8500 by E cores.
22000 / 8 = 2750, but this includes HT
2750 / 1.25 = 2200

As for how, it may be the new FADD included with Golden Cove.

 

[Hulk]

I'm assuming if you run Cinebench in ST mode you get the same score regardless if HT is on or off?

 

[Hulk]

Going by the Cinebench R20 (different version) scoring model of Alder Lake by OneRaichu from a bit back, the i9-12900K gets (in R20) 72% score with P cores, and 28% with E cores. Translating to R23, 30500 total score means 22000 by P cores, and 8500 by E cores.
22000 / 8 = 2750, but this includes HT
2750 / 1.25 = 2200

As for how, it may be the new FADD included with Golden Cove.

Okay this explains it. I forgot about SMT during the MT test.

Assuming 2000 score for Golden Cove and +25% for HT, then the Coves "earn" 20,000 in R23 MT, with Gracemont picking up the remaining 10,000 among the 8 cores.

A 2000 R23 ST score seems quite reasonable given the architectural improvements.

Given the prevalence of thought around here that most software doesn't utilize more than 8 cores very well those 8 Coves are looking to be strong performance, with the Skylake level Monts taking care of housekeeping, or helping with compute with apps that scale to higher thread counts. We have a real horse race here and we are the winners either way!

 

[deasd]

Assuming 2000 score for Golden Cove and +25% for HT, then the Coves "earn" 20,000 in R23 MT, with Gracemont picking up the remaining 10,000 among the 8 cores.

What you said already pushed by Abwx at #13195. The small cores offer 33% of total score. And I saw somewhere that 12900k in geekbench has singlethread score of little above 2000 after tweak. but I forget where did I saw the leak. Everything is showing about 10% IPC advantage against last geneneration of both intel/amd

 

[coercitiv]

I don't have much insight into this and would love to know more?

You already enumerated some of the main factors, but as I said we'll have plenty of time to talk about this in detail after the ADL-S launch.

 

[Det0x]

Okay this explains it. I forgot about SMT during the MT test.

Assuming 2000 score for Golden Cove and +25% for HT, then the Coves "earn" 20,000 in R23 MT, with Gracemont picking up the remaining 10,000 among the 8 cores.

A 2000 R23 ST score seems quite reasonable given the architectural improvements.

Given the prevalence of thought around here that most software doesn't utilize more than 8 cores very well those 8 Coves are looking to be strong performance, with the Skylake level Monts taking care of housekeeping, or helping with compute with apps that scale to higher thread counts. We have a real horse race here and we are the winners either way!

Some new Cinebench ST info here:

There is also a leak from HXL, who posted mysterious screenshots from Cinebench R20 and R23. Just two days ago we had a leak featuring Core i9-12900K with 30K+ multi-core in Cinebench R23, but so far no one has posted proof that the CPU would also score 2000 points in the single-core test.

Should this screenshot indeed be demonstrating the real performance of the 12900K, then this CPU would offer the highest single-core performance out of all known desktop CPUs. In comparison to AMD Ryzen 9 5950X at stock settings, we are looking at a 21% performance advantage of Alder Lake over AMD Zen3 and Intel Rocket Lake architectures, or alternatively 40%+ over Core i9-10900K (Comet Lake-S).

Since Cinebench R23 software is relatively new, it has not been used in Ryzen 5000 series reviews when it came out. However, ComputerBase has a community-based ranking featuring nearly all popular processors:

.. But there is also a rumor going around that says the reason for Alderlakes very good Cinebench performance is because the program almost fits perfectly in Golden Coves 1.25mb L2 cache per core, and the overall performance upgrade over rocketlake in anything that is latency-depended (games) is much lower..
(Almost 100ns memory latency with even overclocked ddr5 in gear4 mode don't sound too good)

 

[Zucker2k]

Another 12900k vs 5950x comparison. Sisoftware.

https://twitter.com/x/status/1441340651782565898

Edit: The linked 12900k in the twitter post seems borked. Here's a better link: https://ranker.sisoftware.co.uk/sho...d5edd8e0d4e4d7f183be8ea8cda895a583f0cdf5&l=en

 

[inf64]

Based on these results, Golden Cove should have around 11% higher IPC than Zen3 and ~19% higher IPC than Cypress Cove ( Zen3 is around 7.5% faster per clock than Cypress Cove, on average). Not great, not terrible. Bad news for intel is that Zen3D is coming soon which will be a massive boost in games. Competition is great!

 

[lobz]

Based on these results, Golden Cove should have around 11% higher IPC than Zen3 and ~19% higher IPC than Cypress Cove ( Zen3 is around 7.5% faster per clock than Cypress Cove, on average). Not great, not terrible. Bad news for intel is that Zen3D is coming soon which will be a massive boost in games. Competition is great!

In SOME games. AMD can decide to acknowledge this and price accordingly, or just "stay strong" till Zen 4. Hoping for the first but expecting the latter.

 

[eek2121]

It is amazing that Intel was able to pull this off. I feel like some won’t ever appreciate the fact that Intel is bringing some serious innovation to the desktop and laptop markets.

I look forward to seeing what mobile ADL-S can do in particular.

I feel like ADL-S is Intel’s “Zen” moment. AMD is going to have to really step it up to compete, RPL-S is rumored to add another 20% to single core performance and more than 50% to multicore performance.

I look forward to seeing AMD’s response. Competition is grand!

 

[uzzi38]

RPL-S is rumored to add another 20% to single core performance

No it's not. RPL-S is a much smaller improvement to single-thread performance than Alder Lake is.

 

[dullard]

No it's not. RPL-S is a much smaller improvement to single-thread performance than Alder Lake is.

Raptor Lake has small increases in IPC (I assume due to more L2 cache), but also it is rumored to have: 3.8% faster clocks, higher allowable PL2 power in performance mode, and faster memory support. While the IPC changes might not be much, the combination of all of the above might get close to 20% gains in the right scenarios.

 

[Hulk]

It is amazing that Intel was able to pull this off. I feel like some won’t ever appreciate the fact that Intel is bringing some serious innovation to the desktop and laptop markets.

I agree. Not long ago Intel was looking pretty dead. Let's face it the Rocket Lake backport was pretty underwhelming and Zen 3 was a smashing success. AMD brought out a fantastic core on a 7nm process in 8, 12, and 16 core designs. A pretty darn perfect and timely product.

Now it looks as though Intel will be producing a hybrid design up to 16 cores on the ESF 10nm process. At first the hybrid design seemed kind of strange but when you consider the importance of single threaded performance and that Intel will be on top or on equal footing for a while is a pretty nice comeback. In addition, having 8 additional Skylake level efficiency cores to take care of housekeeping and help with highly multithreaded apps, as well as adding power efficiency to mobile designs is a smart move.

Now I'm thinking AMD will still be more power efficient on the desktop and there will be applications that will be suited for 16 equally "big" cores, but the point is I think now we're looking at two great designs that will trade blows and we can pick the better choice depending on our personal workloads. Before ADL it was an easy choice, Zen 3 was better in pretty much every scenario if you were looking for leading edge desktop performance.

One point of contention I think I have with some here is that I don't believe disabling the small cores will in any case increase performance except for the niche case where that somehow enables AVX512 in the Coves. But we shall soon see. I don't think I've been looking forward to a CPU release this much in quite a few years.

 

[inf64]

Raptor Lake has small increases in IPC (I assume due to more L2 cache), but also it is rumored to have: 3.8% faster clocks, higher allowable PL2 power in performance mode, and faster memory support. While the IPC changes might not be much, the combination of all of the above might get close to 20% gains in the right scenarios.

If they get 10% cumulative ST performance gain with Raptor lake I'd call it a massive success. Still, that part will go against Zen4 3D and the parts based on the new Zen core will have a clear IPC advantage (based on the rumors).

 

[uzzi38]

Raptor Lake has small increases in IPC (I assume due to more L2 cache), but also it is rumored to have: 3.8% faster clocks, higher allowable PL2 power in performance mode, and faster memory support. While the IPC changes might not be much, the combination of all of the above might get close to 20% gains in the right scenarios.

Higher PL2 doesn't affect ST performance, and neither does better memory support unless the testing is done with tuned memory, which isn't a sensible way of testing anyway.

 

[JoeRambo]

Higher PL2 doesn't affect ST performance, and neither does better memory support unless the testing is done with tuned memory, which isn't a sensible way of testing anyway.

Depends on what the baseline "stock" testing is. If baseline is DDR5 4800 and they move support to say DDR5 6400, that is going to increase performance in a way Skylake scaled from 2133 to 2933 official support.

 

[epsilon84]

Based on these results, Golden Cove should have around 11% higher IPC than Zen3 and ~19% higher IPC than Cypress Cove ( Zen3 is around 7.5% faster per clock than Cypress Cove, on average). Not great, not terrible. Bad news for intel is that Zen3D is coming soon which will be a massive boost in games. Competition is great!

IPC on its own isn't useful without taking into account the operating frequency. If the rumoured 5.3GHz turbo boost is accurate, that is a rather significant ~20% higher than the 4.4GHz all core turbo on a competing 5900X, for example.

If Zen3D can achieve slightly higher clocks AND avg +15% in gaming then we definitely have some serious competition on our hands. If its stuck at mid 4GHz territory then I'm not convinced that will be enough to overcome ADL-S, but we'll see soon enough!

Agreed that this renewed competition is great! Really looking forward to seeing how this all pans out in the coming months. Also looking forward to finally retiring my 8700K that has gone through 3 GPU upgrades already!

 

[Zucker2k]

My prediction, with clock regression of Milan-X, is that Zen 3D may end up like Rocketlake vs Cometlake, but on the reverse. Outside of memory sensitive scenarios, Zen 3D could actually end up less performant compared to Zen 3. Yes, 7nm is mature now but that humongous cache is going to demand some juice so we'll see with the thermals as well.