The x86 CPU Maker You’ve Never Heard Of…

Ever heard of Intel? AMD? Of course you have! Those are your two options for CPUs, at least x86 CPUs anyway, and they are everywhere. They are in every gaming rig and office PC the world over. But what if I told you there’s actually a THIRD x86 CPU maker? Considering how fierce the litigation between Intel and AMD had to get before they finally created a patent cross licensing agreement – full video on that in the cards above by the way – you’d be forgiven for not believing anyone else in the entire world would be allowed to make x86 CPUs, and yet, there is. The company is called Zhaoxin, a Chinese maker of x86 CPUs. This is the fascinating history of China’s top CPU maker.

Our story actually starts, as all CPU stories seem to, in the USA with a little company called Cyrix. Cyrix was founded in Texas in 1988 from a small band of Texas Instruments employees, and started as most CPU companies seem to, by making Intel compatible chips. Their first chip was a maths coprocessor, specifically for floating point operations – AKA numbers with decimal points. They called it the “FasMath”, and it was a 80387 pin-compatible option that provided 50% more performance AND consumed less power at idle too. They also designed a 80287 compatible version which came out in 1991. Their first actual CPU was a rather interesting one. Their 486SLC and DLC models were pin compatible with the older 386 chips, despite offering a few 486 features like onboard L1 cache and the 486 instruction set. The 486SLC and DLC models sat somewhere between Intel’s 386 and 486 chips in terms of performance, although since the 486SLC and DLC fit in 386 motherboards it was more akin to Intel’s “overdrive” chip options where they too made pin-compatible versions of their newer processors for their older motherboards. Cyrix’s 486 chips were hounded for essentially false marketing as the chip couldn’t exactly live up to the performance its name suggested. They did later launch 486SRX2 and DRX2 chips which were double-clocked versions that finally offered proper 486 performance, and were marketed as 386-to-486 upgrades. I found this datasheet advert in Farnell’s archives that is just amazing. It includes benchmark figures, installation instructions, and the brag that “In about 15 minute, you can transform your 386 computer into a 486-class machine.” Amazing. The following year, that being 1993, they did launch an actual 486 compatible chip, the Cx486S and DX. The difference between the S and DX versions by the way is that the DX includes a floating point unit (FPU) onboard, whereas the S versions need an external maths coprocessor chip like the FasMath. These were pretty late to the party though, even behind AMD’s 486 chips – despite their lengthy legal battles – and benchmarked a touch lower than their AMD and Intel counterparts. 

1995 saw Cyrix launch the 5×86, a bit of a placeholder chip to hold off Intel’s Pentium chips, as their own Pentium competitor wasn’t ready for release. The 5×86 was again a drop in replacement for a 486 chip, although since it was actually a cut down version of their Pentium competitor M1 core, it implemented some Pentium-like features, like branch prediction. I found another datasheet for this which details the branch prediction feature, supposedly a feature with 80% prediction accuracy, although it turned out that that feature, and a few others, were actually disabled by default due to enormous instability. You could use a special software to enable them, although especially on the earlier versions you’d likely just make your system incredibly unstable. At least Cyrix’s 5×86 was a newly designed core, unlike AMD’s 5×86 chip which was just a quad-multiplied clock version of the 486. The 5×86 from Cyrix didn’t last long – just six months – as they were finally ready to launch their true Pentium competitor, the 6×86. If you’ve ever heard of Cyrix, this is likely the chip you’ve heard of from them. The full M1 core, the full feature-set, and stellar performance meant it got pretty popular. They actually named models with Pentium-equivalent performance – ie P166+ meaning it’s faster than a Pentium running at 166MHz – despite running at much lower clock speeds. Cyrix initially tried to charge a premium for these better performing chips, but much like AMD, their floating point performance was nowhere near as good as the Pentium, so Cyrix pivoted to offering discounted chips instead, helping them ship units. The key problem with the 6×86 – and the same issue AMD’s K5 ran into – was that Id Software wrote Quake for the Pentium and its dual pipeline design. The Pentium could do both integer and floating point operations at the same time, which Id used to their advantage by doing perspective corrections in the background while texture mapping – something that the 6×86 and AMD K5 just couldn’t do, and Id didn’t bother to write a fallback function if the user didn’t have a Pentium, nor a toggle to disable the feature entirely, meaning if you wanted to play Quake, you basically had to have a Pentium. That meant if you were buying (or building) a PC to game on, you basically had to buy the Pentium. 

Cyrix’s next hit – and the thing that finally meant OEMs bought their chips – was the MediaGX, a budget all-in-one chip that integrated an aging 5×86 core, plus sound and video, all on one chip. While it wasn’t exactly fast, the idea that you could have everything built onto a single budget-friendly chip meant Compaq and Packard Bell taking on Cyrix for their lowest end computer models – such as the Presario 2100 and 2200 from Compaq. Cyrix also upgraded the core to its newer 6×86 design later, and even their MMX supported cores too. The following version, Media GXi, used their newest core design, codenamed Cayenne, with a dual issue FPU – basically two instructions per cycle can be processed rather than one at a time – support for AMD’s 3DNow instruction set, and 256KB on on-die L2 cache – a first for Cyrix. These chips were also the last to be designed and sold by Cyrix – at least fully. Why? Well, you can mostly blame Intel’s litigious ass. 

Intel, not content with just being the market leader and industry innovator, spent a lot of the 90’s suing everyone and anyone who looked sideways at their CPU designs and patents. AMD ran afoul of this anger, of course, but so did Cyrix. The catch is that Cyrix, unlike AMD, was never a true Intel second-source. Cyrix never outright stole Intel’s designs, they meticulously reverse engineered everything, and often went their own way when designing their cores and products – hence the discrepancy in performance, both good and bad throughout the years. Intel saw red and sued them anyway, although Intel lost that case as Cyrix was able to prove they had just reverse engineered the designs – but Cyrix counter-sued for antitrust violations, and Intel settled giving Cyrix $12 million and a permanent license to make x86 products – so long as they were fabricated by foundries that already had an x86 license from Intel such as IBM, SGS Thomson (now ST) and National Semiconductor. During the suits, and after the Media GXi chips launched, in late 1997 National Semiconductor merged with Cyrix, with Cyrix becoming a wholly owned subsidiary of National. The theory was that Cyrix would have freer access to National’s fabs and resources, although that went poorly as National ran into money troubles not long after the merger. Cyrix did actually sue Intel later, claiming Intel infringed on three of Cyrix’s key patents for power management and register renaming – specifically in the Pentium Pro and Pentium II. Intel settled that too, giving Cyrix a full 10 year patent cross licensing agreement.

By 1999 National Semiconductor wanted rid of their CPU arm, having thoroughly run it into the ground by pushing Cyrix to make the low power Media GX style chips AND despite Cyrix having designed a whole new core, the M3, or Jalapeno which came with out of order execution, a dual issue FPU, 256K of L2 cache, an on die memory controller using RAMBUS technology which helped the onboard GPU immensely, along with the GPU having access to the L2 cache, but this was never finished, at least at the hands of National Semiconductor. Instead, a little company called VIA bought the scraps of Cyrix for themselves (before AMD got the last of it, no less) for a cool $167 million. VIA actually started as an American company a year before Cyrix, although in 1992 they decided to move the company to Taiwan to form closer ties with the growing industry there. Initially they focused on making chipsets for Intel and AMD although in 1999 they decided to branch out into CPU development – specifically x86 CPU development as they bought BOTH Cyrix and another x86 CPU maker called Centaur Technology, with Centaur costing just $51 million. Centaur started in only 1995, funded by Integrated Device Technology (or IDT for short), and they managed to produce two versions of a brand new core design under the brand “WinChip”. These were Socket 7 chips – think AMD K5 and K6 – although the WinChips were a tiny die and low power with about the same horsepower as an 80486. The WinChip 2 did improve a little, especially in floating point performance, but it was still a far cry from the K5 and K6 options you could have put in the same motherboards. 

VIA’s first CPU came in 2000, which was an altered version of the Gobi Cayenne core they called the Cyrix III, now codenamed Joshua. VIA, however, didn’t fancy Cyrix’s core design all that much as they very quickly made alternative versions of the VIA Cyrix III with the Samuel core from Centaur. Those models clocked higher at 650 and 677 MHz, versus 500 and 533 MHz for the Cyrix designed cores. Only a month later in February 2001 VIA announced there would be another version of the Cyrix III, this time with Samuel 2 cores and a new 150 nm process node (0.15 micron). This was actually a world’s first on the process node front, not that it helped the performance overly as at least according to Anandtech, the Cyrix III – even with the faster Samuel 2 cores – couldn’t keep up with Intel’s Celeron 566 and AMD’s Duron 600 in any performance test – Quake III Arena included. And just as a final screw you to Cyrix – despite paying triple for Cyrix as they did for Centaur – VIA renamed the Cyrix III to just “C3”. To be honest, apparently they initially picked “Cyrix III” as the name as they believed that the Cyrix brand name was more recognisable and popular even above their own brands, so opted for that instead. VIA bumped the core design to “Ezra”, mostly as a die shrink and added compatibility with Intel’s Pentium III “Tualatin” cores, with the main design philosophy still being low power consumption.

The next big core design was codenamed “Nehemiah”, which frustratingly was still called “C3” publicly. Apparently the Linux kernel refers to this as the C3-2 which is pretty funny. VIA realised when designing this core that their chips weren’t great for desktop, but they were great for embedded applications, and so started to tailor the C3, and subsequent chips, to that purpose. For example, having two onboard random number generators. One limiting factor was that the C3 still was stuck on Socket 370 motherboards, limiting their front-side-bus speed to single data rate 133MHz, whereas their Nehemiah+ core can as a ball-grid-array package and sported a 200 MHz FSB instead. By 2006 VIA’s cross licensing deal for Socket 370 ended, so production of C3 chips had to halt, although VIA pivoted to sell the C7 instead, using Socket 479 or BGA instead. It did feature a new core uarch, “Esther”, which follows the embedded design scheme even more, offering hardware encryption and hashing support, extremely low TDP and power consumption – supposedly less than 1 watt of average power consumption with a total of 20 watts of TDP – larger L2 cache, and dual PLLs to almost instantly switch clock speed for extreme efficiency in operation, among other changes. 

VIA then pivoted to their final line, VIA Nano, starting in 2008. Nano used a new core, Isaiah, which in particular was the first x86-64 processor from VIA – AKA 64 bit. Isaiah bumped the clock speeds and cache, and even added support for virtualisation and ECC memory, but still with the focus on low power, high efficiency chips. The first of those were the L2000 chips, like the L2100 which the register benchmarked against a comparable Intel Atom part and found them to be a pretty decent match. The L3000 series promised a 20 percent bump in performance, and a corresponding 20 percent drop in power consumption, with some parts claiming an idle power consumption of just 100 milliwatts! 2011 saw VIA’s first dual core CPU, the Nano X2 E, which was only 6 years behind Intel and AMD, labelled as the L4000 series. You could also get a dual-die version of the Nano X2, called the Nano QuadCore, which, funnily enough, was essentially a ‘glued together’ quad core chip (still 3 years behind Intel on that one). 

This, barring a revised version of the Nano QuadCore in 2015, the Nano C, was all she wrote when it comes to VIA CPUs. What happened? Well, VIA entered into an agreement with the Shanghai Municipal government to basically spin off their CPU division into the subject of this video, Zhaoxin – which apparently means “million core”. Zhaoxin was formed in 2013, and just a year later they launched their first CPU, as part of the ZX-A family. These were basically just VIA Nano X2 C4350AL chips rebranded, which isn’t all that surprising considering the company was only a year or so old. These were made with VIA Isaiah cores on a 40 nanometre process, with clock speeds between 533MHz and 1066MHz. Not exactly top performers, but that’s fine. There were also ZX-B chips which were identical, and ZX-C and C+ which was basically just the VIA QuadCore chips rebranded again. ZX-D – which is also where they rebranded to KX for desktop and KH for servers – is where we start to see some genuinely new work. The Wudaokou cores in here offered dual channel DDR4 support, PCIe 3.0, USB 3.1, and a full system-on-a-chip (including graphics with DirectX 11 support). These KX-5000 chips were made by TSMC, and were a significant departure from the VIA style of operation. No longer do you need the chipset to do literally everything, now with the CPU being an SoC the memory controller, PCIe, SPI and the likes are all onboard the CPU – much like modern AMD and Intel CPUs. Side note, I just love the optimism in this Hexus article (RIP Hexus) – “Neck and neck with AMD by 2019?” is just hilarious, especially considering this article is a) post Ryzen and b) from 2018. That’s an awfully short time to go from slower than an AMD Athlon2 x4 (by a decent margin according to this amazing article by Geri I’ll link in the description) to ‘neck and neck’! 

Interestingly, the graphics core onboard these chips – and their subsequent lines – are made in-house too. In 2001 VIA established a joint venture with S3 Graphics which had a rather interesting history. They were attempting to be a graphics card maker – just like NVIDIA and ATI – but failed to get any significant market share. They bought another 90’s GPU maker, “Number Nine”, in 1999, then merged into Diamond Multimedia – NVIDIA’s first big customer, video about that in the cards too. VIA created a joint venture with the GPU arm in 2001 and that’s how Zhaoxin have been able to make decent quality GPUs for their CPUs.

ZX-E, or KX-6000 brought another core revision, Lujiazui (loo-jee-owzway) which brought a firm 8 cores, up to 3GHz clock speeds, and a slightly modified core to better support the smaller 16nm TSMC process node they were using to manufacture the chips. A later “G” revision in 2022 added a Glenfly GT-10C0 integrated GPU – from that S3 Graphics/VIA joint venture. They also launched up to 32 core server CPUs at the same time, the KH-40000, which also have the option of being run with dual sockets for a total of 64 cores per board. That’s not quite as impressive as AMD’s 192 core chips which can also be run in dual socket mode for a whopping 384 cores and 768 threads on a single board, but still. Supposedly these have a peak of 2.7GHz clock speed, support up to 4TB of ECC DDR4 RAM, and have 128 PCIe lanes to go around. While benchmarks of these are pretty limited, it seems like the KX-6000G parts are likely to run about the same performance as a Core2Quad Q6600, while at least drawing considerably less power – 15W of TDP instead of 105W. As Tom’s Hardware says, they “perform like CPUs from the late 2000s but use[s] less power”. 

The most recent crop of chips are, you guessed it, KX-7000. These have a redesigned “Century Avenue” core for the much smaller 7nm process node, much higher 3.7GHz peak clock speeds, DDR5 and PCIe 4.0 support, and still up to 8 cores. These were meant to launch considerably earlier, but don’t seem to have started shipping until mid to late 2023, and performance is… rough. A Japanese site managed to get their hands on one, and compared to even a Core i3-8100 the KX-7000/8 (the 8 core part) struggled – and a Ryzen 5600G just trounced it entirely. Surprisingly, even when it comes to power consumption the KX-7000 doesn’t exactly amaze. The 8100 draws almost half the power despite offering basically the same performance all-core. Even idle power consumption is almost double both the 8100 and 5600G, which is quite the departure from the embedded ultra-low-power VIA approach. Of course those have been replaced by ARM chips for the most part, but still. Apparently their aim for this architecture is to match AMD’s Zen 2 performance – that being Ryzen 3000 series chips. Unfortunately it looks like the missed the mark there quite considerably – at least looking at 8 core to 8 core parts in Cinebench. 3,600 points all core for the KX, and 12,355 points for the Ryzen 7 3700X just isn’t quite there. Hell, even the Ryzen 3 3300X – a 4 core part – still offers basically double the performance all core at 6,754 points. Clearly, they’ve got a ways to go before they are “neck and neck with AMD”, huh Hexus?

Zhaoxin did also make a dedicated GPU by putting their iGPU on a dedicated card, with a 45W TDP, and going by the same Glenfly Arise GT-10C0 name. With just VGA and HDMI out, a small fan and no need for anything but motherboard power, this isn’t exactly a 5090 competitor, but considering it’s now fully home-grown in China, I’m sure they are happy to at least have an option that isn’t NVIDIA or AMD. 

Zhaoxin’s development seems to be on the slower side for sure. While the chips are definitely improving, it isn’t anywhere near the rate AMD, or even Intel, are improving. But then again, I’m not sure that’s the point. I think the point is self-reliance. Having a supplier of essentially in-house CPUs and GPUs is key for China politically, and at least on that front it seems like they are going alright enough. With 32 core server CPUs and 8 core desktop parts, while they don’t compare well to outsider options, the fact they have the option is already pretty valuable for China. For the rest of us it seems like we’ll be sticking with Intel and AMD, but at least you now know more about the third x86 CPU maker still in business!