This is only a preview of the December 2026 issue of Silicon Chip. You can view 37 of the 104 pages in the full issue, including the advertisments. For full access, purchase the issue for $10.00 or subscribe for access to the latest issues. Articles in this series:
Items relevant to "Advanced Cable Analyser":
Items relevant to "Inrush Current Limiter":
Articles in this series:
Items relevant to "NeoPixel RGB LED Analog Clock, Part 2":
Items relevant to "The Magic Candle":
Purchase a printed copy of this issue for $15.00. |
Part 2 by Dr David Maddison, VK3DSM
Image source: www.pexels.com/photo/vintage-computermotherboard-with-cpu-and-ram-slots-37256941
THE HISTORY OF
AMD
trials and tribulations
We started this series last month by describing how AMD was
started by a few disgruntled Fairchild employees (like Intel...)
initially operating from a garage. It grew quickly, to the point of
building clones of most of Intel’s processors. We continue from
1993, with the Pentium on the horizon.
T
he period covered in the first article in this
series was roughly 1969 to 1995. During
that time, its main products were
copies of other company’s designs,
usually under license, sometimes via
reverse engineering. However, AMD
did design its own chips, some of them
quite complex.
One such example was the Am2900
chipset from the mid-1970s, a powerful (for the time) processor solution
that was used in minicomputers and
mainframes. From that, its ambition
only grew, and AMD would soon start
producing RISC processors that were
14
Silicon Chip
well ahead of their time, such as the
Am29000.
The Am29000
By 1993, AMD had produced the
Am386 (1991) and Am486 (1993). To
understand the next part of the AMD
journey, we need to take a step back for
a moment. By 1975, the microprocessor market was split into two philosophies: cheap and slower single-chip
CPUs like the Am9080/Intel 8080, or
expensive but faster/more powerful
multi-chip processors like the Am2900
family.
Australia's electronics magazine
The multi-chip Am2900 family used
bit-slice technology to build powerful processors for markets where
cost was no object, such as high-end
minicomputers, military hardware
and graphics applications. Chips like
the Am9080 were used for hobby and
consumer PCs.
Systems built using AMD’s Am2900
chips ran extremely fast (for the time)
at around 20-40MHz, compared to x86
processors like the 8080 (1974) and
8085 (1976), which ran at 2-3MHz.
Because they were so fast, they
could even run emulators to mimic
the 8085, which could run up to ten
times faster than the actual 8085 chip.
As processor technology advanced,
AMD moved away from the multichip bit-slice concept to produce the
single-chip Am29000, also known as
the 29k (see Figs.13 & 14). Like the
Am2900, it was a RISC processor.
When it was introduced in 1987/1988,
it was one of the fastest 32-bit processors in the world, running 3-5 times
faster than CISC chips of the time.
For example, the first version of the
Am29000 could achieve 17-23MIPS
(million instructions per second),
while CISC processors like the Intel
80386 or Motorola 68020 would run
at only 3-5MIPS.
Although the Am29000 was not an
x86 processor, for reasons that will
be explained later, this chip allowed
AMD to compete with the new Intel
Pentium (released in 1993) which Intel
refused to license to AMD (along with
the 80386 and 80486; AMD reverse-
engineered them as mentioned previously).
In the 1990s, x86 processors were
dominant in the PC market even
though the 29k was significantly faster.
The 29k beat Intel in the embedded
processor market for such devices as
HP Color LaserJet printers and graphics terminals, but as the 29k was not
an x86 processor, how could it be used
in the PC market if it didn’t ‘speak’ the
x86 language?
Despite being faster, the 29k was losing money. Therefore, in 1995, AMD
officially cancelled the entire 29k family of processors and moved the engineering team to its PC division. Since
AMD already had an advanced, superfast RISC processor in the form of the
29k, it would have been foolish to
waste the design effort, so they recycled elements of the design into what
they called the K5.
siliconchip.com.au
This was to be AMD’s next big processor, designed to compete against
Intel’s Pentium (1993), specifically the
P5 and P54C versions. “K” in K5 stands
for the fictional element Kryptonite,
which is Superman’s only weakness
(by analogy, Superman was Intel).
The “5” is a reference to the Pentium,
which was so-named because it was
next in the series after the 286, 386
and 486.
Developing the K5
The architecture of the Am29000
(29k) released to customers in 1988
was completely unrelated to the Intel
x86 architecture. The 32-bit Am29000
was a RISC processor. However, AMD
needed something to compete with
the Intel Pentium, which was released
in 1993.
To build a competitive processor,
they took elements of an unreleased
version of the Am29000 codenamed
“Jaguar”, such as the mathematical
engines and out-of-order execution
logic, and attached an x86 decoding front-end, which translated x86
instructions into 29k-style RISC
instructions on-the-fly (which they
called RISC-ops).
For the back-end, they used the
physical execution units, arithmetic
logic units (ALUs) and maths coprocessors from the 29k. The K5 contained
five ALUs from the Jaguar microarchitecture.
To handle speculative out-of-order
execution, where the processor can
execute instructions ahead of time
based on what it predicts the software
will do next, it relied on the Reorder
Buffer (ROB) from the Jaguar, with
minor modifications to keep track
of x86 registers rather than 29k registers. The floating-point unit (FPU)
was taken from the Am29050 to be
used in the K5.
The K5 also adopted the concept of
the branch target cache (BTC) from the
Am29000, which minimised processor
pipeline stalls by storing the targets of
predicted program branches (eg, conditional jump instructions).
The BTC is a tiny, high-speed memory buffer that instantly feeds the first
few instructions of a detour in code
(such as an if/else statement that forces
the computer to break from a linear
sequence to a different set of instructions), giving the slower main memory enough time to catch up without
stalling the processor.
siliconchip.com.au
RISC vs CISC
RISC stands for Reduced Instruction Set Computer, while CISC stands for
Complex Instruction Set Computer. In the 1970s and 1980s, these represented
two different approaches to designing processors, although the distinction has
become much less clear over time.
The idea behind RISC processors was to use a relatively small set of simple,
regular instructions. These were often encoded using fixed-length instruction
words, making them easier to decode and pipeline efficiently.
In contrast, CISC processors used larger and more complex instruction sets,
often with variable-length instructions and many addressing modes. Some
instructions could do quite a lot of work, but might take multiple clock cycles
to complete. This was considered programmer-friendly, especially when more
software was written directly in assembly language.
Arm and RISC-V are modern examples of RISC-derived instruction set
architectures. The main surviving CISC architecture in general-purpose
computers is x86, as used by Intel and AMD processors.
However, modern Intel and AMD x86 processors do not execute most x86
instructions in a simple, literal way. Instead, they decode them into internal
micro-operations, which are then scheduled and executed by a highly optimised
processor core.
These internal operations are often described as RISC-like, although they
are private implementation details rather than a separate RISC instruction set.
The AMD K5 was an early example of this approach.
Over time, the distinction between RISC and CISC has blurred. CISC
processors such as x86 adopted many RISC-like implementation techniques,
including pipelining, register renaming, out-of-order execution and internal
micro-operations.
Meanwhile, RISC architectures became more complex, adding features
such as compressed instruction encodings, SIMD/vector extensions, memoryordering features, privilege modes and specialised instructions. As a result, RISC
and CISC now mostly describe the software-visible instruction set architecture,
not a simple division between how processors work internally.
Figs.13 & 14:
the AM29000
chip and
die. Sources:
https://w.wiki/
RQYm
https://w.wiki/
RQYn
Australia's electronics magazine
December 2026 15
Fig.15: the AMD K5 block
diagram. Note the four
instruction decoders to
convert x86 instructions to
RISC-like instructions.
The K5 was designed to be pin-
compatible with the Pentium; a user
could simply unplug their Pentium
from a Socket 5 or Socket 7 motherboard and drop in a K5.
However, the K5 turned out to be a
disappointment. It suffered from serious design flaws: lower-than-expected
clock speeds and poor performance
compared to Intel’s Pentium and Pentium Pro.
The K5’s fatal flaw
Original source:
https://w.wiki/RQYe
In essence, the K5 was less of a Pentium clone and more of an Am29000
masquerading as a Pentium.
Independent Architecture and the
GHz race (1996-2003)
After the legal battles of the 386/486
era, AMD made a strategic decision
to stop relying on reverse engineering and commit to designing its own
next-generation x86 processors from
the ground up. This led to the K5,
K6, and K7 (Athlon) processor families that would ultimately transform
the company.
The 32-bit K5 (Figs.15-17) was
released in 1996. It was AMD’s first
independently designed x86 microprocessor, intended to compete with
Intel’s Pentium.
AMD’s core RISC architecture was
highly efficient, but it struggled to meet
the clock speeds of Intel chips, which
the marketplace viewed as a competitive feature. Because of the high efficiency of the RISC architecture compared to CISC, it is not necessary to
have the same high clock speeds for
equivalent performance.
However, to overcome this marketing deficiency, AMD introduced
the concept of PR or Processor Rating numbers. For example, an AMD
K5-PR100 ran at a physical clock speed
of only 66MHz but AMD claimed it
was equivalent to an Intel Pentium
running at 100MHz, hence 100 in the
name of the K5. AMD also sold the
PR100 at much higher prices than the
100MHz Pentium.
The K5 had superior raw integer processing performance to the Pentium,
which made it better for running typical office software, but it had an inferior
FPU, which is necessary for complex
3D mathematical operations as used in
games. It happened that, just after the
K5 was launched, 3D games like Quake
were released, which ran much better
on the Pentium’s superior FPU.
Thus, the Pentium retained the loyalty of gamers and power users, for the
time being at least.
The chip taught AMD painful but
valuable lessons about the extreme
difficulty of designing competitive
x86 processors. Nevertheless, as
AMD’s first fully in-house designed
x86 CPU, it was an important milestone for them.
Introducing the K6
AMD recovered quickly with the
32-bit K6 (1997), which was based on
technology acquired from NexGen –
see Figs.18 & 19.
AMD made a strategic acquisition
by purchasing NexGen in 1996. It was
a small but innovative semiconductor company founded in 1986. NexGen had developed one of the most
advanced x86 processor designs of
the era with its Nx586 and Nx686
architectures, which used a RISC-like
internal core that translated complex
Fig.16: the AMD K5
physical package.
Image sources.
Source: https://w.
wiki/RQYf
Fig.17: the AMD
K5 die. It’s
significantly more
complex than
its predecessors.
Source: https://w.
wiki/RQYd
16
Silicon Chip
Australia's electronics magazine
siliconchip.com.au
x86 instructions into simpler micro-
operations, much like the K5.
NexGen extensively patented this
concept, releasing its Nx586 in 1994,
before the K5. Struggling with the disappointing performance of its own
K5 processor, AMD acquired NexGen
for approximately US $850 million
in stock.
The acquisition proved to be a turning point. AMD scrapped much of
the troubled K5 project and adopted
NexGen’s superior microarchitecture,
using the Nx686 as the foundation for
the K6 processor, adding a floating-
point unit and adapting the chip to
fit standard motherboards. It then
launched the K6 in 1997.
Many of NexGen’s top engineers also
joined AMD, bringing critical expertise that would later contribute to the
successful Athlon (K7) and subsequent
generations. The NexGen acquisition
gave AMD both advanced technology
and vital engineering talent at a crucial moment in its history.
The K6 was a much stronger performer than the K5, finally giving AMD
a credible competitor to Intel’s Pentium II (1997). It introduced features
such as MMX (multimedia extensions)
support and a larger internal memory cache.
While still behind Intel in raw
single-threaded performance and manufacturing process technology, the K6
allowed AMD to regain market share
through aggressive pricing and excellent floating-point performance.
The K6 series proved that AMD
could produce competitive processors, but it also highlighted that the
company was still playing catch-up
to an extent.
The K7
A breakthrough came with the 32-bit
K7, codenamed Athlon, launched in
Complicated Athlon naming schemes
AMD’s K7 chips were sold over 1999-2003, while the K8 chips were sold during
2003-2007. The 32-bit K7 was named AMD Athlon, with later models being
Athlon XP. They competed against the Pentium III and early Pentium 4.
For the K8, AMD decided to keep the trusted Athlon name, which had gained
great respect and market success, while adding 64 to indicate this was the
world’s first mainstream 64-bit processor. It was named Athlon 64 and later
Athlon 64 X2 (dual-core).
K7 and K8 were internal names, not official names, while K5 and K6 were
official names; the “K” series was discontinued after the K8. Anther name used
for the K7 is the Duron, which was the budget version, launched in 2000. It had
a smaller L2 cache to reduce its cost.
The K8 used the names Opteron for its enterprise/server model, which
competed against the Intel Xeon; Athlon for mainstream models, which
competed against the Pentium; and
Duron or Sempron for the budget
versions, which competed against the
Intel Celeron. The Turion 64 was used
for laptops and competed against the
Pentium M. See Table 4 overleaf.
Fig.20: an AMD Athlon (Thunderbird)
PGA socket CPU. For this generation,
they abandoned the heat spreader for
better thermal contact between the
die and the heatsink. It worked, but
unfortunately it was possible to crack
the die if too much or uneven clamping
force was applied. Source: https://w.
wiki/RQZG
1999 – see Figs.20 & 21. The Athlon
represented a complete architectural
overhaul and was AMD’s most ambitious project to date.
Its key innovations included a completely new superscalar, out-of-order
execution microarchitecture with nine
execution units; a revolutionary EV6
bus (originally developed by Digital
Equipment Corporation [DEC]) instead
of Intel’s GTL+ bus, allowing much
higher bandwidth and easier multiprocessing; and excellent floating-point
performance.
The original “Classic” Athlon processor was in the Slot A format, like
Intel’s Pentium II (see Fig.22).
In March 2000, AMD achieved a
historic milestone when it released
the Athlon 1GHz. It was the first x86
processor in the world to reach 1GHz
(1000MHz or one billion instruction
clock pulses per second). From August
1999 until January 2002, it was the
fastest x86 chip in the world. This was
a major psychological and marketing
victory over Intel.
The Athlon not only matched but
often exceeded Intel’s Pentium III
(1999) performance, particularly in
gaming and content creation. For the
first time in its history, AMD was
widely regarded as having the faster
desktop processor. The success of the
Figs.18 & 19: an
AMD K6-233APR
processor and die.
The metal can
on top is a heat
spreader that also
physically protects
the silicon die
underneath. Most
modern CPUs use a
similar approach.
Sources: https://w.
wiki/RQZC
https://w.wiki/
RQZD
siliconchip.com.au
Australia's electronics magazine
December 2026 17
Athlon showed that AMD was no longer just a cheaper alternative; it led
the market in performance, at least
for a while.
This coup was achieved through a
combination of talented individuals,
hard work, careful planning and calculated risks. The relative success of
the K6, thanks to the acquisition of
NexGen, generated the breathing room
and cash flow to attempt something
more ambitious.
AMD had its own technical experts,
the people they brought on from NexGen, and also benefited from talent
connected with DEC/Alpha, including
Dirk Meyer and Jim Keller. The K7 they
developed had strong out-of-order execution, a much better floating-point
unit than AMD’s earlier chips and a
bus architecture influenced by DEC’s
Alpha EV6 bus.
Intel was also heading toward the
Pentium 4/NetBurst era, where it
would chase clock speed very aggressively, with limited success. During
1999-2003, Intel was not making flawless architectural choices, and AMD
exploited that gap.
AMD used partnerships to compensate for being smaller. Jerry Sanders’
“virtual gorilla” strategy said that since
AMD could not match Intel’s internal
resources, it could partner its way into
competitiveness.
For example, they partnered with
Motorola to develop a copper interconnect process. Athlon was among
Fig.21: the AMD K7 “Athlon” architecture, again a step up in complexity but
bringing significant performance benefits. Source: https://w.wiki/RQZE
Fig.22: the original Athlon processor, the Athlon Classic in Slot A configuration.
The chips on both sides are L2 cache memories. Like Intel, AMD later abandoned
the slot form factor due to frequency limitations. Source: https://w.wiki/RQYg
18
Silicon Chip
Australia's electronics magazine
the first commercial processors to
use copper interconnect technology,
helping AMD push frequency and
efficiency.
Wired reported in August 1999 that
AMD claimed the Athlon 600 was
about 9% faster on average than a
600MHz Pentium III and 21% faster
in graphics tests. Custom PC’s retrospective on the Athlon 600 backs up
those figures.
Innovation leadership (2003-2007)
Encouraged by the Athlon’s success, AMD went further into the highend market. In 2003, the company
launched the Opteron processor, the
first implementation of K8 architecture
(see Figs.23 & 24).
The Opteron introduced AMD64
(later known as x86-64), a 64-bit extension to the x86 instruction set that
AMD developed.
Basically, AMD decided to unilaterally extend the 32-bit x86 instruction
set to 64 bits, allowing vastly larger
amounts of memory to be addressed
linearly (the 32-bit limit was 4GiB)
while maintaining full backward compatibility with 32-bit software. This
proved to be highly successful, as Microsoft, the Linux kernel developers
and other operating system vendors
quickly added support.
Most importantly, Intel itself eventually adopted AMD64, initially
calling it EM64T and later Intel 64,
rather than continuing with its own
incompatible 64-bit architecture (Itanium/IA-64). By 2006, Intel had fully
embraced AMD64, making it the universal standard for 64-bit x86 computing.
AMD64’s success gave AMD a temporary technological lead, and it also
ensured the long-term survival of the
x86 architecture into the 64-bit era.
Unlike Intel’s earlier failed Itanium
architecture, AMD64 offered full
backward compatibility with 32-bit
software while adding true 64-bit
addressing and performance. The
Opteron quickly gained traction in the
server market, particularly for scientific computing, databases and high-
performance clusters. It gave AMD its
first serious inroads into the enterprise
server segment, a market Intel had long
dominated.
While the 1GHz K7 Athlon relied
on a traditional Front Side Bus (FSB)
to access RAM via a separate memory
controller chip on the motherboard,
siliconchip.com.au
Table 4: a comparison of some AMD K7/K8 variants and competing parts from Intel
AMD processor
Year
Target market
Intel competitor
Outcome
Athlon (K7)
1999
High
performance
Pentium III Slot 1
Athlon Slot A (cartridge), 500-1000MHz.
Faster than the Pentium III.
Athlon (K7)
2000
High
performance
Pentium III
Socket 370
Athlon Slot A, equivalent in performance
to the Intel model. AMD crossed the 1GHz
barrier one day before Intel.
Athlon XP (K7)
2001
Budget
desktops
Early Pentium 4
Popular socketed version.
Athlon MP (K7)
2001
High-end
workstations
and servers
Early Xeon
processors
based on the
Pentium III and
early Pentium 4
Designed to support dual-CPU
motherboards for workstations.
Duron (K7)
2000
Budget
desktops
Celeron
Duron was superior to Celeron.
Athlon 64 single-core (K8)
2003
High
performance
Single-core
Pentium 4
AMD had the superior processor.
Athlon 64 FX Ultrapremium single-core (K8)
2003
Gamers,
overclockers.
Pentium 4
Extreme Edition
The FX was the pinnacle of single-core
gaming performance at the time.
Opteron (K8)
2003
Servers/data
centres
Pentium 4 Xeons
AMD had superior architecture and
performance in this era.
Athlon 64 FX dual-core
and Quad-FX models (K8)
20032006
High
performance
Core 2 Extreme
6800, Core 2
Quad
The Quad FX needed a special motherboard
to run two dual-core FX chips at the same
time; Intel’s Quad model was a single chip.
Sempron (K8)
2004
Budget
desktops
Celeron D (based
on the Pentium
4)
A competitive “value” battle.
Athlon 64 X2 early dualcore models (K8)
2005
High
performance
Pentium D 800
and 900 series
“Presler”
AMD’s dual-core chip was superior to Intel’s
dual-die approach for the Pentium D.
Turion 64 (K8)
2005
Laptops/mobile
Pentium M
Turion was 64-bit vs the 32-bit Pentium M,
but the Pentium offered superior battery life.
Athlon 64 X2 later dualcore models (K8)
20062009
High
performance
Core 2 Duo
To remain competitive, AMD pushed the K8
to 3.2GHz (Athlon X2 6400+).
Figs.23 & 24: the dies of a four-core Opteron “Barcelona” (left) and six-core Opteron “Instanbul” (right) processor.
Sources: https://w.wiki/RQZH & https://w.wiki/RQZJ
siliconchip.com.au
Australia's electronics magazine
December 2026 19
the K8 removed this bottleneck by
building the memory controller
directly into the CPU and using AMD’s
“HyperTransport” technology for fast
communication with the cores.
Internally, the K8 was an evolution
of the K7. Its performance gains came
mostly from architectural improvements, such as the integrated memory
controller, HyperTransport system bus
and AMD64 extensions. By reducing
memory latency and improving the
way the processor communicated with
the rest of the system, the K8 could
often do more useful work per clock
than earlier x86 processors.
Early Slot-A versions of the K7/Athlon used separate Level 2 cache chips
mounted inside the processor cartridge, rather than placing the cache on
the CPU die itself. Later K7 versions,
starting with Thunderbird, moved the
Level 2 cache onto the processor die,
where it could run at the full processor speed. The later K8/Athlon 64 and
Opteron processors also had on-die
Level 2 cache.
In addition, later versions of the K8,
like the Athlon 64 X2, had two processor cores on a single piece of silicon.
The Athlon 64 X2 was not the first
consumer dual-core CPU, as Intel’s
Pentium Extreme Edition 840 and
Pentium D reached the desktop market slightly earlier. However, AMD’s
implementation was widely regarded
as the cleaner and more effective
design, and it was AMD’s first native
dual-core desktop processor.
The K5-to-Opteron era (1996–2003)
was a major transformative period in
AMD’s history. The company completed its transition to a genuine innovator. Although the K5 and early K6
had difficult teething problems, the
engineering and other lessons learned
during those years directly enabled
the spectacular success of the Athlon
and Opteron.
By 2003, AMD had proven it could
design world-class processors, competing fiercely with Intel on both performance and price.
AMD’s entry into the graphics
market
AMD’s entry into the graphics market began with another of its strategic
moves. In July 2006, AMD acquired
ATI Technologies for US$5.6 billion.
This was one of the largest acquisitions in the semiconductor industry
at the time. The deal was intended to
transform AMD from a CPU specialist
into a full-platform provider, capable
of competing with Intel across both
processors and graphics.
The CEO at the time, Hector Ruiz,
believed the future of computing lay
in “heterogeneous integration”. This
is the combining of CPU and GPU
capabilities on the same chip or package, where both processors can work
together on general-purpose computing tasks, not just graphics.
By owning ATI’s Radeon graphics
technology, AMD hoped to accelerate the development of integrated
“Fusion” processors (later called
APUs, accelerated processing units)
that would merge the CPU and GPU
on the same chip.
ATI acquisition
By acquiring ATI in 2006, AMD
gained several critical assets:
1. The well-established Radeon
graphics brand.
2. Advanced GPU architectures,
including the highly regarded R300
family (Radeon 9700/9800) that had
pioneered programmable shaders,
with excellent image quality.
3. Substantial expertise in high-
performance graphics for both consumer gaming and professional workstations (FireGL/FirePro).
4. Early work on combining graphics with other processing functions.
This acquisition transformed AMD
from a CPU-only company into a true
CPU + GPU competitor almost overnight. While the integration proved
difficult and expensive in the short
term, which contributed to AMD’s
financial struggles in the following
years, it laid essential groundwork for
later successes, including Ryzen APUs
and the Instinct line of AI accelerators
(described later).
The ATI acquisition was therefore not merely about entering the
graphics card business. It was a longterm strategic gamble that the future
of computing would involve tightly
coupled CPU and GPU silicon working together. This vision would take
more than a decade to fully mature,
ultimately becoming one of AMD’s
greatest strengths.
Intel was increasingly integrating
basic graphics directly into its processors (Intel GMA, Graphics Media
Accelerator) during 2004-2006, aiming
to control the entire PC platform and
reduce the need for discrete graphics
cards. But for AMD, CEO Hector Ruiz
believed the future of the company
and computing in general was with
heterogeneous computing.
Heterogeneous computing
After acquiring ATI, AMD began
working toward processors that
combined x86 CPU cores and
Fig.25: the ATI Radeon
HD 2900 XT launched
in May 2007, and was
one of the first cards to
be released after AMD
had acquired ATI in
2006. Source: www.
vgamuseum.info/index.
php/home/item/666-atiradeon-hd-2900-xt
20
Silicon Chip
Australia's electronics magazine
siliconchip.com.au
Radeon-
derived GPU cores on the
same die (APUs).
This fitted into the broader idea
of heterogeneous computing, where
different types of processor cores are
used for the tasks they handle best. A
CPU is well-suited to general-purpose
sequential code and operating-system
tasks, while a GPU can process many
similar operations in parallel.
Programming systems such as
OpenCL (Open Computing Language)
were designed to let software use this
mixture of CPUs, GPUs and other processors in a more portable way.
This vision took time to mature.
NVIDIA’s CUDA software platform
became dominant in GPU computing, helped by strong hardware, early
developer support and a growing software ecosystem. OpenCL remained
important, but it did not become
the main programming model for
high-performance GPU computing
in the way its supporters (including
AMD) once hoped.
Even so, AMD’s Fusion/APU thinking was important. It helped establish
the idea that the GPU should was a
major computing resource inside the
system.
Dark years & the Bulldozer
architecture (2008-2016)
Up to 2007, AMD was in a very
strong position with the success and
popularity of the Athlon series of processors, culminating in the Athlon X2,
successfully competing with Intel’s
offerings. Then things took a turn for
the worse during the so-called Dark
Years, roughly between 2007-2016.
While the Athlon X2 wasn’t discontinued until 2009, sales had already
slumped by 2008, and the later X2s
were marketed as budget processors.
Despite strong momentum up to
that point, the company started to
face a series of strategic, technical
and financial difficulties that lasted
nearly a decade and brought it close
to collapse.
The acquisition of ATI was seemingly the correct move based on the
future of heterogeneous computing,
but it added significant debt to the
company. Also, integrating two large
engineering organisations with different cultures proved more difficult
than expected. The development of
the first “Fusion” APUs took longer
than planned and consumed substantial resources.
siliconchip.com.au
On top of those considerable challenges came the global financial crisis of 2008. PC and server demand
weakened sharply, hurting both AMD
and Intel. AMD was more vulnerable because it carried a heavier debt
load, had fewer financial reserves, and
was less dominant in the high-margin
server market. The company reported
a large loss in 2008 and was forced to
cut costs aggressively.
In October 2008, AMD announced
plans to split off its manufacturing
operations into a new company, which
was formally launched as GlobalFoundries in March 2009. This marked
AMD’s transition away from being an
integrated device manufacturer, like
Intel, toward becoming a fabless chip
designer.
From this point on, AMD would
design processors but rely on external foundries, initially and primarily GlobalFoundries, to manufacture
them. The move reduced AMD’s capital burden, but it was not an instant
escape from manufacturing problems.
AMD remained closely tied to
GlobalFoundries through wafer-
supply agreements, so GlobalFoundries’ process delays and yield problems would later cause difficulties of
their own. In the longer term, however, becoming fabless gave AMD
more flexibility, eventually allowing it
to take advantage of TSMC’s (Taiwan
Semiconductor Manufacturing Company) increasingly advanced manufacturing technology.
At the same time, two of AMD’s
major technical initiatives from this
era, Bulldozer and Fusion, produced
disappointing or mixed results. Fusion
helped establish the APU concept,
but Bulldozer failed to deliver the
expected performance and efficiency
gains, leaving AMD to flounder.
Bulldozer
AMD suffered a major technical
setback with its first completely new
CPU architecture since the K8, the
Bulldozer architecture (October 2011).
It introduced a modular design in
which two integer cores shared a single
floating-point unit and other resources
(Fig.26). The flagship of the Bulldozer
architecture upon its release was the
FX-8150 on a single 315mm2 silicon
chip. The multi-chip 16-core Opteron
“Interlagos” had two 8-core chips
in one CPU package (AMD kept the
Opteron name over several processor
generations).
The goal was to improve multithreaded performance and power
efficiency while reducing die area. In
practice, however, the design delivered disappointing results. Single-
threaded performance was often lower
than the previous generation of processors.
Fig.26:
an AMD
Bulldozer
architecture
die (32nm),
possibly the
FX-8150. Each
Bulldozer
module had
two integer
cores sharing
a single
floating-point
unit. Thus,
four modules
gave eight
cores. Source:
https://w.wiki/
RQYj
Australia's electronics magazine
December 2026 21
Technical problems included that
branch prediction was weak, cache
latency was high, and the chip struggled to reach competitive clock speeds.
The FX-8150, despite having eight
cores, frequently underperformed
Intel’s quad-core Sandy Bridge processors in many consumer workloads.
Power consumption was also higher
than expected.
AMD attempted improvements to
Bulldozer with Piledriver (2012) and
Excavator (2015), which brought some
improvements in efficiency and clock
speed, but they did not close the performance gap with Intel’s Haswell and
Broadwell processors.
As a result, AMD lost considerable
market share in both the desktop and
server segments. By 2013, Intel held
roughly 80-85% of the x86 microprocessor market, while AMD’s share fell
into the low teens.
Due to the Bulldozer, Piledriver and
Excavator failures, AMD’s financial situation became critical. They reported
net losses in multiple years between
2011 and 2015. The company’s stock
price dropped below US$2 in 2015;
market analysts questioned whether
AMD could even survive as an independent company.
Multiple staff layoffs occurred, and
AMD was forced to restructure several
times. Manufacturing dependence on
GlobalFoundries (the former AMD
spin-off) also became a disadvantage,
as GlobalFoundries lagged behind
TSMC and Intel in process technology
during this period.
Fusion APUs – an idea before
its time?
While Bulldozer struggled, AMD
pushed forward with its Fusion processors. This was the direct result
of the 2006 ATI acquisition and part
of AMD’s heterogeneous computing
strategy. Fusion was the marketing
name, but AMD gave them the technical name of APUs, hence they were
called AMD Fusion APUs.
However, the Fusion name had to
be phased out starting in 2012 due to
trademark conflicts, leaving the processors to be called APUs. As a result,
they phased out the name Fusion System Architecture (FSA) and started to
use the term Heterogeneous System
Architecture (HSA).
Fusion aimed to combine CPU and
GPU cores on the same die, AMD’s
vision for heterogeneous computing
22
Silicon Chip
at the time (before moving to the chiplet concept).
These early APUs offered very good
integrated graphics and power efficiency, but their CPU performance
(with Bulldozer-derived cores) was
mediocre. They were decent for budget laptops and desktops but couldn’t
compete effectively with Intel’s stronger CPU cores at the time.
Despite these difficulties, AMD
maintained the vision of investing in
its heterogeneous computing strategy.
The company released its first APUs
(later rebranded as Ryzen with Graphics), such as the processors with the
internal development names of Llano
(2011), Trinity (2012), Richland (2013),
Kaveri (2014) and Bristol Ridge (2016).
These generations were all sold
under the name A-Series, with different designations under that.
Transition to chiplets
The progression from the Fusion
concept to a chiplet architecture in
2017 represents one of the greatest
transitions in semiconductor history.
It marked AMD’s shift from a failed
‘all-in-one’ monolithic processor philosophy (CPU, GPU and other parts all
on one piece of silicon) to a modular
‘Lego-brick’ style strategy that fundamentally saved the company.
The successful production of
chiplet-based CPUs allowed for cost
reduction, better yields, process node
optimisation (only using expensive
nodes for parts of the processor) and
excellent scalability.
Throughout the Dark Years, the
Fusion strategy relied on monolithic
dies. As processors get larger, the
chance of a manufacturing defect
increases exponentially. If one tiny
part of a massive monolithic chip is
broken, the entire expensive chip has
to be thrown away (although, in some
cases, sections can be disabled and it
can be sold as a less capable part for
less money).
The scaling problem
Under Fusion, AMD had to spend
immense resources designing completely different monolithic chips for
laptops, desktops and servers. Short of
cash, AMD simply couldn’t afford the
research and development to compete
with Intel across all these markets.
A crucial leadership transition
occurred in 2014. Dr Lisa Su, who had
joined AMD in 2012 as head of global
Australia's electronics magazine
Fig.27: current (since 2014) AMD CEO
Dr Lisa Tzwu-Fang Su. She joined at
the tail end of the ‘dark years’ and has
overseen AMD’s partial overtake of
Intel, thanks to the Zen architecture.
Source: https://w.wiki/3q$z
business development, was promoted
to CEO in October 2014. Su was a
highly respected engineer with a strong
technical background and immediately
began a major restructuring.
She simplified AMD’s product roadmap, reduced operating expenses and
refocused the company on engineering execution. Most importantly, she
initiated the development of a completely new CPU architecture called
Zen, which aimed to address the fundamental weaknesses of Bulldozer.
By 2015-2016, AMD was operating
in survival mode. The company had
managed to preserve its graphics business and made progress on APUs, but it
still remained far behind Intel in CPU
competitiveness.
Observers at the time gave AMD
few chances of a meaningful recovery.
The period from 2007 to 2016 tested
the company’s resilience and forced it
to confront deep structural problems
in design philosophy, manufacturing
strategy and organisational efficiency.
These difficult years also created the
conditions for renewal. The painful
lessons from Bulldozer, the integration challenges from the ATI acquisition, and the financial pressure from
that and the global financial crisis ultimately pushed AMD toward a disciplined, engineering-focused approach
that would define its later success
under Su’s leadership.
Infinity Fabric Architecture
AMD started developing Infinity
Fabric technology around 2012-13,
unveiling it in 2017. They realised
siliconchip.com.au
Fig.28: how Infinity Fabric is used to connect the core complexes (CCDs) and I/O die (cIOD) in a modern AMD processor.
Source: www.igorslab.de/wp-content/uploads/2022/09/zen4guide_slide_4.jpg
that Zen would require a scalable,
modular connection methodology to
make it work. The actual engineering
work was done by Jim Keller and Mark
Papermaster, with Dr Su, then Senior
Vice President (2012-14), providing
executive leadership, engineering
insight and corporate strategy.
Infinity Fabric (IF) is the internal
‘data highway’ for all modern AMD
chips. It is AMD’s proprietary highspeed interconnect technology, introduced in 2017 with the first Zen-based
Ryzen processors (see Fig.28).
It enables modular chiplet-based
designs by providing a means to connect multiple independent silicon
dies, such as CPU and GPU cores, I/O
components, cache memories, memory controllers, AI accelerators and
others, to communicate with each
other at high speeds. It is the main
reason AMD’s chiplet designs are so
effective.
Apart from communication within
the processor module, it also allows
system-wide communication with
peripherals like PCIe cards and USB
devices. It is even scalable to data
centres and exascale supercomputers.
This communication protocol
allows the use of multiple chiplets
in processor designs, which can be
‘mixed and matched’ to produce different variations of processors without having to design new chips from
scratch.
AMD developed Infinity Fabric
to replace HyperTransport, which it
introduced in the early 2000s. While
HyperTransport was advanced for the
siliconchip.com.au
time, it was not flexible enough for the
new chiplet-based design philosophy
AMD wanted to implement.
Infinity Fabric was built to support:
• Multiple chiplets connected
together on the same package
• High-bandwidth and low-latency
communication
• Scalability across desktop, server
and data centre products
Infinity Fabric is a combination of
physical hardware and software-like
communication protocols.
The Physical Layer defines the
wires, circuits and interfaces embedded within the chip to transmit electrical signals. These can be copper or
aluminium wires directly on the silicon chip, on-package traces embedded
in the package substrate that carries
the chiplets, or off-package lanes comprising physical tracks in the motherboard or between CPU or GPU sockets
on a server.
The Control Layer defines the rules.
This is the communication protocol
defining how packets are formatted,
addressed and verified. It also manages cache coherency to ensure that
different parts of the system have the
Table 5: AMD CEOs
CEO
Years in charge
W. J. “Jerry”
Sanders III
1969-2002
Hector Ruiz
2002-2008
Dirk Meyer
2008-2011
Rory Read
2011-2014
Dr Lisa Su
2014-present
Australia's electronics magazine
same idea of what’s in each memory
location despite changes not propagating instantly.
The Control Layer also uses embedded sensors across the silicon chip to
dynamically adjust clock speeds, voltages and data routing based on temperatures and workload demands. The
two main electronic protocols of the
Control Layer are:
1. The Infinity Fabric Data Path
(IFD) is the high-bandwidth pathway
responsible for moving data between
components. It uses high-speed SerDes
(Serialiser/Deserialiser) for converting
between parallel and serial formats. In
Zen-based processors, these links typically run at very high frequencies, up
to 32GT/s (gigatransfers per second) in
recent generations.
2. The Infinity Fabric Control Path
(IFC) handles configuration, synchronisation, power management and coordination between the different dies
and subsystems.
For AMD’s heterogeneous computing philosophy, Infinity Fabric is a key
enabling technology. It allows CPU
cores, GPU cores, I/O dies, memory
controllers and other functional blocks
to exchange data efficiently, both
within a single chip package and, in
some systems, between separate processors or accelerators.
Today, Infinity Fabric is one of
AMD’s most important architectural
advantages. It helped AMD scale EPYC
server processors to high core counts
using chiplets, while also supporting
advanced products such as the AMD
Instinct MI300 series, where CPUs,
December 2026 23
GPUs and high-bandwidth memory
can be tightly coupled inside a single package.
The architecture of Infinity Fabric is
shown in Fig.29. The following abbreviations are used in the diagram:
• 32G SerDes: refers to processors
(like the 4th Gen EPYC series) that support IF lanes/links operating at speeds
up to 32Gbps.
• 2P: two-processor (dual-socket).
A server board setup containing two
physical AMD EPYC CPUs linked via
dedicated IF lanes.
• G Links: dedicated PCIe I/O lanes
originating from the processor that are
explicitly designed to connect GPUs,
storage devices and networking cards.
• P Links: dedicated PCIe I/O lanes
optimised for single-socket (1P) servers. A CPU part number ending in “P”
(eg, EPYC 9654P) is designed to use
all available SerDes links as PCIe I/O
rather than inter-CPU IF.
• D5: a DDR5 (5th generation double data rate) random access memory
(RAM) module.
• 3L/4L: 3-link/4-link specifies how
many physical IF connections are
bound together between computing
resources (such as multiple graphics
dies or separate server dies) to multiply the total bidirectional bandwidth
(up to 400Gbps).
• Fabric + IO: the central I/O die
and Infinity Fabric infrastructure used
to connect the CPU cores, memory
controllers, PCIe/CXL interfaces and
inter-socket links.
• 1DPC: one DIMM per channel,
meaning that each Dual Inline Memory Module has a dedicated communication channel rather than sharing
one between two.
AMD has also worked with industry
partners such as Broadcom to extend
this approach beyond direct chip-tochip links. Broadcom has discussed
support for AMD’s XGMI/Infinity Fabric traffic across future PCI Express
switching technology, including PCIe
Gen 7-era links.
This is associated with AFL, or
Accelerated Fabric Link, a proposed
way to carry AMD fabric traffic, including XGMI, across switched physical
links rather than only direct connections.
Other interconnect
technologies
NVIDIA’s closest competing technology is NVLink, used to connect
GPUs and other processors with a high
bandwidth.
Meanwhile, the broader industry is
developing open interconnect standards such as CXL (Compute Express
Link), UCIe (Universal Chiplet Interconnect Express) and UALink (Ultra
Accelerator Link).
CXL is aimed mainly at coherent
communication between processors,
accelerators and memory devices.
UCIe is designed to connect chiplets
inside a package. UALink, or Ultra
Accelerator Link, is an open accelerator interconnect intended for direct
high-speed communication between
large numbers of AI accelerators in
data centres.
The UALink effort is backed by a
wide group of companies, including
AMD, Apple, AWS, Broadcom, Cisco,
Google, HPE, Intel, Meta, Microsoft
and others. Its goal is to create an
open alternative to proprietary accelerator interconnects such as NVIDIA’s
Fig.29: the Infinity Fabric architecture as used in the AMD EPYC 9004 Genoa. Source: www.servethehome.com/amdepyc-genoa-gaps-intel-xeon-in-stunning-fashion/amd-epyc-9004-genoa-infinity-fabric-overview
24
Silicon Chip
Australia's electronics magazine
siliconchip.com.au
NVLink, reducing dependence on any
single vendor’s ecosystem.
Table 6: AMD’s current products
More on heterogeneous
computing
Product
Heterogeneous?
Features
Ryzen AI
Yes
Ryzen AI 300 series is CPU + GPU +
NPU
Ryzen Desktop/
Mobile APUs
Yes
With Radeon Graphics, CPU + GPU
Instinct MI300/
MI350 series
Yes
Multi-chip CPU + GPU + HBM (High
Bandwidth Memory)
EPYC (many
models)
Highly
High-core count CPUs with integrated
graphics and AI acceleration
Versal Adaptive
(SoCs)
Highly
Arm CPU cores + GPU-like AI engines
+ FPGA fabric
Kria SOMs
Yes
Embedded modules designed for
heterogeneous ‘edge’ AI
Radeon RX
7000/8000 series
No
Discrete graphics cards (shared
driver/software ecosystem with
Ryzen)
Some EPYC and
Ryzen chips
No
Some CPUs lack integrated graphics
AMD later developed the heterogeneous computing idea in several directions: APUs for mainstream PCs and
consoles, Radeon and Instinct accelerators for non-graphics workloads,
and increasingly advanced interconnect technologies like Infinity Fabric.
Chiplets are another major part of
AMD’s modern approach. Rather than
building every processor as one large
monolithic die, AMD can combine
several smaller, specialised dies inside
one package. This improves manufacturing yield, allows different parts of
the processor to be made on different
process technologies, making it easier
to scale core counts.
Intel is also taking a chiplet-like
approach now, albeit much later than
AMD and using different technologies
(we described their system in detail in
the April 2026 issue).
The same broader move toward
heterogeneous and scalable computing is also visible in AMD’s Instinct
MI-
series accelerators. These high-
performance data-centre GPUs are
designed primarily for workloads such
as AI, scientific simulation and high-
performance computing, rather than
conventional graphics output.
So while Intel and AMD both offered
integrated graphics, AMD’s Fusion/
APU strategy was different in emphasis. Intel’s integrated graphics were initially about platform integration and
cost reduction, while AMD wanted
the GPU to become a more powerful
and programmable partner to the CPU.
That vision took many years to
become practical, but it now sits at the
centre of AMD’s processor, accelerator
and chiplet designs.
Today’s computer graphics
market
The computer graphics market
today is dominated by two primary
players: NVIDIA, the clear leader,
and AMD, with Intel trailing as a distant third through its Arc GPU range.
Unfortunately, Arc’s future is looking
uncertain at the moment.
Modern graphics processing units
(GPUs) do far more than simply rendering images on a computer monitor.
They have become the basic hardware
for artificial intelligence and machine
siliconchip.com.au
Table 7: technical terms for modern processor chip designs
Term
Meaning
Usage
APU (Accelerated
Processing Unit)
AMD’s official term for a processor that
tightly integrates CPU cores and GPU
cores (often with an NPU as well) on the
same chip/package.
Primarily
AMD
products
Heterogeneous
processor
Any chip that combines different types
of processors (CPU, GPU, NPU, DSP
[digital signal processor] etc).
Generalpurpose
Heterogeneous
SoC
A System-on-Chip (SoC) that contains
different processing elements.
Generalpurpose
Integrated
Accelerator
Processor
Emphasises the presence of specialised
accelerators (especially artificial
intelligence).
AI chips
learning because they have thousands
of simple processing cores that excel
at performing highly parallel mathematical operations, particularly matrix
multiplications and floating-point
arithmetic.
These operations form the core
workload of neural networks during
both AI training and inferencing tasks.
This massive parallelism gives GPUs a
huge performance advantage over traditional CPUs for AI tasks.
Modern processors increasingly
integrate powerful GPU cores directly
onto the same package or die, making
it less meaningful in many cases to
think of the GPU as a completely separate unit.
While AMD’s current strategy is
based on the idea that future performance gains will come from tightly
integrating different types of processing engines (CPU cores, GPU cores,
Australia's electronics magazine
AI accelerators etc), they still maintain plenty of traditional products
(especially high-end discrete CPUs
and GPUs) to serve specific markets,
generating revenue while the heterogeneous vision matures.
Table 6 indicates the mix of heterogeneous and traditional products. NPU
stands for Neural Processing Unit;
AMD’s name for their NPU architecture is XDNA. For general terminology, see Table 7.
Next month
We will start the third and final part
of this series next month by looking at
the Zen series of chips that has allowed
AMD to take over a significant portion
of the desktop and server CPU market,
as well as make inroads into the mobile
space. We will also give more details on
chiplets, 3D V-Cache, Ryzen AI, Versal,
SC
Radeon/RDNA GPUs and more.
December 2026 25
|