AMD Could BLOCK Sale of Intel!

AMD could actually block the sale of Intel to Broadcom, how crazy is that! But wait, Broadcom is buying Intel? And how the hell could AMD block that? Well let’s backtrack a little bit and explain. To be clear, the idea that Intel is going to be bought out is just rumours – much in the way there were endless rumours AMD would be bought out back in their rough days, now that AMD is flying high and Intel is in the dumps, well the rumours have flipped. As for Broadcom, again, that’s still somewhat rumour mill stuff, although a WSJ report says that Broadcom has, “informally discussed with its advisers making a bid but would likely only do so if it finds a partner for Intel’s manufacturing business” – which is where the rumour that TSMC would take over Intel’s fabs, while Broadcom takes the chip designs for themselves. Again, the word even in that quote, “informally” suggests the sincerity of this, but the actual sale isn’t the most interesting part here, that would be the fact that AMD could sink any deal to buy Intel, so how the hell does that work?

Well, to properly understand that we need to look back how it came to be that we have only two* x86 CPU makers. To be pedantic there’s actually three, AMD, Intel, and Zhaoxin (formerly VIA), although the latter really only sells in-house in China and at least by comparison seems to lag behind in performance and capabilities, who themselves only acquired an x86 license by buying Cyrix, but for this story they aren’t all that important. The first CPU Intel made was the 4004, with a whopping 4 bit bytes and a max clock speed of a blistering… 750 KHz. For context, just in clock speed alone, a modern Intel chip like the 285K is 7,600 times faster than the 4004. That’s just insane. Anyway, that was pretty basic, although for 1971 it was pretty sweet. It took until 1978 for Intel to make their first x86 chip, the 8086. These were 16 bit chips, and, significantly, this was the first Intel chip with ‘microcode’. We’d likely just call that ‘firmware’ today, but basically it’s software that’s built into the chip itself that means you can alter how the chip works after building it, and importantly it also counted as copyrightable software, not only patentable hardware. Even more importantly, in 1980 Intel secured a deal with IBM for the one, the only, the IBM PC. 

The big problem with that deal is that back then the sorts of people who bought microprocessors – people like IBM – required chip makers to offer a second-source chip maker, alongside the first-source, in this case Intel. That was partly due to silicon manufacturing space being incredibly limited and therefore difficult to meet demand, and partially as a redundancy precaution in case the first-source chip maker went bust or stopped making the chip they needed. Luckily for Intel, AMD, founded by another ex-Fairchild employee, Jerry Sanders, had already been second-sourcing Intel products including the 8080 – AMD called their slightly improved version the 9080 – with a full cross licensing agreement in 1976, which was then extended in 1982 for another ten years to help Intel fulfil their 8086 and 8088 chip contracts with IBM. Unfortunately, things went downhill pretty quickly after that. AMD happily second-sourced Intel’s 80186 and 80286, but Intel felt they were getting the short end of the stick in this arrangement as AMD was getting these high-tech, well selling CPUs, and Intel was getting peripheral chips with no major sales volume. By 1985 Intel was ready to launch the much improved 80386 – known more as just the 386 so I’ll stick with that – which was a pretty significant leap forward in performance, especially thanks to being a 32 bit chip. It found its way into IBM PC clones like the Compaq DeskPro which sold incredibly well, and since IBM had moved manufacturing of their own 286 clone in-house and weren’t all that interested in the 386, Intel opted to purposefully offer no second-source options. Considering AMD still had that 10 year deal – less than 3 years old at that point – AMD took Intel to the forced arbitration included in the deal. Intel responded by terminating the agreement with one year’s notice.

The arbitration took a grueling five years, but in the end Intel was found to have breached the contract. AMD quickly moved to sell their own 386 clone, the AM386, but through what can only be described as an amazing cover story or the most insane coincidence, Intel found out AMD were launching the AM386 early and sued to try and halt production. Basically Intel’s story is that two different Mike Webb’s were staying at the same Sunnyvale Hilton hotel at the same time, one from AMD and one from Intel, and when AMD send confidential documentation to their Mike Webb, well the hotel got it wrong and gave the documents to the Intel Mike, not the AMD one. That’s Intel’s story, although AMD struggled to believe that and suspected foul play or outright espionage. Regardless, a judge ruled that “386” was a generic term not protected by copyright, and therefore let AMD bring their 386 clone to market – a full six years after Intel’s original chip. Amazingly, this is what caused Intel to ditch straight chip numbers and switch to names, with the 386 being renamed to the i386 so it could be copyrighted, and just two generations on they swapped from numbers entirely, to a name you’re likely very familiar with: “Pentium”. 

That wasn’t the end of the litigation, oh god no. AMD and Intel sued each other six times in the 90’s, with Intel suing AMD in 90, 91 and 93 over microcode copyright infringement in AMD’s 80C287 math coprocessor, along with the AM386 and AM486 respectively. That is significant, because in 1989 in another Intel lawsuit against NEC, a judge ruled that microcode was software and therefore was copyrightable (even though NEC’s V20 and V30 chips didn’t actually violate that copyright meaning intel technically lost that lawsuit, but they won the war as that decision would be imperative for their fight against AMD). AMD sued Intel in 91 with a federal antitrust case claiming Intel was trying to create a monopoly, and again in 92 under California unfair competition laws because, and I can’t believe Intel had the gall to do this, but they were demanding that buyers of the AM386 and AM486 chips had to pay Intel PATENT ROYALTIES. Like what the hell Intel, that’s both a big brain and big cojones move. In 95 both teams decided enough was enough and settled all their cases, with Intel paying AMD $18 million for breach of contract, and AMD paying Intel $58 million for patent infringement.  

In the midst of this battling, Intel launched their Pentium lineup – starting in 1993, and later with the Pentium MMX chips in 96 – these chips are a pretty big leap forward again. The 486 chips introduced a feature called pipelining which is an attempt to keep the whole chip busy at once, by adding parallelism to the chip’s operation. Instead of one operation blocking access to the whole chip, each part of the chip – fetch, decode, execute, etc – is able to operate simultaneously, so instruction one runs it’s fetch command on the first clock cycle, then on clock cycle two instruction one takes that fetch and starts decoding, while the next instruction in the queue starts it’s fetch. Next cycle instruction one gets executed, while two gets decoded, and a third instruction gets fetched. That cycle repeats for every stage, meaning instead of having to wait for instruction one to fetch, decode, execute, access memory and write out the result before the next instruction starts, you can get way more throughput per cycle making the chip way faster. Pentium introduced an upgrade to this, which is being ‘superscaler’. This basically takes the idea of pipelining to the next level by executing two instructions at the same time with two different execution units. For context, Intel’s latest Arrow lake chips, at least the performance cores anyway, have a total of 18 execution ports. The Pentiums had one main pipeline, and a simpler pipeline, meaning that in most cases it could handle two operations per cycle. Again, that was a pretty big step forward. 

AMD, already half a decade late to the party, couldn’t directly compete with these new Pentium chips – at least on the design front. Funnily enough AMD were actually ahead of Intel in the nanometer game (sound familiar?) and brought out a 500 nanometer, then 350 nanometer, version of their AM486 they called the DX4. That was actually faster than the early Pentiums, despite being a considerably older chip. That did give AMD some breathing room to create their ‘pentium class’ chip, the K5, in 96. The K5 changed the game entirely, such that we still use its core design philosophy today. That being uOPs, or micro-operations, basically each x86 instruction the chip is sent gets split up into RISC instructions that the CPU then acts on. You might have heard of RISC before, but only in the context of ARM or RISC-V chips, with RISC standing for reduced instruction set computing, rather than x86’s CISC, or complex instruction set computing. This simplifies the chip design into a smaller set of base instructions, with a front end component that handles the CISC to RISC conversion. This was a huge step forward, meaning AMD’s K5 chips, even at lower clock speeds, were faster than the Pentiums in almost every workload. Except Quake with its weird floating point stuff that favoured the Pentium. 

Intel responded with the Pentium Pro, which also featured uOPs, along with a bunch of other features including on-chip level 2 cache, and physical address extension or PAE which is what lets you address more than 4GB of RAM, to name a few. It was amazing for the time – albeit flew a little too close to the sun because it struggled to run 16 bit code – something that Windows 95 and 98, along with like every game available at the time was riddled with, making it seem like a less-than-impressive bit of kit. AMD, on the other hand, bought out NexGen and rebranded what they’d be working on as their new K6 chip, which at least fixed the floating point performance so Quake ran well. Intel’s second go at the Pentium Pro, the Pentium II, was better with 16 bit code, and here’s where we start to see Intel taking more aggressive anti-competitor, anti-consumer moves. They put the damn CPU on a proprietary card! Before the Pentium II, any socket 4 chip would work in a socket 4 motherboard, be that Intel, AMD, Cyrix, NexGen, whoever. If it fits, it sits. Wait no that’s cats. If it fits, it works. Hell Intel would even sell you “overdrive” versions of their newer chips to fit in your older boards, they were that compatible. But Intel wanted a way to shut everyone else out, AMD included, so they split the market. If you wanted a Pentium II, you had to buy a slot 1 motherboard. This is where we also see the start of Intel offering multiple SKUs of their own products – not just newer versions that effectively make the older ones outdated, but purposefully lower spec versions of their newest, hottest products. Enter, “Celeron”. This is also where “Xeon” comes in, this is a slightly higher spec version of the main chip, mostly focused on the server market. Clearly though Intel didn’t think too far ahead with slot one, so created slot 2 to accommodate things like multi-processor for the server market, further frustrating motherboard makers.

AMD opted to follow that with the K6-III which introduced the three level cache we know and love today (and that Intel has just moved away from with their new four level cache…), and segmented their own product line with the Duron. We’re into 2000 now, so we’re at least in this century. Hallelujah!  In 2001 AMD launched the Athlon, which was considerably faster than the Pentium III in basically all regards, while still being cheaper. Needless to say the Athlon was a big success for AMD. Unsurprisingly, Intel struggled getting their 250 nanometer process node to work (although later refined the Pentium III with the Coppermine and later Tualatin architectures at 180 and 130 nanometer), and were just about to absolutely light money on fire with the infamous “itanium” chips. 

Itanium was the first 64 bit option from, well, anyone, and featured a bunch of new architecture changes, like EPIC – explicitly parallel instruction computing and VLIW or very long instruction word – basically you could pass a bundle of instructions to the processor which could then run that whole bundle in parallel. The problem with that is you need the software to be explicitly compiled with this in mind. The software needs to be able to bundle the right instructions together and send that bundle off to the CPU, rather than the usual queue. The problem? Well no one, not even Intel, could make a compiler that could actually do that overly well. So while potentially good in theory, in practice it was… let’s just say poor. Even more hilariously, Intel actually had a deal with HP to provide Itanium chips until at least 2017, with the last deliveries being in 2021, so Intel has been burning cargo-ship-sized bundles of money every year keeping this titanic failure afloat. They once predicted up to $38 billion per year in revenue from Itanium servers, but the actual figures are just… woof. 

AMD, on the other hand, played it much smarter They didn’t try and reinvent the wheel, they just spruced it up with a little 64 bit magic. Enter X86_64, or AMD64, an updated 64 bit version of x86, or Intel386 as it’s often called. These chips were fully compatible with 8, 16 and 32 bit code – along with 64 bit of course – and considering the choice between a completely backwards compatible chip family, or something that was practically dead-on-arrival, well it was an obvious choice to opt for the opteron (pun intended). Intel saw the writing on the wall here and tucked their tail between their legs and politely asked AMD if they could license AMD’s new instruction set to be able to build the Pentium 4, and later Celeron and Pentium D chip families. 

Amazingly, at least according to PassMark, in 2006 AMD had an all time high 48.4% market share in the desktop processor market. That is still 10% higher than even today with their Ryzen dominance (albeit that doesn’t include the Core Ultra versus 9800X3D from late last year and early this year). AMD was on fire! This is also when dual core CPUs really hit the market, with Intel’s Pentium D – a dual core Pentium 4 – and the Athlon 64 X2, although in a beautiful piece of irony, the Pentium D was actually just two Pentium 4 dies on one substrate, wereas the Athlon X2 was a ‘proper’ dual core, ie both cores on one die. Oh how the times have changed… Naturally the Athlon X2 was more performant, but it was more expensive too, compared to the cheaper, but slower, Pentium D. Naturally the Athlon X2 was more performant, but it was more expensive too, compared to the cheaper, but slower, Pentium D. The even funnier thing was that Intel was really only staying in the game by bumping the clock frequencies over and over to try and keep up with AMD, so for the next wave of chips they knew they had to do something different. And different they did! They stole Transmeta’s dynamic clock frequency and voltage tech! Seriously, Transmeta, one of the other x86 CPU makers from the early 2000’s invented the whole turbo boost clock and voltage behaviour, and while most other people licensed the design from them, Intel just straight stole it. Transmeta sued Intel, and in 2007 Intel settled for $250 million – which in today’s money is almost double at $383 million. Anyway. They took the core design from their mobile version of the Pentium III, the front side bus from the Pentium 4, and the stole speed step tech, and turned that into the Core and Core Duo.

At the same time, AMD seemed to be a little too comfortable up on that perch, and spent all their R&D money on acquiring a little video card company called ATI. This wasn’t the best time for AMD to spunk all their money, because Intel swiftly launched the Core 2 Duo which AMD’s offerings just couldn’t keep up with. AMD pumped the clock speeds again, to little success, especially when Intel dropped the absolute bombshell that was the Core 2 Quad, a quad core version of their already successful Core 2 design. AMD did attempt to compete with the Phenom line, including the Phenom II X4 and later X3 and X6. The fun fact about those especially three core parts is that the early versions had a potentially amazing feature – you could re-enable one, two or all three missing cores right in the BIOS. Basically the way that this sort of product segmentation works is that the big silicon wafers aren’t perfect, and the individual dies that make up that wafer can have defects. Some are outright buggered, but some have more minor defects in, say, just one core. So, AMD tests the dies, find one has a broken core, so they disable two or three of those cores and sell it as a lower end part, therefore saving that bit of silicon from the bin. Now the ‘proper’ way to disable those cores is to laser them off from the rest of the chip, but I guess AMD forgot to do that, meaning they just disabled them in software, so enterprising nerds could re-enable them after-the-fact. Now there’s a good chance they were disabled for a good reason, but some reported that they bought an X3 and unlocked it to a seemingly fully functional X6, and that seemed like a killer deal at the time! 

Sadly for AMD, even the Phenom II X6 couldn’t really keep up with Intel’s Sandy Bridge chips, the Core i7 2700K in particular. This is the start of the dark days for AMD as their market share slipped to just over a quarter – 27.2% according to PassMark – and that’s only the beginning. AMD struggled pretty heavily on the financial front, so in 2009 opted to do the unthinkable – sell off their fabs. At the start of 2009 AMD spun off their own foundries into GlobalFoundries – Mubadala, the investment fund of Abu Dhabi (the capital of the United Arab Emirates) – pumped $700 million into AMD to buy a majority share in their foundry business, with $1.2 billion of AMD’s debt being transferred to GlobalFoundries, freeing up AMD to, at least in theory, come up with something brilliant. 

They didn’t. Their FX lineup, starting with Bulldozer, was a disaster, sending AMD on a death spiral – but unlike most other important but financially ruined companies, AMD couldn’t be bought out and saved. Why? Well that’s where the key thing for this story comes in. When AMD sold their fabs off, Intel got their knickers in a twist about ‘AMD sub-licensing Intel’s property without consent’. Because GlobalFoundries were no longer direct subsidiaries of AMD (although at least initially AMD was still a minority owner), Intel saw that as a breach of the peace and started being rather feisty. By November of 2009 AMD and Intel had struck a new patent cross licensing agreement, which allowed AMD indefinite access to Intel’s patents – namely the original x86 architecture – and allowed Intel access to AMD’s patents – namely AMD64. Sounds great, right? Well there’s a few killshots in there, with the primary of which being clause 5.2. That describes the different reasons the agreement would be automatically terminated. Obviously breach of contract would be one, but bankruptcy and importantly here, “Change of Control” too. That means if either company were to be bought out, the agreement is automatically terminated. 

This one clause is credited as basically the sole reason AMD was never bought out during their DECADE of barely scraping by, because any buyer would have to either immediately fight Intel on patents, or at best strike a new cross licensing deal with them, which was not a sure thing. Intel likely wanted that clause initially because they were the goliath in the market and saw AMD’s dire state and wanted to give them a little poison pill just to be sure, but now it may well be coming back at them. If Broadcom and TSMC want to buy – and split up – Intel, they would have to negotiate with the new champion, AMD, and that isn’t a sure thing either. AMD, with it’s new found might, could very well squeeze the comparatively struggling Intel and block any potential deals. While personally I don’t think they would, I’m sure they’d rather leverage the position to get a very favourable deal out of it and avoid having to deal with anti-trust monopoly lawsuits, it is technically a possibility and that’s kind of mad considering where AMD was a decade ago. From death’s doorstep to the top CPU maker (well trading places with Qualcomm anyway, albeit with different markets) is one hell of a comeback, but it’ll yet to be seen if Intel can regain their throne or if their current woes are and AMD-esque death spiral. I don’t think Intel is in anywhere near as dire a state as AMD was in the 2010s, so it feels a little premature to be discussing buyouts for them, but if nothing else I hope you’ve enjoyed learning about the history as much as I did while writing this video!