Note: the following is a modified and shortened version of the transcript Jim used for this video. This transcript has been revised for an article format and is not a 1:1 copy. The original video can be watched here.
Alright guys how’s it going?
Prophecy, by its very nature is often hit and miss. So when I make a video on what I feel is likely to happen in 2020, I’m under no illusion that I can’t be entirely wrong. On the flip side I might be entirely right or even just a little bit right.
Obviously, I’d rather be entirely right all the time, however I won’t be and can’t be. What I can do though is analyze the current and past situations in order to form what I believe will be the most likely conclusion.
I’ll start by recapping the current situation.
So, AMD and Intel are essentially fighting 4 different CPU battles.
- First and most important from the financial perspective is the battle for the Data Center, where Intel’s Xeon fights against AMD’s Epyc
- The second battle is on the High-End Desktop where Intel’s Core X series and W Series Workstation CPUs fight against AMD’s Threadripper series
- The third battle is on desktop, both consumer and business, where Intel’s Core and Pentium Series fight against AMD’s Ryzen and Athlon
- And the fourth battle is in laptops, where again it’s Core and Pentium vs Ryzen and Athlon
Now let’s talk about size. There are people out there who believe Intel and AMD are on a reasonably similar level regarding stuff like structure and finances. A quick look over the important facts and figures quickly dispels that notion:
- Intel employs around 110,000 people. AMD employs around 10,000.
- Intel’s 2018 revenue was around $71 billion, AMD’s was around $6.5 billion.
- Intel’s 2018 net income was around $21 billion, AMD’s was $337 million.
- Intel has $128 billion in assets whereas AMD only has $4.5 billion.
Now in fact, Intel are only 1 year older than AMD, being founded in 1968 whereas AMD were founded in 1969. You might wonder how then Intel came to be so much larger than AMD. As I’m going to avoid getting too political in this video I’ll just point you to one potential answer to that question, by suggesting you watch my Intel Anti-Competitive, Anti-Consumer, Anti-Technology video.
But regarding scale, Intel has 11 times more employees and 11 times more revenue which is a huge difference. However, for me the real difference between the companies is the cash. Intel made 62 times more profit than AMD last year and they’ve been doing that every year for decades, which is a major reason why they have 28 times the value in assets.
To sum this one up simply, Intel are a far bigger company than AMD. So much larger that the latter really shouldn’t have any chance of competing.
So why then, are AMD competing better than ever before? One major reason is focus. Intel have 11 times the manpower, however they also have a huge number of products outside of x86 CPUs, from FPGA’s to networking and memory – among many others. Before Lisa Su took over at AMD, AMD were very unfocused – they were dabbling in some ARM stuff, phones and tablets before they sold their handset division to Qualcomm. They also had graphics of course.
At some point, around 2013, they were getting perilously close to bankruptcy – their products were not good at all and after Papermaster and Su came in, they canceled a bunch of projects and refocused the company on high performance compute. The graphics division did suffer as a result, however graphics and AI are just too important to ditch altogether. Neglect – is the word I would use for their GPU division.
Essentially though, Intel are huge but spread pretty thin – if you look at their huge product portfolio. AMD were small and spread too thinly, however now they are small and focused. There’s no question that Intel’s entire x86 design and production capacity still dwarfs AMD’s, however it’s not as crushing as some of the numbers might have you believe at first glance.
What is overwhelming though is Intel’s ability to… spend their way out of trouble. Again I don’t want to get too political in this one and by now I’m quite sure you must be bored of seeing this “Financial Horsepower” slide, but this is just a fact of the CPU environment. This $3 billion “spend your way out of trouble” budget for 2019 is nearly 10 times AMD’s entire 2018 profit, and if Intel needs to spend more, then they certainly could, as $3 billion is only one seventh of their entire 2018 profit.
Let’s get back to those 4 battles in Data Center, High-End Desktop, consumer/business desktop and laptops.
Right now, AMD has the largest lead in server CPU that they’ve ever had. Core counts rule in the Data Center and realistically they have 64 cores versus Intel’s 28 cores. They are basically twice as fast, these 56 Core 9200’s that Intel has don’t really exist in any kind of volume. AMD’s prices are also lower. Slam dunk you say? While it should be, it’s a far cry from that. In fact Intel still had 93% of Data Center sales last quarter. In that quarter their Data Center group saw record revenue with a mix of high-performance Xeon processors and growth in every subsegment.
AMD’s Data Center sales are increasing too, but again it’s very much a case of that 11 to 1 difference in company size also being seen in sales. This stuff moves quite slowly, however I feel that given AMD’s crushing performance advantage, they should be doing a lot better than they are.
In the High-End Desktop segment, AMD just launched their 32 and 24 core Threadripper 3000 series while Intel launched their max 18 core Cascade Lake X series. I’m sure you’ve seen the resulting benchmarks and you probably heard about the fallout between Intel and many in the tech press over how Intel tried to pull their launch forward by a few hours so the comparison wasn’t made. That ploy failed somewhat as the tech press today are much wiser to these moves.
Regarding market share, we don’t have any numbers on this, however AMD’s Robert Hallock did say during a call that at certain points during previous Threadripper product launches, they had been #1 on sales. AMD probably has around 20% of the High-End Desktop market currently, the market itself though is only worth around $1.5 billion per year.
In consumer desktop, Intel still has over 80% share, the latest numbers I had was 82% vs 18%. If you’re wondering how that can be the case, when we are constantly being bombarded with Mindfactory and Amazon number showing Ryzen 3000 eating Coffee Lake for lunch, my numbers do include the entire spectrum from the lowest end Pentiums up to chips like the 9900KS. Intel also does very well in business PC’s and with OEM’s while AMD tends to do a bit better with self-builders and power users. AMD is gaining share steadily… but again slowly. The consumer desktop arena isn’t quite as cut and dried as the Data Center and High-End Desktop are. The fact here is that gaming is important and Intel still holds a small lead at the very top, and everywhere else is a close run thing, and AMD’s chips are actually a bit more expensive, if you only care about gaming.
Even with that, I personally find Intel CPUs difficult to recommend due to all the security issues. We saw in the last video that they basically recommend disabling hyper-threading on all their chips previous to the 8th generation. I personally had an i7 6700K which I wouldn’t be particularly happy about turning into an i5. Who knows what vulnerability will strike Intel next?
And finally, in laptops Intel has around 85% share I believe, though AMD has been gaining… steadily, quarter after quarter for the past 2 years actually. But, that has come from a very low base.
So put all that together and Intel’s CPU share is between 93% and 80% – depending on segment. Again, a crushing advantage in market share advantage– however AMD is gaining in every segment.
Now for this next part I’m going to talk about manufacturing.
One major reason for Intel’s previous success was their manufacturing prowess, both in technological superiority and in capacity. Going back to around the 2004/2005 timeframe, after a mediocre 90nm node they went on an incredible roll starting with 65nm helped by their new Core architecture. Core was shrunk to 45nm then followed by Nehalem on the same process, which was another great leap forward for Intel. AMD’s 65nm was poor though, as was their Phenom architecture. However, most of the problems were fixed at 45nm with Phenom II, which only lagged Nehalem by a month.
Intel followed up on Nehalem and 45nm with a strong 32nm die shrink, called Westmere. They then followed up on 32nm with the incredible Sandy Bridge architecture. Back then Intel were still in the real “Moore’s Law” of a new node every 2 years or so.
But at 32nm things started to go badly wrong for AMD. Their first proper APU, “Llano” launched with very poor yields and clocks of only 3GHz. That was followed by their disastrous Bulldozer architecture which was large, power hungry and slow.
Intel on the other hand, had realised that time was running out for clock speeds on the old processes and opted to use FinFETs at 22nm. What’s telling here though is that even with FinFETs, they were only able to match the Sandy Bridge 2700K’s clock speeds of 3.5GHz base and 3.9GHz turbo with the new Ivy Bridge 3770K. Haswell followed a year later, on the same process and with the exact same clocks. Intel did ramp it up some time later with the 4GHz base – 4.4GHz boost 4790K.
AMD toiled with their awful architecture which was scaled down to 28nm, what was known as a “half node” in the industry. In other words it was marginally better than GlobalFoundries’ poor 32nm process but it honestly didn’t matter much as no process could possibly have saved Bulldozer.
So this was 2012-2013 and at this point in time, Intel appeared to have an almost unassailable lead. None of the foundries – be that TSMC, GlobalFoundries, SMIC or UMC – planned for FinFets on their own 20nm nodes, which lead to that node being barely better than their previous 28nm. Nvidia infamously had a real go at their partner TSMC during this period, slamming their 20nm as worthless.
Next up for Intel was 14nm and their Broadwell architecture. There were murmurs of a problem but nothing concrete until they finally launched the oddly named i7 5775C. Clocks were poor – only 3.3GHz base and 3.7GHz turbo – and the TDP was only 65W, well down on the usual circa 80W+ of previous flagships. The chip also included a large 128MB of L4 cache which did help a fair bit in gaming.
But during the Anandtech review we were treated to this slide and an admission – of sorts – that 14nm hadn’t been smooth. The forecast was that this would improve to be in-line with 22nm, but we’d discover much later on that that wasn’t how it went. If you note, on this slide Q1 ’14 is where it starts. What Intel didn’t show was the disaster they were having before that point.
Another year and another claim of 14nm yield trending to match 22nm but by this point nobody was buying it. For the first time in history, Intel’s manufacturing division looked vulnerable. They followed up on their 14nm with 14nm+, later admitting that their old “tick/tock” model was finished, replaced instead by “process/architecture/optimization”.
TSMC and GlobalFoundries meanwhile had delivered FinFETs at their 16/14nm nodes. Both had arrived at FinFETs the same way, by using a hybrid approach of 20nm and 14nm. This meant that the foundries’ “14nm” was not truly on par with Intel’s 14nm, which was on its way to becoming 14nm++, or “process/architecture/optimization/optimization”.
It was no surprise that in late 2016 rumours of even more problems on their next node – 10nm – began gathering steam, and it was clear that Intel were having real trouble maintaining Moore’s Law.
Since early 2017 the headlines have been dominated by Intel’s inability to deliver a working 10nm process, while TSMC went from strength to strength. Their 7nm is a “true” 7nm process and clearly better than Intel’s 14nm and any amount of pluses. Zen 2 has overwhelmingly superior performance per Watt. What’s more, TSMC’s 7nm+, using extreme ultraviolet lithography is coming very soon, while Intel looks forward to their own 7nm at some point in the distant future… 2022 looking most likely today.
That’s the history of how we got to this point in very late 2019. In a nutshell, Intel’s inability to get a working 10nm process has meant they’ve lost the manufacturing lead that they held for decades and AMD now has an enormous lead in performance per Watt.
But what about 2020, and beyond? Intel are huge – too big to fail – and it’s only a matter of time before they’re back in the lead, both in process and architecture. Right? Possibly, but certainly not in 2020.
When Threadripper 3000 series launched at the end of November with 24 and 32 Cores, AMD also stated that a 64 Core version was coming in 2020. I believe that they’ll launch it in a few weeks at CES. Simply put, this is 64 Cores vs 18 for Intel next year in the High-End Desktop space. Intel does have a 28 Core workstation CPU, the W-3175X, however that is based on their very best silicon die that they use for Xeon. Furthermore, it gets soundly beaten by the 32 Core Threadripper anyway.
The only thing saving Intel in High-End Desktop is AMD’s pricing. 64C vs 18C or 28C, it just doesn’t matter. That’s an impossible task for them to even get close to next year. There’s really nothing more to say about it.
Moving up a tier to Data Centers, and for the first half of the year the situation is really the same because the server chips and High-End Desktop chips are really the same. AMD will continue to gain Data Center market share, possibly at an increasing rate. I say that because TSMC will bring more 7nm capacity online throughout the year.
This leaked slide from Asus shows us that Intel has something more coming in the 2nd quarter, something being 48 Cores with Cooper Lake. There was a rumour suggesting this is a 3 die part – 2 core dies and an IO die – plus finally for Intel, PCIe 4.0. That rumour appears to be incorrect regarding PCIe 4.0 at least, according to this slide. Intel also said Cooper Lake would max out at 56 Cores, not 48, so make of that what you will. It is 14nm and 300W for the 48C part, which is about the absolute maximum that you’d expect to see in an OEM box. However, I did receive leaked information that new systems were being worked on, and water cooling will be required.
As if that wasn’t strange enough, only one quarter later Intel are supposed to be launching Ice Lake, their first 10nm server part, on the same platform. As if that wasn’t strange enough, Ice Lake has only 38 Cores, down 10 from Cooper Lake. The TDP however is only down by 10%. I think this part is even stranger as we’re surely looking at a dual chip design and Intel surely didn’t deliberately create a chip with 19 cores. That means this is very likely to be a salvage part. One year ago Intel showed off the CPU during their architecture day and it’s clearly a large package. I think putting it all together, this is likely to be the 3 chip part – 2 10nm core chips with possibly a 14nm IO die and this time for sure we’ll be seeing PCIe 4.0.
Why is intel doing this? I would guess at 270W that’s just about coolable on air, so perhaps this 38-core Ice Lake part is the best Intel can do under normal conditions. Remember, as the Ice Lake core will have a pretty nice IPC uplift of over 15% so even if clocks are down on the 14nm parts, it could be a much more competitive part against Rome, though it won’t get near the 64-Core 7742. AMD now has an even faster part, the 280W 7H12 which has a 2.6GHz base clock and is also designed for water cooling.
The problem for Intel is that Ice Lake won’t be up against Rome – it will be up against Milan. From this High Performance Computing conference back in October we learned that Milan would be a 64 core part, same as Rome. I found three parts of the presentation most interesting – first of all was that Milan appeared to be doing away with the CCX layout of 4 cores and 16MB of L3 cache each. Now the chip appears to be a true octo-core with shared 32+ MB of L3 cache.
The second interesting part was when the speaker noted that Zen 2 was still bound by memory, saying that the Rome core is so efficient that they simply cannot get enough data to it. This is actually a nice problem to have. Back in the days of Sandy Bridge and Bulldozer, Sandy Bridge gained a lot of performance with faster memory while Bulldozer didn’t. The reason for that was Sandy Bridge had a great core whereas Bulldozer didn’t.
A month or so ago AMD’s Forrest Norrod dropped something of a bombshell on Intel with his comments regarding Milan’s IPC, noting that:
Zen 3 will be based on a completely new architecture.
and will
deliver performance gains “right in line with what you would expect from an entirely new architecture.”
before stating:
[…] at a time when Intel is promising double-digit IPC gains for future microarchitectures – AMD is “confident [in] being able to drive significant IPC gains each generation.”
As I noted in a recent video, Zen 3 – Milan – is already sampling so it’s not like he’s guessing. The fact here is we have the obvious change to the architecture which could enable the IPC increase – the new cache layout – and we heard how Zen 2 was bound by memory. Recently we’ve seen that Ryzen 3000 makes decent gains with tighter memory timings too. We shouldn’t be surprised by a 20%+ IPC increase with what is effectively, a doubling of L3 cache at least.
Ice Lake gains 18% IPC but perhaps loses clock speeds, while Milan gains 20%+ IPC and a little clock speed too in all probability. 64 cores vs 38 cores, again it’s just no contest and if anything, AMD could be even further ahead if 20% IPC is low-balling it.
The 3rd interesting part of the presentation was the roadmap for all this. During this video at first I thought that these production areas didn’t make a lot of sense. Naples launched in June 2017 and this production appears to be right at the end of Q2, which would have been June. Of course the thing here is Naples came after Ryzen 1000, which launched in March. The Zen chips that both Ryzen and Epyc were based on must have been produced long before then.
Rome “production” can just about make some kind of sense for June though as it was August before Rome launched.
What’s really important here though is that compared to both Naples and Rome, Milan’s “production” is pretty deep in Q3. If we extrapolate based on Rome then Milan is starting to look like a very late 2020 part, so perhaps Intel could get 2-3 months lead in terms of time to market with Ice Lake SP. They’re gonna need all the help they can get though, as a few days ago Charlie over at SemiAccurate claimed that their entire Data Center roadmap was delayed again.
Intel refuted the claims, but they’ve been doing that for years. From this slide we can see Ice Lake was due early in the 2nd quarter of 2020. And as recently as a few days ago Intel’s Murthy Renduchintala stated at the Global TMT Conference that Ice Lake server would be coming at the end of 2020, so as usual there’s some truth in the delay talk.
Moving on to the segment which will likely be most relevant to us and certainly to me, the Desktop. Next up for Intel is Comet Lake and 10 cores, and another recent leak showed the i5 10600 with 6 cores and 12 threads.
3.3GHz is not a fantastic base clock however the chip ramps to 4.7GHz turbo. What’s interesting about this is previously we caught wind of another leak by “momomo” on Twitter showing the same 3.3GHz base and a 4.8GHz turbo. On paper, these look like decent competing parts for the Ryzen 5 series… depending on the price.
There was something else interesting here, as this Comet Lake SKU leak didn’t include any I7’s or I9’s… reminding me of some leaked information I received early in October which said “Intel is having major issues with their internal power regulation”. That’s the kind of thing that is likely to affect the higher-end parts, so something that is worth looking out for if the i5’s and i3’s launch first.
Even if the i7’s and i9’s launch, it appears that Comet Lake is essentially Coffee Lake Refresh with 2 more cores, which as we know was Coffee Lake with 2 more cores, which was Kaby Lake with 2 more cores which was Skylake with a clock bump.
10 Skylake cores on 14nm is not all that interesting and versus the 12 and 16 Zen 2 cores, it all seems rather pointless to me.
Where things do get interesting is with what AMD does next. As I described earlier, Zen 3 does appear to be coming late in 2020, if the roadmap is anything to go by, and this is for server which should be prioritized. Zen 3 is being manufactured on TSMC’s 7nm+ node, which uses EUV and promises 10% performance at the same power or 15% less power at the same performance as 7nm. Also important is the 20% reduction in area which could allow for more cache, as suggested by the 32+ MB L3.
Looking at this amazing die shot of Zen 2’s core die which is 10.32mm x 7.34mm, or 75.75 square mm, the L3 – these 8 blocks in the middle – will be around 40% of the die or around 30 square mm. We could perhaps see 64MB of L3 per chiplet – with all the cache available to every core remember – and while that would be a slightly larger die than the Zen 2 chiplet, it would still be under 100 square mm.
What about the yield on the new process though? We know that new processes start off with lower yield, however TSMC’s N7+ is still 7nm, it just uses EUV for some of the layers. According to TSMC, it has similar yield to 7nm. How good or bad is that yield? Well this is where it gets very interesting, as recently Wikichip reported on TSMC’s 5-nanometer update, including this amazing slide. As you can see, N7 in red has a defect density of only 0.09.
Let me run this through the wafer calculator quickly to show what that means. Remember the Zen 2 chiplet is 10.32 by 7.34 mm and we need to use 300mm wafers. We now set the defect density to 0.09 and that gives a 93.45% yield rate – that’s full working 8 core chiplets, and an enormous 720 good dies per wafer.
Remember, AMD is struggling to supply their markets even with this yield rate – so either they have a very limited number of wafers or something else is up.
Where this story gets even more interesting is, this 0.09 defect rate is not the actual defect rate of AMD’s process, nor is it the defect rate in November. In fact this 0.09 is the defect rate from July – before we even had Zen 2 on the shelves – and it’s actually for the mobile process.
How do I know that? Looking back at Wikichip in July I found this slide… looks familiar right? It’s exactly the same as the November slide except without the defect rate, but all points are exactly the same. What’s more, we can see “N7 Large Die” beneath which means…
“With the company pushing into HPC, they have started reporting defect densities separately for mobile customers and HPC customers with die sizes of 250 square millimeter dies and larger.
I can’t think of any reason why they’d specify large dies, however we know that Navi 10 is 251 mm2 and for sure counts as part of this “Large Die” defect rate… which is even lower than the mobile process. Zen 2 is also on the high performance node so it must also be included in this “Large Die” defect rate… which must be around 0.06 or 0.07. In fact, that doesn’t make a huge difference to the yield on Zen 2 as it’s already near max, but every little counts.
So, yield on 7+ for Zen 3 is likely to be north of 90% so you might think that moving everything over to 7+ asap is what makes sense, however that will come down to factors like capacity and ramp time. Historically EUV suffered from a lack of throughput, around half of a non-EUV machine. However EUV allows for a reduction in mask count which should claw back some of that disadvantage by allowing for faster production cycles.
I have no idea what percentage of TSMC’s 7nm machines will be switched over to EUV in 2020, but I would guess at under 20% and possibly under 10%.
The point here, if you recall we were discussing consumer-grade CPUs, is that I’m just not certain that we’ll see Zen 3 on the desktop on 7nm+ in 2020. It’s important to note that seeing 2 new architectures one year after the other is pretty rare these days. Bulldozer was refreshed with Piledriver the year after, Zen was refreshed with Zen+ the year after.
What AMD could do – and I’ve alluded to this before – is launch Zen 3 – Milan – in Data Centers while refreshing Zen 2 on the current 7nm process… OR use this N7P optimization which unlike N7+ uses the same design rules as N7 – so no need for any reworking, and offers 7% performance improvement at the same power, or up to 10% lower power at the same speed, compared to N7.
That sounds a bit like what the move from 14nm to 12nm brought with Zen+. In this case, with the very best chiplets also going to Ryzen 3000 as the new server chips will be moving on to Zen 3, AMD could get a double benefit from such a move, gaining some decent clocks speeds and performance per Watt.
Another alternative is that they “backport” Zen 3 from N7+ to N7 or N7P. That way the IPC gains from Zen 3 will be achieved at the cost of slightly lower clock speeds and performance per Watt. The gains however would be much more capacity on the current node.
Near the beginning of this video I gave figures on market share. At some point AMD needs to start fighting a market share battle on the Desktop and they won’t do it by moving on to 7+ EUV next year. That will simply give us the same situation regarding supply that we saw this year.
The requirement for a real push in volume is even more obvious given Intel’s dreadful supply issues. At the recent talk Murthy again went over Intel’s 14nm supply issues and only last week we learned that they were resurrecting 22nm Haswell’s for the Pentium market. While that is a pretty desperate move, it will help to alleviate some of the pressure on Intel, as will their 10nm Ice Lake server products coming late next year.
It’s been over 2 years of shortages for Intel and this shortage could easily continue until 2022. Somebody at AMD surely did the calculations and figured that there is a large swathe of desktop market just begging to be taken over. It’s fairly telling that Intel’s CEO Bob Swan recently stated that holding 90% CPU market share is no longer the goal, and that Intel will focus on winning 30% of the silicon share. Holding 90% CPU share is simply no longer possible for them if AMD is anywhere near as aggressive at going after it.
So for this final part of the video – let’s be honest it’s pretty rosey for AMD and pretty awful for Intel, but that’s the reality of the CPU landscape. The reality is that Intel is still massively larger, selling many more parts and making much more money, and this will continue in 2020 and beyond.
In mobile they have Ice Lake on 10nm which looks like a good part, and according to Anandtech’s Surface laptop showdown, it runs circles around Zen+. Of course we’re expecting Zen 2 mobile chips soon, sadly still with Vega graphics though.
I wouldn’t expect to see a big difference materializing in the laptop market share over 2020. Intel has their Athena Project and throws a lot of money at the ecosystem. Whether or not they have enough 10nm supply to supply the market in the way they have been is unlikely, however they’ll get by with a combination of 10nm Ice Lake and 14nm parts at the high-end.
AMD on the other hand will have a lot of 7nm wafers, however, a lot of that will be moving over to the consoles. Laptops are a fight they aren’t really fighting and that makes sense.
I could continue with this video and I had planned to talk about 2021 and beyond, however it’s already running on far too long so I’ll call it here.
2020, gonna be interesting for a variety of reasons. Clearly though, Zen 3 is the big one and unless Intel has back-ported Ice Lake to 14nm, they could be in a hell of a trouble on the desktop late next year… or possibly early 2021.
I’ll be back with something different I think, before the Christmas break. In the meantime check out the website, Anubias has written some nice technical pieces on the inner workings of CPU which is well worth a read, currently at part 4, with more to come.
I’ll catch you later guys.
