The infamous Intel Core i9 processor... finally real?

[IllogicalGlory]

Much higher IPC and clock will do the job.

The base clocks aren't that much higher on Intel. We'll have to see what clocks it will run at under load on both sides.

 

[Hail The Brain Slug]

The base clocks aren't that much higher on Intel. We'll have to see what clocks it will run at under load on both sides.

Also the potential to overclock. The very high turbo boost clocks rumoured for the i9's would seem to imply that if you can keep it cool, 4.3-4.5+ will be possible. I doubt Whitehaven will overclock better than the 1800x, probably worse. A 500+ MHz advantage plus the IPC difference should make up for the core count difference.

 

[Arachnotronic]

Also the potential to overclock. The very high turbo boost clocks rumoured for the i9's would seem to imply that if you can keep it cool, 4.3-4.5+ will be possible. I doubt Whitehaven will overclock better than the 1800x, probably worse. A 500+ MHz advantage plus the IPC difference should make up for the core count difference.

Yes, in Intel chips, all of the cores must be validated to work at the rated Turbo Boost 2.0 speed. So if TB 2.0 frequency is 4.3GHz, then with good cooling your odds of running the chip at 4.3GHz on all cores is quite high.

 

[Hail The Brain Slug]

Yes, in Intel chips, all of the cores must be validated to work at the rated Turbo Boost 2.0 speed. So if TB 2.0 frequency is 4.3GHz, then with good cooling your odds of running the chip at 4.3GHz on all cores is quite high.

Thats precisely what I am referring to. Overclocking all cores to turbo boost max is kind of the "gimme" overclock for intel. I expect a few hundred MHz over that will be possible, as well. I would not be surprised to see 4.5 as an average overclock for the 10 and 12 cores with adequate cooling.

 

[ehume]

Does calling them i9 mean that the HEDT now come with iGPU?

 

[Topweasel]

Much higher IPC and clock will do the job.

Why do people assume it will have that much better clocks? I mean it will, but will it be enough. The process for Ryzen is as a generality designed for power efficiency and is more capped in top clocks. But Also remember that on top of it Ryzen is already an 8 core chip that is competing against a 4 core chip on clocks. It has amazing power efficiency up to about 3.6 GHz. There is good chance that AMD can keep a MCM'd 2 Zeppelin die close to if not over 3.2-3.5GHz and stay in the power envelop. Intel's highest clocked 10+ core chip is a 12 core at 3.2GHz. Sky lake at 3.7 GHz (for the server) is still an 80w CPU. So probably 3.5GHz for 2 Skylakes (that's even if they are doing MCM, and that they are doing two 6c Skylakes) to get to 140w. At worst for Intel they would be tied, at best maybe a 300MHz difference. Not like the nearly a GHz when comparing a 7700 against a 1800x. So comparison wise IPC would go down, clock advantage would go down. Though the would be closer in core count, but in systems where the core count usually means more than outright clock speed (or worse case just interchangeable).

We will see but I expect this to be a close fight. So close that Intel might want to worry more about beating the 12c Threadripper and not the 16c.

 

[blue11]

Why do people assume it will have that much better clocks? I mean it will, but will it be enough. The process for Ryzen is as a generality designed for power efficiency and is more capped in top clocks. But Also remember that on top of it Ryzen is already an 8 core chip that is competing against a 4 core chip on clocks. It has amazing power efficiency up to about 3.6 GHz. There is good chance that AMD can keep a MCM'd 2 Zeppelin die close to if not over 3.2-3.5GHz and stay in the power envelop. Intel's highest clocked 10+ core chip is a 12 core at 3.2GHz. Sky lake at 3.7 GHz (for the server) is still an 80w CPU. So probably 3.5GHz for 2 Skylakes (that's even if they are doing MCM, and that they are doing two 6c Skylakes) to get to 140w. At worst for Intel they would be tied, at best maybe a 300MHz difference. Not like the nearly a GHz when comparing a 7700 against a 1800x. So comparison wise IPC would go down, clock advantage would go down. Though the would be closer in core count, but in systems where the core count usually means more than outright clock speed (or worse case just interchangeable).

We will see but I expect this to be a close fight. So close that Intel might want to worry more about beating the 12c Threadripper and not the 16c.

If you want to play the TDP game, the target to beat is the Xeon Platinum 8180 (28 cores at 2.5 GHz turbo in 205 W). Intel HEDT CPUs have never been remotely energy efficient, nor have the base clocks ever been important. A 6800K has twice the TDP of the 7700 (non-K) and delivers far less than twice the performance. Even a 6950X is barely on par with 7700 in energy efficiency, and the 7700 is not even the most energy efficient client SKU. The only reason to ever buy an Intel HEDT CPU is if you absolutely must have both high frequency and higher core count than client.

The only clock speed anybody cares about for HEDT is OC frequency. Since the Turbo Boost 2.0 frequency has been announced as 4.3 GHz, we already know every core on Skylake-X will reach at least 4.3 GHz. Given the Turbo Boost 3.0 frequency of 4.5 GHz, we know that there is a rather high probability that all cores will reach 4.5 GHz as well. On the other hand, "Threadripper"/Naples are not new silicon but rather repackages of Ryzen, so we already know that they will not OC past 4.1 GHz. Furthermore, we also know that the MCM nature of the high core count Zen packages will have inferior scalability (CCX problems on steroids) to the unified core found in Skylake-SP, so it certainly is no stretch to infer that a 12-core Skylake-X may outperform "Threadripper" even in multi-threaded workloads, particularly those working on shared data.

16 cores is only 33% more than 12 cores. It's easy to see how that margin could be made up:
15% clock advantage * 5% IPC advantage * 10% scalability advantage => 33%

This is without considering Amdahl's law (more cores always scale worse than faster cores).

...

Of course, Intel could well price the 7920X at $3000 and make the comparison a moot point.

 

[mikk]

Why do people assume it will have that much better clocks?

The overclocking capability of Ryzen 8C is rather low, it barely goes above 4.0-4.1 Ghz with OC, it won't get better out of a sudden with a 12C or 16C version. You should expect the opposite, another clock drop because it will be much harder for a 12C or 16C version. Intel is shipping their 10C SKU with 4.3-4.5 Ghz on default, AMD can only dream of this because they can't reach this even when oced. Clock headroom is way higher on SKL-X.

 

[Topweasel]

The overclocking capability of Ryzen 8C is rather low, it barely goes above 4.0-4.1 Ghz with OC, it won't get better out of a sudden with a 12C or 16C version. You should expect the opposite, another clock drop because it will be much harder for a 12C or 16C version. Intel is shipping their 10C SKU with 4.3-4.5 Ghz on default, AMD can only dream of this because they can't reach this even when oced. Clock headroom is way higher on SKL-X.

Did you even read my post. I heavily suggested both would see reduced clocks. I just don't get why people would think a CPU with 200% more cores would clock as High. I think people are over estimating the Intel clocks.

 

[guskline]

Let's keep in mind that this is BOTH a new processor and a new chipset so the performance gains should be significant from the X99 chipset and Haswell-E/Broadwell-E processor line.

I own a wonderful 5960x at 4.4Ghz that is very fast. My new Ryzen 1800x at 4Ghz on all cores is nearly as fast but cost 1/2 of the Haswell-E.

I expect Intel to reassert their dominance. Price becomes the issue.

 

[farosis]

I just don't understand why they think all core running at 4.3G is certainly happening.

Did you even read my post. I heavily suggested both would see reduced clocks. I just don't get why people would think a CPU with 200% more cores would clock as High. I think people are over estimating the Intel clocks.

 

[Topweasel]

I just don't understand why they think all core running at 4.3G is certainly happening.

Yeah. If I had to guess, 3.5GHz. I'll leave the window open for 3.7. But it's could be as low as 3GHz for power reasons. 4GHz is going to be 200w+.

 

[farosis]

far from 4.3 G on all core so far from userBenchmark:
https://hardforum.com/threads/10-core-skylake-x-appears-on-userbenchmark.1933606/

tbh: if intel can do that within TDP budge, AMD naples will be doomed.

 

[Topweasel]

If you want to play the TDP game, the target to beat is the Xeon Platinum 8180 (28 cores at 2.5 GHz turbo in 205 W). Intel HEDT CPUs have never been remotely energy efficient, nor have the base clocks ever been important. A 6800K has twice the TDP of the 7700 (non-K) and delivers far less than twice the performance. Even a 6950X is barely on par with 7700 in energy efficiency, and the 7700 is not even the most energy efficient client SKU. The only reason to ever buy an Intel HEDT CPU is if you absolutely must have both high frequency and higher core count than client.

The only clock speed anybody cares about for HEDT is OC frequency. Since the Turbo Boost 2.0 frequency has been announced as 4.3 GHz, we already know every core on Skylake-X will reach at least 4.3 GHz. Given the Turbo Boost 3.0 frequency of 4.5 GHz, we know that there is a rather high probability that all cores will reach 4.5 GHz as well. On the other hand, "Threadripper"/Naples are not new silicon but rather repackages of Ryzen, so we already know that they will not OC past 4.1 GHz. Furthermore, we also know that the MCM nature of the high core count Zen packages will have inferior scalability (CCX problems on steroids) to the unified core found in Skylake-SP, so it certainly is no stretch to infer that a 12-core Skylake-X may outperform "Threadripper" even in multi-threaded workloads, particularly those working on shared data.

16 cores is only 33% more than 12 cores. It's easy to see how that margin could be made up:
15% clock advantage * 5% IPC advantage * 10% scalability advantage => 33%

This is without considering Amdahl's law (more cores always scale worse than faster cores).

...

Of course, Intel could well price the 7920X at $3000 and make the comparison a moot point.

4.5Ghz is going to be nearly 300w all cores. 4.2w probably closer to 270w. Those boards are going to need to be made like tanks. It's one thing to understand that each core can hit that high and having all the cores run at that speed by defualt.

Best case scenario is Intel targets like 180w. Are people really willing to pump 75% more power into a even a $1.5k CPU. The people doing that are going to be in the single digits.

I will say this. It's a really interesting platform. I can't wait to see what Intel pulls off. The next couple years are going to be so tense. This is like 99-2002 again.

 

[IndyColtsFan]

Thats precisely what I am referring to. Overclocking all cores to turbo boost max is kind of the "gimme" overclock for intel. I expect a few hundred MHz over that will be possible, as well. I would not be surprised to see 4.5 as an average overclock for the 10 and 12 cores with adequate cooling.

Yep, and this is the exact reason, IMO, that people are going to be disappointed when Intel announces pricing - they will still be able to command a premium due to the clock speed advantage, overclocking, IPC, etc.

At best, I think we will see the HEDT CPUs slide down one rung in the pricing tiers (10 core will occupy the current 8 core slot) but the pessimist in me notes that there is a large gap in the current pricing model ($1000 8 core and $1700 10 core) so I wouldn't be shocked if Intel created a $1200-$1300 tier and put the 10 core model there.

 

[Topweasel]

far from 4.3 G on all core so far from userBenchmark:
https://hardforum.com/threads/10-core-skylake-x-appears-on-userbenchmark.1933606/

tbh: if intel can do that within TDP budge, AMD naples will be doomed.

I don't see how they can do that within the budget. I wouldn't be surprised if Intel put numbers up themselves. I mean you are right if Intel can pull it off that would be amazing and will nearly kill AMD's mid range server options. But I remember Intel demoing a cherry picked 5GHz Prescott before it's release using extreme cooling, before actually launching at barely faster then I think it was Northwood.

This isn't even a demo, it's a system people can't even see running posting a number to webpage. It's the inverse of the GPU numbers around where chances are they running slower than production and have much more testers for driver qualifications causing more leaks.

 

[guskline]

I have my 5960x at 4.4 custom water cooled with an EK Evo Supremacy block.
At full stress on 8 cores I measure via Aida64 @ 235-240W. My Ryzen 1800x at 4Ghz on all 8 cores is 135-140W.

I suspect if a Skylake-X 12C is OC'd to even 4.4 it's hard to believe it would be under 250W.

However, keep in mind this will be a high end chipset/processor that will be robust.

 

[blue11]

I don't see how they can do that within the budget. I wouldn't be surprised if Intel put numbers up themselves. I mean you are right if Intel can pull it off that would be amazing and will nearly kill AMD's mid range server options,

I don't understand why you keep talking about TDP here. The X299 boards will be designed to deliver as much power as needed to reach maximum OC frequencies. The ability to deliver way more power than Intel specifies is essentially how the motherboard manufacturers compete with each other. If the chips need 1 kW to run at OC (which they won't), boards capable of delivering 1 kW will be made. Since each core is validated by Intel at the factory, reaching the turbo frequency on all cores is merely an issue of power delivery.

None of this has anything to do with servers. As I said in the previous post, for servers, the target AMD will have to beat is:

28 cores
2.5 GHz base
2.8 GHz (estimated) turbo
Unified cache
205 W TDP

 

[ehume]

TDP will be an issue as a potential user considers cooling solutions.

 

[Topweasel]

I don't understand why you keep talking about TDP here. The X299 boards will be designed to deliver as much power as needed to reach maximum OC frequencies. The ability to deliver way more power than Intel specifies is essentially how the motherboard manufacturers compete with each other. If the chips need 1 kW to run at OC (which they won't), boards capable of delivering 1 kW will be made. Since each core is validated by Intel at the factory, reaching the turbo frequency on all cores is merely an issue of power delivery.

None of this has anything to do with servers. As I said in the previous post, for servers, the target AMD will have to beat is:

28 cores
2.5 GHz base
2.8 GHz (estimated) turbo
Unified cache
205 W TDP

In what world do think TDP doesn't matter? This isn't a magical world where just because you imagine it, it can happen. The CPU will have a TDP wall in Intel's design. They aren't going to make a 300w CPU. Just isn't going to happen. I used the servers to get a bit a bit of a TDP comparison. I think they will up the power usage, but it will have a cap, best guess is 180w. To get all cores to 4.5 it will need 300w. That isn't reasonable on the mobo side the cooling side and who is going to creat a mobo for a 300w CPU.

Think of the design that has to go into the board. Think of the price of the CPU. Think of the amount of people buying this CPU. Think of the amount of extra power people are going to send to the CPU at this level at that price. It's just not reasonable. It's all academic. Nobody who cares about their machine is going to do this. What I do know is that it's default clock isn't going to be as high as people think and for those willing to blow up a 1.5k+ CPU it's power usage is going to increase incredibly with ever 100MHz. Intel intelligently moving a core or two up 4.5 is only 20% of the CPU. I just don't think a sustained 4GHz+ overclock is going to be manageable without having the motherboard and the cooling within an inch of their capabilities.

This is where AMD I think can keep up on clocks. We already see a 3.7GHz clock at 90w. We know that Ryzen just plainly will not run reliably beyond 4GHz. But it can stay closer to 4GHz with 75% of core count. 12c at even at 160w TDP could be clocked as high as 3.6 GHz.

I want to be clear I am not saying that the Intel arch can't be clocked as high as 4.5 GHz. It can But I can't see a platform or actual users in a situation that makes it possible. Just look at the KL 5GHz. Part if it is the TIM used for the heat spreader but the requirement is beyond 99% of the users out there. Now take a CPU with 3 time the cores, a weird design not quite SL but not yet KL, 5 times the price, trying to keep it within 10% of the speed of a CPU people have to butcher to get up to the speed. Just seems unreasonable to expect that kind of performance. That ain't even counting the cooling and mobo requirements.

 

[blue11]

In what world do think TDP doesn't matter? This isn't a magical world where just because you imagine it, it can happen. The CPU will have a TDP wall in Intel's design. They aren't going to make a 300w CPU. Just isn't going to happen. I used the servers to get a bit a bit of a TDP comparison. I think they will up the power usage, but it will have a cap, best guess is 180w. To get all cores to 4.5 it will need 300w. That isn't reasonable on the mobo side the cooling side and who is going to creat a mobo for a 300w CPU.

Of course the TDP doesn't matter. It doesn't matter if Intel "makes" a 300 W CPU. If the max OC frequency needs 300 W, motherboard manufacturers will produce designs that will deliver 300 W.

Think of the design that has to go into the board. Think of the price of the CPU. Think of the amount of people buying this CPU. Think of the amount of extra power people are going to send to the CPU at this level at that price. It's just not reasonable. It's all academic. Nobody who cares about their machine is going to do this. What I do know is that it's default clock isn't going to be as high as people think and for those willing to blow up a 1.5k+ CPU it's power usage is going to increase incredibly with ever 100MHz. Intel intelligently moving a core or two up 4.5 is only 20% of the CPU. I just don't think a sustained 4GHz+ overclock is going to be manageable without having the motherboard and the cooling within an inch of their capabilities.

The LGA-3647 socket was designed for Xeon Phi with TDP of 250 W. That means it can momentarily draw as much as 350 W on turbo frequency. Skylake-X will be on the smaller LGA-2066 socket, but there is no reason why such power draws can't be sustained. Haswell-E/Broadwell-E already consume over 200 W. Even the puny LGA-1151 socket is capable of delivering over 200 W for OC Skylake and Kaby Lake client processors.

This is where AMD I think can keep up on clocks. We already see a 3.7GHz clock at 90w. We know that Ryzen just plainly will not run reliably beyond 4GHz. But it can stay closer to 4GHz with 75% of core count. 12c at even at 160w TDP could be clocked as high as 3.6 GHz.

Ryzen does not run at anything near the TDP listed on the box.

I want to be clear I am not saying that the Intel arch can't be clocked as high as 4.5 GHz. It can But I can't see a platform or actual users in a situation that makes it possible. Just look at the KL 5GHz. Part if it is the TIM used for the heat spreader but the requirement is beyond 99% of the users out there. Now take a CPU with 3 time the cores, a weird design not quite SL but not yet KL, 5 times the price, trying to keep it within 10% of the speed of a CPU people have to butcher to get up to the speed. Just seems unreasonable to expect that kind of performance. That ain't even counting the cooling and mobo requirements.

You are already assuming that 16-core Zen-based CPUs will be running at their OC frequency of 4.1 GHz. Why would you not assume that the corresponding Skylake-based CPU will be running at its maximum OC frequency?

EDIT: Also, in the worst case, you could likely set a lower AVX OC frequency. Haswell-E/Broadwell-E draw far less power than their TDP when running non-AVX workloads. Zen doesn't even support AVX, let alone AVX-512.

 

[Insert_Nickname]

Zen doesn't even support AVX, let alone AVX-512.

That is just plain wrong. Zen supports AVX2. As my Ryzen will testify.

In fact, AMD already supported AVX with Bulldozer, and their own XOP.

 

[Topweasel]

Of course the TDP doesn't matter. It doesn't matter if Intel "makes" a 300 W CPU. If the max OC frequency needs 300 W, motherboard manufacturers will produce designs that will deliver 300 W.

Sure they can just snap their fingers and have motherboard manufacturers create a consumer enthusiast motherboard requiring twice the out put they currently supporting and 33% more than they have been supporting on max over clocks. All within fitting current case and power supply standards.

The LGA-3647 socket was designed for Xeon Phi with TDP of 250 W. That means it can momentarily draw as much as 350 W on turbo frequency. Skylake-X will be on the smaller LGA-2066 socket, but there is no reason why such power draws can't be sustained. Haswell-E/Broadwell-E already consume over 200 W. Even the puny LGA-1151 socket is capable of delivering over 200 W for OC Skylake and Kaby Lake client processors.

I can see this might be a bit pointless to explain but just like clocks on a CPU. The higher you get the requirements get that much more extreme. That 50% increase in Power will cost that much more to develop and implement than the 200w before it. You can't compare a consumer platform with a platform made to fit specific use cases that is tons more expensive and made to order.

Ryzen does not run at anything near the TDP listed on the box.

Point. Neither of them do. TDP is about cooling requirements and limited max power output isn't as important figuring out the cooling requirements. The point is that Ryzen can sit close to 4GHz with 8 cores (double a 7700K) at the power output of the 7700 at 4.5. A 12 core Ryzen is 50% more cores, a 12 core i9 would be 3x the cores. My point was AMD keeping the HEDT cores at the retail speeds of of their R7 lineup would be much more manageable.

You are already assuming that 16-core Zen-based CPUs will be running at their OC frequency of 4.1 GHz. Why would you not assume that the corresponding Skylake-based CPU will be running at its maximum OC frequency?

I assumed nothing. When did I say that? Like I noted was that Ryzen and it's 8 core dies sit in a really good spot efficiency wise at mid 3GHz. Intel probably as well but they clock their stuff up so high because they can. But like I have noted 90w on a 4c CPU clocked at 4.5 isn't going to carry over as well to a 12 core setup. That's my point Intel can ratchet up a single core Turbo, heat will be spread out and the power can be dispersed by clocking the other cores down. But they will clock the default clocks down. But if you think that Mobo manufacturers will be able to pull off a a consumer mobo for <$500 that will have the overhead to support a 300w sustained CPU, that will fit the ATX or even ATX-E standard, and work with current PSU's, and people can get good enough cooling for without going LN2, then sure than a 4.5GHz OCs could be attainable.

 

[mikk]

TDP will be an issue as a potential user considers cooling solutions.

TDP alone doesn't say anything about heat dissipation. If you have a big Die it might be easier to dissipate heat from the chip than a smaller Die with smaller TDP. Broadwell-E is a good example.

That is just plain wrong. Zen supports AVX2. As my Ryzen will testify.

In fact, AMD already supported AVX with Bulldozer, and their own XOP.

But only with 2x128 Bit. AVX2 performance is very poor therefore if properly supported.

 

[Insert_Nickname]

But only with 2x128 Bit. AVX2 performance is very poor therefore if properly supported.

Again wrong. Ryzen performs well enough in AVX software. Also, compared to a 4C Intel CPU, Ryzen might have half the throughput, but it has twice the hardware. Which make it a complete non-issue. HEDT is slightly different, but if you have a workload that can actually take full advantage of AVX(2), you're properbly using such a CPU in the first place.

See this thread:
https://forums.anandtech.com/threads/ryzens-halved-256bit-avx2-throughput.2501158/

Most consumer software is still stuck at SSE2, at most SSE 4.2. Intel make sure of that by disabling AVX on their Pentium/Celeron line.