What Intel tried to do is create a node that would give them a 3 year advantage over the competition that while "10nm" has aspects where important to CPU products that are a lot smaller - competing with 7nm and even beating it in some areas but overall it isn't comparable in every regard to a 7nm node. Somewhat ironic that it looks like not only are they going to not have a 3 year march on the competition at this rate they are going to be 3 years behind before they've even [truly] got to 10nm let alone their 7nm process.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Another day, another broken Intel process node: Intel announces 7nm will be delayed
- Thread starter Grim5
- Start date
What Intel tried to do is create a node that would give them a 3 year advantage over the competition that while "10nm" has aspects where important to CPU products that are a lot smaller - competing with 7nm and even beating it in some areas but overall it isn't comparable in every regard to a 7nm node. Somewhat ironic that it looks like not only are they going to not have a 3 year march on the competition at this rate they are going to be 3 years behind before they've even [truly] got to 10nm let alone their 7nm process.
Soldato
I'd say it was quite the opposite. AMD had a long way to catch up coming from Bulldozer and its derivatives, it's only by virtue of Intel being so far ahead and holding back advancements they could've dropped years ago have they managed to stay in the position they're in. The gap between the two companies really was that large.
But Comet Lake is the first proper sign that Intel haven't got anything left in the tank, and that 10 core is a one-off (Rocket Lake caps at 8 cores by all accounts). They're wringing the neck of 14nm, clocking the **** off it by pumping a bazillion watts through it, and a lot of the gains the Rocket Lake design has are practically nullified by it being backported to 14nm.
Zen 2 is the point where AMD closed the gap. Zen 3 is the point where they finally move past Intel. Zen 4 is where they maintain their lead.
Associate
- Joined
- 28 Sep 2018
- Posts
- 2,259
Ring bus doesn’t scale well past 8 cores. Even the 10core has a worse core to core latency than a 9900k. Anand posted the technical details and you can prove this by having the same memory overlock on a 9900k and a 10700k/10900k and the 9900k will have better latency. A few of us tested this.
that’s not a 14nm issue. That’s an architectural issue.
Not really, they had a massive lead on AMD which is pretty much down to nothing now, and an even bigger lead on ARM/Apple which has closed down significantly.
They were more or less unrivalled, now they have two or more rivals knocking at the door and at least one (AMD) surpassing them.
Soldato
Precisely, Intel have nothing left in the tank. They could've moved past ring bus a long time ago, but they didn't. At the very least they could've given us 8 cores a long time ago, but they didn't. And now they're stuck at the very end of what they can do, hampered by ring bus and 14nm.
Soldato
In non-gaming applications AMD tends to be far closer than in games,so it's not really ST performance which is the main factor at play here IMHO.
Much of the AMD deficit in gaming with Zen2 is probably down to Zen2 having higher inter-core latency,compared to Intel(CCX-CCX latency),and lots of game engines(and maybe even Nvidia drivers) not being fully "Zen aware",so AMD has to brute force things. With the consoles being Zen2,expect more game engines,etc to be coded to work around any bottlenecks in the design,and that will get Zen closer to whatever Intel has.
This is the problem Intel has,so many games will be now developed with Zen in mind,so why the one area they seem to still be relevant for enthusiasts in,will start to ebb away.
Much of the AMD deficit in gaming with Zen2 is probably down to Zen2 having higher inter-core latency,compared to Intel(CCX-CCX latency),and lots of game engines(and maybe even Nvidia drivers) not being fully "Zen aware",so AMD has to brute force things. With the consoles being Zen2,expect more game engines,etc to be coded to work around any bottlenecks in the design,and that will get Zen closer to whatever Intel has.
This is the problem Intel has,so many games will be now developed with Zen in mind,so why the one area they seem to still be relevant for enthusiasts in,will start to ebb away.
Associate
I think if Intel had been pushed earlier than now then these problems would have been discovered and worked through. They were resting on their 14nm node thinking they could take their time and work to their timeframe. Ryzen changed all that, IPC wise at least as powerful clock for clock. Intel got caught flatfooted. Went to the well to many times and now the well has run dry. Read a interesting tweet from someone that said Intel could get to 3nm before AMD. The replies were that that could only happen by outsourcing or a miracle happens.
As for the AMD brute force approach, I think it's the other way around, it's Intel trying to brute force clockspeed at the expense of heat/power. Adding cores and clockspeed turns their processors to a furnace so a tradeoff has to happen, hence them dropping cores to contain heat.
As for the AMD brute force approach, I think it's the other way around, it's Intel trying to brute force clockspeed at the expense of heat/power. Adding cores and clockspeed turns their processors to a furnace so a tradeoff has to happen, hence them dropping cores to contain heat.
Associate
- Joined
- 28 Sep 2018
- Posts
- 2,259
AMD's main issue in gaming is their IF architecture. They just can't get data moved around fast enough, thus the latency.
It's also why the cache is so big. To further reduce the impact of high latency.
The main reason ringbus exists is because it's really fast for the mainstream desktop use case. It's the fastest method upto 8core by far.
I mean, they have on HEDT and Xeons. It's only their mainstream consumer chips still using the ring bus at this point. The idea that just ditching the ring bus solves everything is wrong though. Both Intel's mesh and AMD's Infinity Fabric have their own problems, especially with latency. Whilst the 10900K has higher inter-core latency than a 9900K due to the larger ring, it's still much lower than both Intel HEDT and Ryzen (at least when Ryzen has to dip out of a CCX, which is why the 3300X fares so much better than the other chips in that respect). The reality is that whilst the ring bus may not be perfect for a ten-core chip, nobody has come up with anything better. Not even for an eight-core or six-core one for that matter. It's possible we might see the latter two in Zen 3 though if the eight-core CCX rumour turns out to be accurate.
Soldato
The 8 core CCX isn't a rumour, it was confirmed by AMD's Milan presentation slides ages ago.
I know Intel use a mesh for the big CPUs, that's not what I'm saying. Granted ring bus does have its benefits, but the point is Intel have always known ring bus would not realistically scale past 8 cores (Comet Lake is a push) yet they've just sat on it all this time, drip feeding inconsequential updates every "new" generation. Who knows when Intel would've even moved past 4 cores on mainstream desktop had Ryzen not **** in their cornflakes. So no, the idea of just ditching ring bus is not wrong (not that I actually said that), it must be done in order for Intel to move their mainstream CPUs past 8 cores. And they can't just use their mesh either because that has serious latency issues as you point out.
This is why I said Intel have nothing left in the tank. 14nm is absolutely maxed, ring bus is absolutely maxed, new architectures are being nerfed by backporting processes to the point there's little point releasing them, their entire mainstream product line has nowhere forward to go now.
I know Intel use a mesh for the big CPUs, that's not what I'm saying. Granted ring bus does have its benefits, but the point is Intel have always known ring bus would not realistically scale past 8 cores (Comet Lake is a push) yet they've just sat on it all this time, drip feeding inconsequential updates every "new" generation. Who knows when Intel would've even moved past 4 cores on mainstream desktop had Ryzen not **** in their cornflakes. So no, the idea of just ditching ring bus is not wrong (not that I actually said that), it must be done in order for Intel to move their mainstream CPUs past 8 cores. And they can't just use their mesh either because that has serious latency issues as you point out.
This is why I said Intel have nothing left in the tank. 14nm is absolutely maxed, ring bus is absolutely maxed, new architectures are being nerfed by backporting processes to the point there's little point releasing them, their entire mainstream product line has nowhere forward to go now.
Associate
- Joined
- 28 Sep 2018
- Posts
- 2,259
You're not understanding the point. Ring is perfectly suited to get the best out of a 8core chip. They can continue to advance the cores while keeping ring there until ring itself is a bottleneck. You'll know this by measuring core to core latency so it's not like they could hide it.
It's good they're using not mesh everywhere. They found the ideal interconnect method for 8 (kinda 10) cores. AMD chose to use the same interconnect method for a 6 core as they use for a 16core+ There's a price to that. That's why you don't see performance scale with frequency in games on AMD chips. They know it's an issue and are trying to address it with Zen 3 (to a point).
Intel's issue isn't ring. Infact, ring is a pretty big advantage for intel and what keeps them clearly ahead in gaming and low core workloads. It's purely a process node issue. You're conflating multiple points that don't necessarily correlate.
Soldato
I'm not misunderstanding anything, but these multiple points do...um...mesh together (sorry 
). Ring may be perfectly fine for 8 cores, but how long are we going to stay on 8 cores? AMD want to push mainstream core counts up, Intel were forced into pushing core counts up before they wanted to do so. But how much can Intel change their underlying tech whilst they don't even have a working 10nm node for the foreseeable future, let alone anything further.
Rocket Lake has lost some of its poke by being backported to 14nm. All of the "Coves" that power the "Lakes" are going to be crippled in some way because they're not being produced on the node they were designed for (except possibly Willow Cove). Ring arguably is already a bottleneck, regardless of how good or bad the cores are. AMD are already on 12 and 16 cores for mainstream products. 8 cores are being pushed into mobile APUs. Zen 4 is likely to push core counts up, possibly meaning that 8 cores is the new baseline in AMD's product stack. And Intel would be capped at 8 cores. That's ridiculous.
). Ring may be perfectly fine for 8 cores, but how long are we going to stay on 8 cores? AMD want to push mainstream core counts up, Intel were forced into pushing core counts up before they wanted to do so. But how much can Intel change their underlying tech whilst they don't even have a working 10nm node for the foreseeable future, let alone anything further.Rocket Lake has lost some of its poke by being backported to 14nm. All of the "Coves" that power the "Lakes" are going to be crippled in some way because they're not being produced on the node they were designed for (except possibly Willow Cove). Ring arguably is already a bottleneck, regardless of how good or bad the cores are. AMD are already on 12 and 16 cores for mainstream products. 8 cores are being pushed into mobile APUs. Zen 4 is likely to push core counts up, possibly meaning that 8 cores is the new baseline in AMD's product stack. And Intel would be capped at 8 cores. That's ridiculous.
intel has their 10nm process but they say that it is not going to be as comparable to their 22nm stuff. Do they mean yield? Or performance. I mean if they got their 10nm processes sorted why not refine that. Anyways, who knows what exactly is going with intel. The parallel of intel’s current problems with regard to CPU to AMD some 10yrs ago is uncanny.
Associate
- Joined
- 28 Sep 2018
- Posts
- 2,259
I'm not misunderstanding anything, but these multiple points do...um...mesh together (sorry
Rocket Lake has lost some of its poke by being backported to 14nm. All of the "Coves" that power the "Lakes" are going to be crippled in some way because they're not being produced on the node they were designed for (except possibly Willow Cove). Ring arguably is already a bottleneck, regardless of how good or bad the cores are. AMD are already on 12 and 16 cores for mainstream products. 8 cores are being pushed into mobile APUs. Zen 4 is likely to push core counts up, possibly meaning that 8 cores is the new baseline in AMD's product stack. And Intel would be capped at 8 cores. That's ridiculous.
It's perfectly reasonable for AMD and intel to have diverging paths. It's fine to have a 8core chip that's got great per core performance and very lower latency vs a chip with very high core count but much lower per core performance.
Then people can determine their use case and choose accordingly. If you gave me a choice between my 9900k today vs a 12900k with 8 core but 30% higher performance per core or a 11950k that's another Skylake but with 12 cores, I'll take the 8 core with 30% per core performance. Why? because for MY use cases, the 8core is going to give me better performance.
You keep saying intel capped at 8 cores. They're not. They have a HEDT platform for higher core counts. There's no reason to have a single platform if a split strategy creates better performance buckets. I would absolutely not want my 9900k replaced with a HEDT equivalent.
Associate
It seems that 8 cores is the limit for Intel while maintaining manageable heat and power using the consumer CPU's ringbus. If they start using the HEDT platform to get cores up then they loose massive margin as then they will have to compete with AMD pricing for that amount of cores. Quite easy for AMD to add another chiplet to up cores but a different matter for Intel where it is not quite as easy to add cores. They (Intel) are in a very tricky situation now where AMD can add cores relatively cheaply. If Intel do the same then it eats into profits.
Also, i'm sure i've read that Intel have had to relax the 10nm node twice to get the yields they have at the moment. So where before they compared their 10nm to TSMC 7nm, where does it compare now?
Also, i'm sure i've read that Intel have had to relax the 10nm node twice to get the yields they have at the moment. So where before they compared their 10nm to TSMC 7nm, where does it compare now?
Soldato
But we're not talking about HEDT, we're talking about mainstream desktop. And in that segment Intel very much are capped at 8 cores. I've honestly lost track of what you're saying now, because it's coming across as narrow-scope pish. Right here, right now, AMD have established 12 and 16 core CPUs in the mainstream market, and it is one of their mission objectives to increase core counts on the desktop. It is a matter of fact that unless Intel can move past their existing ring bus design and scale their mainstream CPUs past 8 cores then they are going to be at a significant disadvantage; Zen 2 cores are already far superior to Skylake cores, and AMD are continually improving their latency issues. Once that happens, Intel literally have nothing left.
But Intel cannot address these issues while stuck on 14nm. As I said, all of the upcoming Coves are losing performance and features by being backported to 14nm. Intel cannot hope to improve their designs and keep up with AMD advancing technology as long as they are stuck on 14nm, the node simply won't allow it.
That is simply the case.
But Intel cannot address these issues while stuck on 14nm. As I said, all of the upcoming Coves are losing performance and features by being backported to 14nm. Intel cannot hope to improve their designs and keep up with AMD advancing technology as long as they are stuck on 14nm, the node simply won't allow it.
That is simply the case.
Zen 3 (4000 series) isn't using N7+ tho is it.. It's using N7P or something. Which isn't a huge leap forwards at all.
https://www.anandtech.com/show/1558...7nm-7nm-for-future-products-euv-not-specified
But the point is that there's nothing better. There just isn't. Even for ten cores. Nobody has come up with anything. You say that Intel have "just sat there" not improving, yet AMD haven't come up with anything better either. The Infinity Fabric is demonstrably worse for a ten-core CPU. The ring bus is the best way to handle at least up to ten cores, and likely a bit further. You say that Comet Lake is an exception and a "push" of the ring bus, yet the 10900K isn't close to being the first CPU with more than eight cores to use the ring bus, even on the desktop (the 6950X was). The fact that its scaling starts to drop beyond eight cores doesn't actually make it bad anywhere beyond that. There's clearly a tipping point where the latency will get so bad that it will make sense to move to the mesh or something like it, but I don't know of any testing that's been done to find out exactly where that is. Ultimately, it's not just as simple as saying "ring bus bad" as you're trying to boil it down to. I'll say it again to be clear - there is no better solution for a ten-core CPU. At least nothing that's been invented yet. At some point you're hard limited by physical factors, such as the speed of light. It will always take longer to travel around a mesh-like interconnect than it will a simple ring. There's no "solving" that problem unless you find a way to break the laws of physics.
Soldato
Really? You say "there is no better solution" and then immediately contradict yourself by saying "at least nothing that's been invented yet". That is entirely the point. Intel has just sat there, drip feeding the tiniest of improvements for a decade. Ring bus may well be the best thing right now for 8-10 cores because Intel had zero intention of changing it, but in the space of a few years they've been forced into the situation where their current multicore implementation for mainstream CPUs is inadequate for future growth. And I'll say it again to be clear: ring bus may be perfectly fine for 8-10 cores, but mainstream CPUs are already surpassing that count, and Intel have nothing to move on to.
So yes, it really is as simple as "ring bus bad" for Intel's future. Are you saying Intel are just going to sit back and stay at 8 cores (Rocket Lake isn't getting 10) because their ring bus works fine? What happens when AMD take 8 cores to the baseline count? What happens when AMD bring 20 or even 24 cores to the mainstream desktop? What happens when AMD's distributed chiplet approach, with all of its failings, outperforms Intel's ring?
Associate
- Joined
- 28 Sep 2018
- Posts
- 2,259
@Aretak 8 is the peak for ring as c2c latency takes a hit at 10. Look up the anandtech article that goes into it nicely. In real world terms, a group of hardcore oc’ers I’m a part of has been testing and comparing mem tuned and at a similar oc like 16/4400 bdie hand tightened on all timings. The 10core can be 2ns slower than the 8core. Is that too much? I don’t think so esp given the imc on cml is a lot better than cfl but it’s certainly measurable and repeatable.