Intel Skylake / Kaby Lake

[lolfail9001]

What people 'feel' about overclocking has no bearing on what Intel or AMD are legally obligated to do.

And what Intel or AMD are legally obliged to do is irrelevant as well. What they do is.

Where is the statement from Intel that says "we are committed to supporting overclocking, hence we give you adjustable BCLK and soldered chips?"

Where is the statement from Intel: "**** you, we will put out every stop possible to make overclocking our chips as hard as possible for average joe". Nowhere, yet that's exactly what they are doing for last 6 years.

I don't think Intel is actively thinking about overclocking one way or the other.

They certainly do, they would not introduce BCLK governor with Skylake if they were not thinking about overclocking. They would not tie BCLK to everything in Sandy Bridge after Westmere if they were not. They would not start putting TIM over 400mm^2 dies if they were not. I am not tripping about clocking CPUs to the limit even.

One (nutty) thread over, it was revealed that new Sandy Bridge CPUs are being made with TIM, when the originals were soldered.

In Sandy Bridge line-up only 2500k/2600k/2700k were soldered iirc, the rest were using TIM from the very beginning. Low end Sandies were 100% TIM since 2011.

 

[tamz_msc]

And what Intel or AMD are legally obliged to do is irrelevant as well. What they do is.

They sell products, you buy them, and using the products under approved conditions means that you are covered under a limited warranty and the customer support that goes with it. That's about all the obligation that they have towards you.

Where is the statement from Intel: "**** you, we will put out every stop possible to make overclocking our chips as hard as possible for average joe". Nowhere, yet that's exactly what they are doing for last 6 years.

LOL what? Since when did a company put out in public a statement with negative connotations like that? That's like saying that Microsoft should say it officially that "we don't want you to stop Windows 10 telemetry, hence we made GPedit useless".

Establishing intent is what matters from a legal perspective.

Nobody admits to things like that in public.

 

[beginner99]

The TIM thing really sucks. And further thinking about it even the 7820X doesn't offer that great value. What pisses me off is the stupid Intel market segmentation.
"OK, we lower our prices a bit for 8-core but in turn you get less PCIE-lanes and TIM". Feels like a one-step forward, 2 steps back scenario.

Who actually cares about AVX-512? if you need that you probably need to code for it and at that point might be better to use OpenCL or CUDA and go GPGPU.

Given this AMD here has a great change to completely invalidate what Intel just released. Keep it soldered, simple and price competitively. The 8-core variant can be sold at 1800x* price or even a bit cheaper but let's just say $420. So $30 more than 7800x but 2 more cores and more PCIe lanes. Then a 12-core variant.goes against the 7820x. $650. You get 50% more cores (albeit lower turbo) and more PCIe lanes for $50 more. And the 16-core part vs 7900x at $1200. So $200 more for 62.5% more cores, more PCIe lanes and probably higher base frequency.

* The 1800x will probably just be priced out intentionally as all the best dies will go to HEDT anyway.

 

[wildhorse2k]

What problems are you referring to? You can already use RAM up to 4000 Mhz. I'm not aware of any other issues that aren't simply design choices to reflect the low prices. A family member has had Ryzen since release day and has had zero stability/reliability issues with both CPU and RAM overclocked. I have been doing dozens of hours of research on Ryzen since release and *almost* bought one until Skylake-X got bumped up to June. If I had to wait until August I would have still gone Ryzen. This is assuming benchmarks and OC potential are as expected of course. I think Ryzen announced one month too early, but it was fully functional/stable from day one and AMD did a pretty phenomenal job getting updates out in a timely manner for OCers and enthusiasts.

After AGESA 1.0.0.6, 4000 Mhz is basically only a BIOS option. 3466 seems to be the highest stable frequency for most people, requiring some fiddling with settings. It depends on CPU quality and motherboard a lot. Nothing stellar. RAM compatibility got greatly improved. But it shows what an early release Ryzen was and buyers don't complain at all. Like people who got scammed denying it.

Notice how for X399 no motherboard manufacturer reported the top supported frequency. For X299, ASRock does it, up to 4400.

Few months ago Ryzen was my favorite choice. But after googling for "ryzen stability", "ryzen bug" and others you can find out how many problems people have. I was not willing to pay for such a product regardless of good performance/cost ratio. Blessed are those who are unaware.

 

[tamz_msc]

After AGESA 1.0.0.6, 4000 Mhz is basically only a BIOS option. 3466 seems to be the highest stable frequency for most people, requiring some fiddling with settings. It depends on CPU quality and motherboard a lot. Nothing stellar. RAM compatibility got greatly improved. But it shows what an early release Ryzen was and buyers don't complain at all. Like people who got scammed denying it.

Notice how for X399 no motherboard manufacturer reported the top supported frequency. For X299, ASRock does it, up to 4400.

Few months ago Ryzen was my favorite choice. But after googling for "ryzen stability", "ryzen bug" and others you can find out how many problems people have. I was not willing to pay for such a product regardless of good performance/cost ratio. Blessed are those who are unaware.

Literally any memory speed outside of JEDEC qualifications is a BIOS option. People are now reporting dual-rank at 3200MHz/single rank 3200MHz and all 4 DIMMs with 1.0.0.6. B350 and X370 both have OC memory support, depending on the manufacturer - on Intel only Z-series motherboards allow XMP, and even 4GHz+ is not guaranteed without the highest-end motherboards like Asus Maximus.

There are literally dozens of people on these forums who have positive things to say about how the memory issue was handled by AMD over the past three months.

Seriously either you have no clue what you're talking about or your bias is undeniable.

 

[krumme]

The TIM thing really sucks. And further thinking about it even the 7820X doesn't offer that great value. What pisses me off is the stupid Intel market segmentation.
"OK, we lower our prices a bit for 8-core but in turn you get less PCIE-lanes and TIM". Feels like a one-step forward, 2 steps back scenario.

Who actually cares about AVX-512? if you need that you probably need to code for it and at that point might be better to use OpenCL or CUDA and go GPGPU.

Given this AMD here has a great change to completely invalidate what Intel just released. Keep it soldered, simple and price competitively. The 8-core variant can be sold at 1800x* price or even a bit cheaper but let's just say $420. So $30 more than 7800x but 2 more cores and more PCIe lanes. Then a 12-core variant.goes against the 7820x. $650. You get 50% more cores (albeit lower turbo) and more PCIe lanes for $50 more. And the 16-core part vs 7900x at $1200. So $200 more for 62.5% more cores, more PCIe lanes and probably higher base frequency.

* The 1800x will probably just be priced out intentionally as all the best dies will go to HEDT anyway.

Intel probably made this decision long ago. If we look at bwe 8c 180-200w powerusage at 4.3, it seems to me this could be a serious issue. Its going to be thermally limited and thats just a shame.
Today i think they would choose differently. What they have going for them outside if avx2 is a tiny bit of ipc and then as the most important part - fmax. Its the fmax that keeps this line strong and give it its performance advantage and then they gimp on it.
They have an extremely strong packaging technology and invest enormous sums here that rivals Amd entire r&d budget and perhaps this move was a way to capitalise on that by saving a $ per cpu?. Greedy ..... shooting themselves in the foot but thats what monopolies also do. And perhaps its a good thing.

Look at tr sp3r2. Its as a lunch box that will dissipate heat like crazy. Put a d15 in it and you are fine for even 16c on 4ghz. I think the contrast will be big. But its very interesting how it ends up. They might have solved it. Who knows.

 

[moonbogg]

I'd just about bet money that Intel is going to come out with an identical line of chips with soldered heat spreaders, charge more for them and give them some fancy branding for overclockers. TIM based chips are good enough to compete with threadripper, but if you actually want your money's worth, you will have to, um, spend more of it for another OCer's special edition chip. I swear to god I expect this to happen.

 

[TheF34RChannel]

I'd just about bet money that Intel is going to come out with an identical line of chips with soldered heat spreaders, charge more for them and give them some fancy branding for overclockers. TIM based chips are good enough to compete with threadripper, but if you actually want your money's worth, you will have to, um, spend more of it for another OCer's special edition chip. I swear to god I expect this to happen.

Nah thinking they might go back to solder for the next line but not this one. Let's see how these pan out first.

 

[krumme]

I'd just about bet money that Intel is going to come out with an identical line of chips with soldered heat spreaders, charge more for them and give them some fancy branding for overclockers. TIM based chips are good enough to compete with threadripper, but if you actually want your money's worth, you will have to, um, spend more of it for another OCer's special edition chip. I swear to god I expect this to happen.

Yeaa its a creative idea. They might come up with something similar. They tend to be pretty good at handling such challenges. They might also use oc mb bundles deals and add their hd cache memory.

If threadripper is scaling like near 2x as shown today its going to be pretty darn hard even for their 18c in mt task.
Take eg the 7900x. Add aprox 10% for mt perf as stated by Intel vs the 6950 and compare it to a 90% uplift of a 1800x in cb? Gaaa for stock;
It just doesnt work
Its a slaughter. The 16c tr might even end up faster that the 18c in many loads like cb mt.
Thats why the oc for fmax is so damn important. So yeaa perhaps you are right. They surely have the tech to do it right now.

 

[Zucker2k]

Yeaa its a creative idea. They might come up with something similar. They tend to be pretty good at handling such challenges. They might also use oc mb bundles deals and add their hd cache memory.

If threadripper is scaling like near 2x as shown today its going to be pretty darn hard even for their 18c in mt task.
Take eg the 7900x. Add aprox 10% for mt perf as stated by Intel vs the 6950 and compare it to a 90% uplift of a 1800x in cb? Gaaa for stock;
It just doesnt work
Its a slaughter. The 16c tr might even end up faster that the 18c in many loads like cb mt.
Thats why the oc for fmax is so damn important. So yeaa perhaps you are right. They surely have the tech to do it right now.

What's the setup and settings of the Threadripper machine AMD demo'ed today? Any idea?

So how could you draw all these conclusions, bro?

 

[krumme]

What's the setup and settings of the Threadripper machine AMD demo'ed today? Any idea?

So how could you draw all these conclusions, bro?

Bro. What if top end retail 16c tr is stock 3.4/3.5 with near perfect scaling in cb? Is that a problem?
Tell me; wouldnt it be great?

 

[tamz_msc]

What's the setup and settings of the Threadripper machine AMD demo'ed today? Any idea?

So how could you draw all these conclusions, bro?

Tom's Hardware reported 13.04 seconds for the Blender demo. If you use the 'correct' version of the demo which the 3.4GHz constant 8 core Ryzen took 26 seconds to complete, then we get perfect 2x scaling which means that clock speeds should be similar.

 

[itsmydamnation]

What's the setup and settings of the Threadripper machine AMD demo'ed today? Any idea?

So how could you draw all these conclusions, bro?

CB doesn't care much for memory clock/fabric clock, so its all down to clock speed. According to the stilt GMI is giving 42GB/s in each direction between each Zeppelin chip( @ 2666mem clock, so 50GB/s @ 3200) which means there is basically no throughput bottleneck between chips on the package, so the question is how much added latency.

But the thing is , all these workstation multi core workloads dont really care that much about the extra latency because SMT+ ILP provides a big enough oooE window for it not to matter.

 

[Zucker2k]

Bro. What if top end retail 16c tr is stock 3.4/3.5 with near perfect scaling in cb? Is that a problem?
Tell me; wouldnt it be great?

It'll be awesome, really. But 1800x is at 3.6ghz clocks at 8 cores and 95 watts. I doubt the 16 core Threadripper will be clocked near that, hence my question about the setup settings eg. cooling and clocks. Maybe XFR at work?
Also, Blender is just ONE among many apps that Threadripper will be running when released so I don't see how it's all of a sudden become the standard measure of ipc.

 

[lolfail9001]

LOL what? Since when did a company put out in public a statement with negative connotations like that?

I am literally describing what they were doing. I do not care about lies their PR comes up with, i am looking at their actual actions that go without commentary from PR. If you fail to see that, then i am disappointed, that's about it.

Establishing intent is what matters from a legal perspective.

Nobody admits to things like that in public.

Did you really think i was pushing for lawsuit here ? Hahahahaha, no. It just fuels my dislike of Intel, little else, i do prefer to vote with wallet after all.

Tom's Hardware reported 13.04 seconds for the Blender demo. If you use the 'correct' version of the demo which the 3.4GHz constant 8 core Ryzen took 26 seconds to complete, then we get perfect 2x scaling which means that clock speeds should be similar.

Except that Blender does not have linear core scaling anywhere, so it means that they most likely ran that demo at different settings for the sake of showing off faster completion.

 

[krumme]

CB doesn't care much for memory clock/fabric clock, so its all down to clock speed. According to the stilt GMI is giving 42GB/s in each direction between each Zeppelin chip( @ 2666mem clock, so 50GB/s @ 3200) which means there is basically no throughput bottleneck between chips on the package, so the question is how much added latency.

But the thing is , all these workstation multi core workloads dont really care that much about the extra latency because SMT+ ILP provides a big enough oooE window for it not to matter.

Is this Keller guy worth his pay or what? Lol
I think it will be as you described and it just shows the power of doing your own thinking. Its conceptually elegant, simple amd extremely costeffective. This is integrated business and engineering thinking at its finest. Granted it demands the fabric tech but compared to the other obstacles you remove by developing it and doing the right cut for diesize its an easy way.

Its easy to recognise when you see it and easy to know in hindsight but the art is doing it forward.

 

[lolfail9001]

Is this Keller guy worth his pay or what? Lol
I think it will be as you described and it just shows the power of doing your own thinking. Its conceptually elegant, simple amd extremely costeffective. This is integrated business and engineering thinking at its finest. Granted it demands the fabric tech but compared to the other obstacles you remove by developing it and doing the right cut for diesize its an easy way.

Its easy to recognise when you see it and easy to know in hindsight but the art is doing it forward.

Yet AMD dropped it's margin forecast for Q2. That elegance apparently does not bring any money.

 

[itsmydamnation]

Yet AMD dropped it's margin forecast for Q2. That elegance apparently does not bring any money.

Sigh..... your so clever.... funny how no chips are on the market that use GMI in Q2. We also dont see the price for Ryzen dropping in Q2 either. For all you know AMD could have been asked to ship more console chips then they expected, there is lower overall margin right there........

Taking a single number in isolation is dumb.........

 

[frozentundra123456]

I agree solder would be a better solution. OTOH, according to silicon lottery, more than half of 7700k chips can reach 5 ghz, so even with six or eight cores, I don't think the use of TIM rules out 4.5 or higher overclock.

 

[krumme]

It'll be awesome, really. But 1800x is at 3.6ghz clocks at 8 cores and 95 watts. I doubt the 16 core Threadripper will be clocked near that, hence my question about the setup settings eg. cooling and clocks. Maybe XFR at work?
Also, Blender is just ONE among many apps that Threadripper will be running when released so I don't see how it's all of a sudden become the standard measure of ipc.

I dont think its a meassure for ipc but as said The Stilt have shown the bandwith and its plenty and we know most of the typical workloads for such type of high core counts hides the latency. Its not dx11 games.
I think they will use the 180w tdp we know the socket is build for. Why not. Its 2x90w. And i think that gives us 3.4/3.5 with a slightly new rev of process plus the ability to bin using the same die all over. Imo it looks elegant and extremely effective from a business perspective.
What is needed to know and what i think is the important unknown that might be a big issue - is motherboard cost. That sp3r2 is serious business and probably there is a serious price to pay. Like serious even for people buying 1000 usd cpu. If its eg 450 vs 250 for sklx mb you can just add 200 usd to the cost. I find cost important...

 

[JoeRambo]

Anyone else noticed that Skylake-X has L2 cache that has same parameters as legendary Athlon64/Opteron stuff. 16-way associativity and 1MB sized.

 

[lolfail9001]

Sigh..... your so clever....

You're.

Anyways, my point is for you to look at title of the thread. Circlejerk over GMI links on CPU that according to you does not exist, let alone related to CPUs this thread is about, is definition of offtopic here.

 

[Sweepr]

If threadripper is scaling like near 2x as shown today its going to be pretty darn hard even for their 18c in mt task.

Beating 16C Threadripper (MCM-based, 4 CCXs) that uses a different TR4 socket compared to Naples will be a walk in the park for a competitively clocked 18C Core i9-7980X in almost every scenario, given that a Broadwell-E CPU with the same core count is faster overall at lower clock speeds. As for the other part you ignore, which is gaming, it will be slaugher considering Broadwell-E already leads by 20% in CPU limited testing (per clock, with HT off for both). Not everyone is buying HEDT only for productivity.

 

[krumme]

You're.

Anyways, my point is for you to look at title of the thread. Circlejerk over GMI links on CPU that according to you does not exist, let alone related to CPUs this thread is about, is definition of offtopic here.

Is it?
They adress the same market. They adress the same workload. They cost aprox the same. They come to market at the same time. The same people is using it for the same.
For most people this is the definition of same.
Commenting on one cpu with a certain label and tech gives no sense if it isnt compared to the other. Its the differences that constitutes the product. That way we can discuss technology.
Eg. Core i5 talk was therefore highly relevant when dicussing 8350 and so was eg. 7700k gaming performance in ryzen thread.
It was no problem back then except for a few. We can handle it now. Consistency.

 

[DrMrLordX]

Anyone else noticed that Skylake-X has L2 cache that has same parameters as legendary Athlon64/Opteron stuff. 16-way associativity and 1MB sized.

I hadn't noticed the associativity at all, but interesting now that you mention it.