A retro-experiment post popped up in my feed: someone boots up a Cyrix MediaGX — a 1997 processor with an x86 core, a "video card," a memory controller, and PCI all on a single die. The author calls it the "world's first APU" and wonders why almost no one remembers it.
What caught my eye was something else: that "world's first APU" claim. AMD coined the term APU—and they coined it in 2011, fourteen years later. So either we're dealing with a rare case of a term catching up to reality with a double-digit delay, or a retroactive label slapped on someone else's work. Figuring this out turned out to be more interesting than either of the two processors.
The text ahead has three storylines: what it was technically, why the market rejected it, and how it came to pass that the design lost on the desktop but won in industry.
The MediaGX launched on February 20, 1997—from Cyrix Corporation of Richardson, Texas. Inside a single 120 mm² die with 2.4 million transistors sat: a scalar Cx5x86 core, 16 KB of unified L1 cache, a DRAM controller, a graphics accelerator, and a PCI interface. There was no longer any need for a north bridge as a class of component.
It ran on 3.3 volts, was manufactured in IBM's 0.44-micron process with three metal layers, cost Cyrix about $35 to produce, and sold for $79 (120 MHz) and $99 (133 MHz) in batches of a thousand.
This fundamentally set it apart from what Intel was doing. In the Pentium, the memory controller, graphics, and PCI lived on separate chips—MediaGX pulled them inside. Analyst Linley Gwennap of Microprocessor Report, who reviewed the new product in March 1997, put it bluntly: Cyrix had rolled out a chip that does more than any Intel product.
The second act of the program was the companion chip, the Cx5510. Inside it lived ISA and USB, IDE, AC'97 audio, RAMDAC, and an MPEG-1 decoder. The two-chip combo replaced a dozen components that made up a standard budget PC of those years. In 1998, BYTE journalist Tom Halfhill described the result this way: the system board shrank to five chips, and nothing else was needed to build a computer (barring power supply, memory, and disk).
Sounds like a triumph. Then come the details that ensured there wasn't one.
The MediaGX had no L2 cache. None at all. There was physically nowhere to connect it: the memory controller was already inside the die, and the design provided no external bus to hook cache memory to.
Cyrix spun it beautifully. Since the memory controller is inside and communicates with the core over an internal 64-bit bus at the core's clock frequency, a DRAM page miss costs about the same as an L2 access on a Pentium-133. The logic sounded convincing: if memory is already at hand, it doesn't need a cache.
Gwennap objected head-on: a Pentium L2 hit will be better than a MediaGX DRAM page buffer hit, which means without caches at both levels, the Cyrix chip only beats a Pentium that also lacks L2. This was confirmed on paper: in tasks that purely load the CPU—spreadsheet recalculations, database searches—the 133 MHz MediaGX lost badly to a Pentium-133 at the same frequency. What saved it was that real home applications mix compute with graphics, and in that mix the chip was competitive.
Here is the crux of the whole story. MediaGX implemented UMA—unified memory architecture: the framebuffer stopped being a separate memory chip and moved into standard RAM.
Cyrix pitched this as cost-saving, and the press release proudly announced: 2 MB of framebuffer was no longer needed. Gwennap looked at it soberly and wrote that the savings were hidden, not realized: the same 2 MB of DRAM was still carved out of the total pool and occupied by graphics. The Compaq Presario 2100 computer sold with 24 MB of memory, of which only 22 MB was available to software. And since a cheap framebuffer was made from the exact same chips as main memory, the monetary benefit turned out to be minimal.
The second nuance was more interesting than the first: UMA inherently degrades performance because the CPU and graphics start fighting over the same bus. The industry benchmark of those years was a drop of 10–15%. Cyrix solved this with on-the-fly data compression: a compressed framebuffer takes up less space in memory, and you have to read it less. In the Winstone 96 test, the ratio was 20:1—meaning 95% of framebuffer accesses from main memory vanished. A move that must be acknowledged as elegant.
But here's what's important: Gwennap in 1997 called UMA a solution that doesn't catch on in PCs—several chipsets supported it, but no one was rushing to build machines on them. Two years later, in a review of the Intel 810 chipset, the very same Microprocessor Report wrote that the 1995-vintage UMA revolution had already failed, and the 430VX chipset with UMA was almost never enabled in that mode.
Now let's shift our gaze fourteen years forward. November 2020, Apple is talking about the M1 and specifically calls out unified memory architecture in a bullet point of the press release—a shared pool of memory in the same package, accessed by all SoC blocks to avoid copying data between them. By this point, the phrase "shared memory" had stopped being a dirty word and became the main selling point in the presentation.
The same mechanics, the same question—"let graphics live in shared memory"—and two opposite verdicts twenty-three years apart.
There were three technical reasons for the failure on the desktop, and they fell one after another like dominoes.
Compression didn't scale. The trick that saved UMA in the era of 2D graphics yields almost nothing in 3D, where textures and depth buffers are pushed through memory. Intel didn't use the Cyrix trick at all in their 810 chipset in 1999—Microprocessor Report notes this without explanation, and I suspect that framebuffer compression was no longer saving the day by then.
Upgradeability vanished. MediaGX only worked on a board specifically designed for it, because the processor was too tightly coupled with the chipset. You couldn't swap the processor or add a graphics card: the design didn't provide for an AGP bus at all.
Integrated graphics couldn't catch up with discrete. The rendering speed of integrated solutions was modest, and in 1999, when Intel released the 810, analysts wrote honestly: the integrated 3D engine is slower than "many of today's cheapest discrete chips."
And then the thing that made everyone forget about MediaGX happened. Intel released the 810—their version of the same idea, with integrated graphics, UMA, and dynamic memory allocation instead of hard-allocating the framebuffer at startup. Microprocessor Report analysts directly linked the two things: the MediaGX UMA approach "led to the sub-$1000 PC market," MediaGX "left the stage," and now Intel was striking back with its own solution. With the Intel brand, the 810 went to the mass market, and the pioneer did not.
At the same time, no one in the analyst camp was under any illusions: the 810 retained all the "old problems of UMA," and to get the performance of a standard machine in a UMA system, the buyer was advised to get a processor one tier faster. The price difference of a tier was $10–30—and that's exactly what ate up all the chipset savings.
Cyrix didn't survive to see these teardowns as an independent entity. In August 1997, National Semiconductor bought it for about $550 million in stock—not so much for the processors as for the rights to them: National had a cross-licensing agreement with Intel, without which Cyrix, as a fabless designer, legally couldn't freely manufacture x86-compatible chips. National's CEO Brian Halla saw the future not in competing with Intel, but in "PC-on-a-chip": fully compatible devices at a $500 price point, and eventually $200. The press announcement of the deal was described as schizophrenic: Cyrix insisted on direct competition, National—on a "computer on a chip."
Two years later, National renamed the result NS Geode. In 2003, the Geode business went to AMD. And then something happened that is more important for history than all the marketing disputes: the chip stopped getting updated, and that's exactly why it lived for twenty-two years.
The subsequent fate of the Geode LX is a set of deferred promises. In 2009, AMD announced that it would not create a new microprocessor, but Geode LX800 would be available for purchase until 2015. In 2016, the deadline was pushed to 2019.
In early 2018, embedded board manufacturers led by congatec negotiated an extension even further—to 2021. The final batches shipped in the first quarter of 2022. Twelve years of announcing its own funeral.
In hardware terms, this is explained by a figure I found in materials from distributor EMAC: the AMD Geode LX800 at 500 MHz consumes 1.8 W, and the entire LX line fits into 1.2–2.6 W with a TDP of 2.8–5.1 W. This kind of heat dissipation can be handled without a fan at all, which means you can put the board in a warehouse terminal, a thin client behind a rack, or a medical device that no one will touch for years.
The chip also managed to be an indirect cause of a specific technical drama. That same Geode LX controls memory refresh itself, and its controller can't handle more capacity: according to EMAC, at 512 MB of DDR333, this takes up 50% of the clock cycles, and at 1024 MB—70%. Hence the direct consequence: above 512 MB, you can't raise the frequency to DDR400, the memory simply doesn't have enough time.
MediaGX/Geode turned out not to be a desktop product, but an industrial one. Three niches kept it alive.
Thin clients. The Compaq T30 on NS Geode GX1—a 2002 device running Windows CE that survived in real workplaces until 2010. A separate irony: inside was a fully-fledged x86 computer with four USBs, PS/2, COM, LPT, VGA, and network, but the stock firmware could do exactly one thing—connect to a remote desktop. No file manager, no notepad. Power choked by closed software.
Educational laptop. OLPC XO-1—the famous "$100 laptop for children in developing countries" project—went into production in November 2007 on an AMD Geode LX-700 clocked at 433 MHz with a power consumption of 0.8 W. The choice was dictated not by speed, but by the fact that the device had to run off a solar panel and have no fan.
Industrial tablets and PC/104 boards. Terminals, kiosks, onboard systems. The Geode LX800 replaced the AMD K5 in this role, and its own replacement—the AMD G-Series—managed to come out in three generations and retire before the Geode itself did.
And one more unexpected consequence: the design proved so resilient that it spawned an entire secondary market for its replacement. Taiwan's ICOP, a manufacturer of embedded boards, still maintains a page with a "drop-in replacement" selector for the LX800—with the caveat that the main difficulty isn't performance, but that new processors don't support legacy I/O and old OSes. In their comparison table, the Geode LX800 at 498 MHz scores 5791 points in CrystalMark versus 24495 for the replacement—and people still refuse to migrate for years, because on the other side is Windows XP and the ISA bus.
Returning to what I started with. AMD released the first lineup under the APU moniker—Brazos, the E- and C-series—on January 4, 2011, at CES. Senior Vice President Rick Bergman called them "the greatest advancement in computing since the introduction of x86 over 40 years ago" in the press release. Six months later, in June, the "Llano" A-series launched, where the wording reached the level of "a turning point for AMD and perhaps the biggest architectural change in the industry since the invention of the microprocessor."
Now let's compare dates. MediaGX with integrated graphics and a memory controller launched in February 1997—14 years before a word was coined for this configuration. Intel put graphics next to the CPU in Clarkdale in January 2010, still on a separate die in the same package, and only Sandy Bridge in 2011 brought both onto a single die. AMD arrived at the single-die configuration last of the three and called it the first appearance of a new category.
And here's the uncomfortable part. Calling the MediaGX the "first APU" is a projection in hindsight. APU as a category includes marketing volume, an ecosystem, and a word that didn't exist in 1997. It's more accurate to say: Cyrix built an architecture that the industry eventually caught up to, and didn't survive to see it, because it was too small for the idea it had come up with.
Three places where the question isn't fully closed.
First—the price tags. I cited launch prices of $79 and $99 from Microprocessor Report dated March 10, 1997. But in a Cyrix press release from June 30 of the same year, where they announce 166 and 180 MHz, the price for 133 MHz is already listed as $63, and for 150 MHz—$75. In four months, the price of a faster chip fell below what the slow one cost, and I don't know if this is explained by volume discounts, a process node change, or just different configurations. It doesn't matter for the story, but substituting one figure for another without a caveat would be a lie.
Second—about "hardware bugs." In popular retellings, they are exactly what explains the MediaGX's death on the desktop. Gwennap in 1997, taking the chip apart piece by piece, doesn't write about bugs at all: he writes about the lack of L2, weak floating-point performance, UMA, and the impossibility of an upgrade. The errors spoken of in hindsight are more like the vehicle Cyrix's reputation was riding in by 1999, rather than a 1997 diagnosis. Where the real defect ends and folklore begins, I can't say: there's no clear engineering errata list in the public domain.
Third—watts. The "1.8 W" for the Geode LX800 from the distributor's materials is the typical core consumption, while the TDP of the entire LX line, according to official documentation, fits into 2.8–5.1 W. The difference is threefold, and for a story about passive cooling, it's fundamental: dissipating 1.8 W without a fan is easy, 5 W—already requires a proper heatsink. Which number is more honest for a specific board, I don't know.
The most useful thought here, in my opinion, is about how to read industry terms. What gets declared "first" in the industry isn't what appeared earliest, but what managed to become a category with a working name. MediaGX did everything technically fourteen years before it came to be called an APU—and didn't make it into the textbooks, because it looked at the task differently. Cyrix optimized not for speed, but for the computer manufacturer's purchase order, and the result of this optimization is measured not in benchmarks, but by the fact that the line lived for twenty-two years in a market where twenty-two years is six generations of Intel.