Imitation To Innovation: AMD’s Best CPUs – Part 2

Duron and Sempron: AMD’s Celerons

CPU makers seem to like names that end in “on.” To compete with the Celeron and back up its Athlon, AMD released the Duron, later replaced by the Sempron. These two budget processors were generally slower than the Athlon and had less cache memory. AMD’s exclusive cache design enabled CPUs with an L2 cache that was smaller than the L1, since the latter was not mirrored in the L2 (unlike the inclusive architecture used by Intel). The Sempron is simply a re-named Athlon XP, with certain versions equipped with less cache memory (256 of 512 KB are disabled in the Thorton).

AMD Duron and Sempron

Code name	Spitfire	Thorton
Date released	2000	2004
Architecture	32-bits	32-bits
Data bus	32-bits	32-bits
Address bus	32-bits	32-bits
Maximum memory	4,096 MB	4,096 MB
L1 cache	64 KB + 64 KB	64 KB + 64 KB
L2 cache	64 KB (CPU frequency)	256 KB (CPU frequency)
Clock frequency	600-950 MHz	1,500-2,000 MHz
FSB	100 MHz (DDR)	166 MHz (DDR)
FPU	built-in	built-in
SIMD	MMX, Enhanced 3DNow!	MMX, Enhanced 3DNow!, SSE
Fabrication process	180 nm	130 nm
Number of transistors	25 million	54.3 million
Power consumption	27-41 W	62 W
Voltage	1.5–1.6 V	1.6 V
Die surface area	100 mm²	100.99 mm²
Connector	Socket A	Socket A

In addition to the Spitfire, AMD also released the Duron Morgan (based on the Athlon XP, with SSE support) and the Applebred (130 nm). The Sempron continued its career with the K8 Sempron 3400+, which is a 64-bit Sempron.

The K8: AMD Moves To 64 Bits

K8 was the first x86 processor compatible with 64-bit addressing. The architecture had other advantages such as an integrated memory controller. AMD has released a veritable army of K8-based processors since then, but we’ll concentrate on the models intended for the mainstream: the Athlon 64s. In practice, the Opteron (the server version), Athlon 64 FX (high-end) and Turion 64 (for mobile PCs) are very closely related. In general, they differ only in the management of the memory controller and cache memory, plus the type of memory used.

AMD Athlon 64

Code name	ClawHammer	Orleans
Date released	2003	2006
Architecture	64-bits	64-bits
Data bus	64-bits	64-bits
Address bus	64-bits	64-bits
Maximum memory	1 TB	1 TB
L1 cache	64 KB + 64 KB	64 KB + 64 KB
L2 cache	1,024 KB (CPU frequency)	512 KB (CPU frequency)
Clock frequency	1,800-2,400 MHz	1,800-2,600 MHz
memory controller	DDR-400, 1 channel	DDR2-667, 2 channel
FSB	800 MHz (HTT)	1,000 MHz (HTT)
FPU	built-in	built-in
SIMD	MMX, Enhanced 3DNow!, SSE, SSE2	MMX, Enhanced 3DNow!, SSE, SSE2, SSE3
Fabrication process	130 nm	90 nm
Number of transistors	105.9 million	81.1 million
Power consumption	89 W (TDP)	62 W (TDP)
Voltage	1.5 V	1.25-1.4 V
Die surface area	193 mm²	103 mm²
Connector	Socket 754	Socket AM2

Athlon 64 processors still use a PR number to indicate their ranking in the product range and there are many different versions, which generally differ in terms of cache memory and/or fabrication process. We highlighted only two models, but there are a dozen or so different K8 versions for the standard Athlon 64 alone.

Athlon 64 X2: AMD’s Dual-Core

In 2005, AMD changed its architecture to offer a dual-core version of the K8, and the Athlon 64 X2 was born. Though made up of two K8 cores, the architecture—using a HyperTransport interface—enabled good performance, unlike the solution used by Intel, with the FSB handling communication between the CPUs in its first dual-core processors. The Athlon 64 X2 exists in different sockets and is still on the market (as of August 2008) as an entry-level solution.

AMD Athlon 64 X2

Code name	Toledo	Brisbane
Date released	2005	2006
Architecture	64-bits	64-bits
Data bus	64-bits	64-bits
Address bus	64-bits	64-bits
Maximum memory	1 TB	1 TB
L1 cache	64 KB + 64 KB x 2	64 KB + 64 KB x 2
L2 cache	1,024 KB x 2 (CPU frequency)	512 KB x 2 (CPU frequency)
Clock frequency	2,200-2,400 MHz	1,900-3,100 MHz
Memory controller	DDR-400, 2 channels	DDR2-800, 2 channels
FSB	1,000 MHz (HTT)	1,000 MHz (HTT)
FPU	built-in	built-in
SIMD	MMX, Enhanced 3DNow!, SSE, SSE2, SSE3	MMX, Enhanced 3DNow!, SSE, SSE2, SSE3
Fabrication process	90 nm	65 nm
Number of transistors	233.2 million	153 million
Power consumption	89/110 W (TDP)	65/89 W (TDP)
Voltage	1.35–1.4 V	1.25–1.35 V
Die surface area	199 mm²	126 mm²
Connector	Socket 939	Socket AM2

As for the Athlon 64, we’re only showing two versions of the K8, though other versions exist. Obviously there are server versions (Opteron), high-end versions (Athlon 64 FX) and mobile versions (Turion 64 X2), and also entry-level versions in the form of the Sempron X2. One final anecdote: AMD got away with using the same code name for a processor as Intel had used: the Santa Rosa (a dual-core Opteron manufactured on a 90 nm process).

The Future Lies With Phenom

Now, in mid-2008, it’s no secret that AMD’s CPUs are struggling to keep up on the performance front. But there are a few promising prospects. Early tests of the 45 nm Phenom show interesting results and the Fusion, a cross between a GPU and a CPU, seems to be making progress as well.

Let’s hope that AMD’s financial problems are only temporary, and that they’ll be around for many more years to compete with Intel in the x86 processor arena.

Intel’s 15 Most Unforgettable x86 CPUs-Part 2

Pentium M: Laptops Flex Their Muscles

In 2003 the market for portable PCs was booming and Intel had only two processors for them: the aging Pentium III Tualatin and the Pentium 4, whose high power consumption made it unsuitable. But a savior was to arrive from Israel: the Banias (alias Pentium M). This processor, based on the P6 architecture (the same as the Pentium Pro) had high performance and low power consumption. It even beat the Pentium 4, while consuming a lot less power. This was the processor used in the Centrino platform and it was quickly followed (in 2004) by the (faster) Dothan model. The Pentium M left its mark on the world of mobility, and the Stealey (A100) still uses the Dothan architecture (with lower frequencies and TDP).

Intel Pentium M

Code name	Banias	Dothan
Date released	2003	2004
Architecture	32 bits	32 bits
Data bus	64 bits	64 bits
Address bus	32 bits	32 bits
Maximum memory	4 GB	4 GB
L1 cache	32 KB + 32 KB	32 KB + 32 KB
L2 cache	1,024 KB	2,048 KB
Clock frequency	0.9–1.7 GHz	1–2.13 GHz
FSB	400 MHz	400, 533 MHz
SIMD	MMX, SSE, SSE2	MMX, SSE, SSE2
SMT/SMP	no	no
Fabrication process	130 nm	90 nm
Number of transistors	77 million	140 million
Power consumption	9-30 W	6-35 W
Voltage	0.9–1.5 V	0.9–1.4 V
Die surface area	82 mm²	87 mm²
Connector	Socket 479	Socket 479

As with the Pentium 4, the FSB actually operated at a quarter of the nominal frequency (QDR). The connector used, the Socket 479, actually had 478 pins, but they were arranged differently from the Pentium 4 Socket 478 (though adapters were made).

Pentium 4 Gets 64-bit And Another Core

In 2005, Intel improved its Pentium 4 twice. First, with the Prescott-2M, and then with Smithfield. The former was a 64-bit processor, based on the Prescott design, and the latter was a dual-core processor. They are fairly similar and have the same problems as other Pentium 4s: low instructions per cycle (IPC) throughput and difficulty in increasing the clock frequency due to current losses. These two processors, intended to limit losses while awaiting the Core 2 Duo, are not among Intel’s most highly regarded. And while the Pentium D (the commercial name of the Smithfield) does have two cores, in reality it’s an assembly of two Prescott dies in the same package.

Intel Pentium 4

Code name	Prescott-2M	Smithfield
Date released	2005	2005
Architecture	64 bits	64 bits
Data bus	64 bits	64 bits
Address bus	64 (actual 36) bits	64 (actual 36) bits
Maximum memory	64 GB	64 GB
L1 cache	16 KB + 12 Kµops	2 x 16 KB + 12 Kµops
L2 cache	2,048 KB	2 x 1,024 KB
Clock frequency	3–3.6 GHz	2.8–3.2 GHz
FSB	800 MHz	800 MHz
SIMD	MMX, SSE, SSE2, SSE3	MMX, SSE, SSE2, SSE3
SMT/SMP	Hyper-Threading	dual cores (Hyper-Threading on certain models)
Fabrication process	90 nm	90 nm
Number of transistors	169 million	230 million
TDP	84-115 W	95-130 W
Voltage	1.2 V	1.2 V
Die surface area	135 mm²	206 mm²
Connector	LGA775	LGA775

An interesting point is that whereas the Pentium 4 processors intended for the consumer market did not use the PAE technology (which enables 36-bit, as opposed to 32-bit memory management) and were therefore limited to 4 GB of RAM, these models can go beyond that limit. In practice, the address bus is still limited to 36 bits (40 bits on the Xeon), but PAE (management in 4 GB pages) is now ancient history—a 64-bit program is capable of making full use of the available memory.

Hyper-Threading, an Intel SMT technology, was available on certain models (Xeon and Extreme Edition). Finally, a 65 nm version (the 9×0 series) of the Pentium 4 was released later, but made no major improvements.

The First Mobile Dual-Core

In 2006, Intel announced the Core Duo. The first dual-core processor for portable PCs boasted excellent performance—much better than the Pentium 4. It was also one of the first x86 processors to be truly dual-core. The cache, for example, is shared (whereas the Pentium D was more like an assembly of two processors in the same package). This processor was part of the Centrino Duo platform and was a huge success. The only drawback was that it was still a 32-bit processor, unlike the Pentium 4.

Intel Core Duo

Code name	Yonah
Date released	2006
Architecture	32 bits
Data bus	64 bits
Address bus	32 bits
Maximum memory	4 GB
L1 cache	32 KB + 32 KB
L2 cache	2,048 KB shared
Clock frequency	1.06–2.33 GHz
FSB	667 MHz
SIMD	MMX, SSE, SSE2, SSE3
SMT/SMP	Dual core
Fabrication process	65 nm
Number of transistors	151 million
TDP	9-31 W
Voltage	0.9–1.3 V
Die surface area	91 mm²
Connector	Socket 479

A Core Solo version with one core was also made available, and the low-power-consumption versions used a 533 MHz bus (133 MHz QDR) instead of 667 MHz. This processor was used in servers (code name Sossaman), which was a first for a processor originally intended for the mobile world. Note that this processor didn’t officially use the Core architecture of the Core 2 Duo, and it was quickly replaced by the Core 2 Duo (Merom) in portable PCs. Also, the Yonah’s Socket 479 is different from the Socket 479 of other Pentium M processors.

Today’s Hotness: The Core 2 Duo

In 2006, Intel released a processor that quickly became a best-seller: the Core 2 Duo. Derived from work done on the Pentium M, this processor uses a new Core architecture. Before, Intel had two lines of processors—the Pentium 4 for desktops, Pentium M for mobiles, and both lines for servers. In contrast, Intel now has a single micro-architecture on which all of its product lines draw. The 64-bit Core 2 Duo is represented from the low end to the high end, for desktop computers, portables and servers.

There are many versions of the architecture, resulting in configurations with a different number of cores (one to four, yielding everything from Solos to Quads), cache memory (512 KB to 12 MB), and the FSB (between 400 and 1600 MHz). The model shown here is the original Core 2 Duo, but faster versions (at 45 nm) exist.

Intel Core 2 Duo

Code name	Conroe
Date released	2006
Architecture	64 bits
Data bus	64 bits
Address bus	64 (actual 36) bits
Maximum memory	64 GB
L1 cache	32 KB + 32 KB
L2 cache	2,048 KB shared
Clock frequency	1.8-3 GHz
FSB	800-1066-1333 MHz
SIMD	MMX, SSE, SSE2, SSE3, SSSE3
SMT/SMP	Dual core
Fabrication process	65 nm
Number of transistors	291 million
TDP	65 W
Voltage	1.5 V
Die surface area	143 mm²
Connector	LGA 775

The mobile versions (Merom) are basically identical (but not as fast, with a slower FSB) whereas the Extreme Edition versions are faster. The Core 2 Duo also exists in a four-core version, which was, in fact, two Conroes in the same package. The 45 nm version of the Core 2 Duo (Penryn) has a larger cache and generates less heat, but is still fundamentally similar to this model.

The Future: Nehalem, Atom, Etc.

Obviously, this is only the first part of a series of articles. The second part, on AMD processors, will follow (along with a piece on AMD’s ATI graphics cards). But the story of the Intel x86 processors doesn’t end with the Core 2 Duo, and obviously other models are planned for the future. Nehalem and Atom are also x86 processors. And a little bird tells us that Intel’s upcoming entry into the graphics market, Larrabee, is also based on a number of x86 cores.

Imitation To Innovation: AMD’s Best CPUs – Part 1

AMD Clones Intel

The year is 1981, and Intel (see history of Intel processors) has just been chosen by IBM to supply the processor for the first personal computer. IBM wanted at least two CPU suppliers for its PC, and forced Intel to license its technology. And so it was that AMD became one of the first companies to sell an 8086 clone. AMD’s first processor went on sale in 1982. Because it was a licensed processor, the AMD 8086 (and 8088) was identical to Intel’s model.

AMD 8086

Code name	?
Date released	1982
Architecture	16-bits
Data bus	16-bits
Address bus	20-bits
Maximum memory	1 MB
L1 cache	no
L2 cache	no
Clock frequency	5-10 MHz
FSB	same as clock frequency
FPU	8087
SIMD	no
Fabrication process	3,000 nm
Number of transistors	29,000
Power consumption	?
Voltage	5 V
Die surface area	16 mm²
Connector	40 pins

Note the “© Intel” on the processor, made by AMD.

Am286: Manufactured Under License, But Faster

AMD’s Am286, a clone of the Intel 80286 manufactured under license, was identical to the chip from Intel, but it had a big advantage: its higher clock speed. Whereas Intel’s 286s topped out at 12.5 MHz, AMD sold 20 MHz versions. Because the 286 was more economical than the 386, whose innovations weren’t fully exploited for several years, AMD was already the value choice more than 20 years ago.

Am286

Code name	?
Date released	1983
Architecture	16-bits
Data bus	16-bits
Address bus	24-bits
Maximum memory	16 MB
L1 cache	no
L2 cache	no
Clock frequency	8-20 MHz
FSB	same as clock frequency
FPU	80287
SIMD	no
Fabrication process	1,500 nm
Number of transistors	134,000
Power consumption	?
Voltage	5 V
Die surface area	49 mm²
Connector	68 pins

Am386: A 40-MHz 386

In 1991, AMD released its 386 processor. Like its predecessors, this model was identical to the Intel versions. AMD was licensed to produce clones of Intel products, right down to the microcode (the CPU’s firmware). This processor had two notable features. First, it was faster than the Intel model—40 MHz compared to a top speed of 33 MHz at Intel—and it was the first to sport the Windows Compatible logo on the package.

Am386

Code name	?
Date released	1991
Architecture	32-bits
Data bus	32-bits
Address bus	32-bits
Maximum memory	4,096 MB
L1 cache	no
L2 cache	no
Clock frequency	12-40 MHz
FSB	same as clock frequency
FPU	80387
SIMD	no
Fabrication process	1,500 – 1,000 nm
Number of transistors	275,000
Power consumption	2 W (33 MHz)
Voltage	5 V
Die surface area	42 mm²
Connector	132 pins

Am486: The Last Clone

The 486 was the last clone of an Intel processor. AMD produced 486s in two different versions—one with microcode by Intel and another with microcode by AMD, because the company was having legal hassles with Intel by that point. In addition to processors sold under the 486 designation, AMD also marketed an AMD 5×86, which was a 486 with a 4x clock multiplier. Running at 133 MHz, this model was compatible with 486 motherboards, but had the performance of a Pentium 75. It was with the 5×86 that AMD began using the famous “Pentium Rating” (5×86 PR 75), which it would stay with up to and including the Athlon 64 X2.

Am486 / 5×86

Code name	?	X5
Date released	1993	1995
Architecture	32-bits	32-bits
Data bus	32-bits	32-bits
Address bus	32-bits	32-bits
Maximum memory	4,096 MB	4,096 MB
L1 cache	8 KB	16 KB
L2 cache	motherboard (FSB frequency)	motherboard (FSB frequency)
Clock frequency	16-120 MHz	133 MHz
FSB	16-50 MHz	33 MHz
FPU	built-in	built-in
SIMD	no	no
Fabrication process	1,000 – 800 nm	350 nm
Number of transistors	1,185,000	?
Power consumption	?	?
Voltage	5 V–3.3 V	3.45 V
Die surface area	81 – 67 mm²	?
Connector	168 pins	168 pins

The K5: AMD’s Very Own Processor

In 1996, AMD released its fifth-generation processor, the K5. Compared to Intel’s Pentium, the K5 was technically more advanced, though it did have some faults. It’s especially interesting because of its RISC-based internal architecture that decoded x86 instructions into micro-instructions before executing them. The K5 had difficulty reaching high clock speeds and its FPU was a little weak. Still, in normal use, the K5 was a better performer than the Pentium and its PR was not just hype—a K5 clocked at 100 MHz was sold as a PR133 chip, meaning that AMD considered it as being equivalent in performance to a 133 MHz Pentium.

AMD K5

Code name	SSA/5, 5k86
Date released	1996
Architecture	32-bits
Data bus	64-bits
Address bus	32-bits
Maximum memory	4,096 MB
L1 cache	16 KB + 8 KB
L2 cache	motherboard (FSB frequency)
Clock frequency	75-133 MHz (PR75 – PR200)
FSB	50-66 MHz
FPU	built-in
SIMD	no
Fabrication process	500 – 350 nm
Number of transistors	4.3 million
Power consumption	11-16 W
Voltage	3.52 V
Die surface area	251 – 181 mm²
Connector	Socket 5 or 7

The use of the PR resulted in such oddities as a K5 PR90 and PR120 running at the same frequency (90 MHz) and a PR100 and PR133 both clocked at 100 MHz. Notice also that the CPU package informed buyers that a heat sink and fan were required—at that time, the use of such cooling devices was not yet common practice.

The K6: AMD Extends Its Range

In 1997, AMD released a new processor: the K6. Unlike the K5, which was created by AMD, the K6 was the result of the work done by NexGen on the Nx686. This processor was compatible with Socket 7 (Pentium) motherboards and offered very good performance compared to Intel’s Pentium II processors, at a much lower price. The K6’s FPU was still a little weak compared to Intel’s. A 250 nm version of the K6, called Little Foot, came out in 1998.

Also in 1998, AMD announced the K6-2, a processor that used a faster bus (100 MHz) and had improved SIMD performance. It also had one more MMX unit than the K6 and a new instruction set, 3DNow!, for floating-point calculations (MMX handled only integers). The K6-2 (400 and up) was a big success because it was a good upgrade solution for owners of Pentium MMX platforms—by using the 2X multiplier on a motherboard with a 66 MHz bus, the processor was in fact operating at 6X (400 MHz), which permitted a significant gain in speed at a lower upgrade cost.

Finally, in 1999, AMD released the third version of the K6, the K6-III. The main difference from the K6-2 version was an on-chip 256 KB cache. The K6-III was very fast, but also very costly to produce, and was quickly replaced by the Athlon (K7).

AMD K6, K6-2, K6-III

Code name	K6, Little Foot (250 nm)	K6-3D, Chomper	Sharptooth
Date released	1997/1998	1998	1999
Architecture	32-bits	32-bits	32-bits
Data bus	64-bits	64-bits	64-bits
Address bus	32-bits	32-bits	32-bits
Maximum memory	4,096 MB	4,096 MB	4,096 MB
L1 cache	32 KB + 32 KB	32 + 32 KB	32 + 32 KB
L2 cache	motherboard (FSB frequency)	motherboard (FSB frequency)	256 KB (CPU frequency)
L3 cache	no	no	motherboard (FSB frequency)
Clock frequency	166-300 MHz	300-550 MHz	400-450 MHz
FSB	50-66 MHz	66-100 MHz	100 MHz
FPU	built-in	built-in	built-in
SIMD	MMX	MMX, 3DNow!	MMX, 3DNow!
Fabrication process	350 – 250 nm	250 nm	250 nm
Number of transistors	8.8 million	9.3 million	21.3 million
Power consumption	12-28 W	13-25 W	10-17 W
Voltage	2.2–2.9 V–3.2 V	2.2–2.4 V	2.2–2.4 V
Die surface area	157-68 mm²	81 mm²	118 mm²
Connector	Socket 7	Socket 7 / Super Socket 7	Super Socket 7

AMD also marketed K6-2+ and K6-3+ processors, mainly for portable PCs. These used a 180 nm fab process and had an on-chip 128 KB (K6-2+) or 256 KB (K6-3+) L2 cache.

K7/Athlon: A Killer

In 1999, AMD released its seventh-generation processor, the K7, later renamed Athlon. This chip did away with the drawbacks of earlier models and finally had an FPU worthy of the name—in fact, it was even better than Intel’s. The Athlon was the fastest x86 processor and had many strong points, including a fast FSB—the EV6, used in the first Alpha processors—and high performance numbers. The only real problem came not from the processor but from the chipsets: neither the AMD nor Via models could compete with Intel’s chipsets (like the famous 440BX). The K7 used Slot A (competing with Intel’s Slot 1) and had a Level 2 cache with a variable divider (1/2, 2/5 or 1/3).

AMD Athlon (K7, K75)

Code name	Argon (K7)	Pluto, Orion (K75)
Date released	1999	1999
Architecture	32-bits	32-bits
Data bus	64-bits	64-bits
Address bus	32-bits	32-bits
Maximum memory	4,096 MB	4,096 MB
L1 cache	64 KB + 64 KB	64 KB + 64 KB
L2 cache	Slot A (1/2 CPU)	Slot A (1/2, 2/5 or 1/3 CPU)
Clock frequency	500-700 MHz	550-1000 MHz
FSB	100 MHz (DDR)	100 MHz (DDR)
FPU	built-in	built-in
SIMD	MMX, Enhanced 3DNow!	MMX, Enhanced 3DNow!
Fabrication process	250 nm	180 nm
Number of transistors	22 million	22 million
Power consumption	42-50 W	31-65 W
Voltage	1.6 V	1.6–1.8 V
Die surface area	184 mm²	102 mm²
Connector	Slot A	Slot A

Just as a side note, it was AMD who was the first to announce (and market) a 1 GHz processor with the Athlon (two days before Intel’s 1 GHz Pentium III).

AMD Improves the Athlon: Thunderbird, XP, and more.

AMD knew it had a winner with the K7 architecture and improved it little by little, increasing the frequency and using finer fab processes. The Thunderbird core employed a 180 nm process and had 256 KB of on-chip cache. The Palomino design introduced support for SSE. The Athlon XP changed the package and reinstated PR numbers. The Thoroughbred was an Athlon XP using a 130 nm fab process (with a 256 KB cache). Barton had a 512 KB cache and also used a 130 nm process. Athlon XP and subsequent models used the PR number instead of a clock frequency designation.

AMD Athlon

Code name	Thunderbird	Palomino/XP	Thoroughbred	Barton
Date released	2000	2001	2002	2003
Architecture	32-bits	32-bits	32-bits	32-bits
Data bus	64-bits	64-bits	64-bits	64-bits
Address bus	32-bits	32-bits	32-bits	32-bits
Maximum memory	4,096 MB	4,096 MB	4,096 MB	4,096 MB
L1 cache	64 KB + 64 KB	64 KB + 64 KB	64 KB + 64 KB	64 KB + 64 KB
L2 cache	256 KB (CPU frequency)	256 KB (CPU frequency)	256 KB (CPU frequency)	512 KB (CPU frequency)
Clock frequency	650-1,400 MHz	1,000-1,733 MHz	1,200-2,250 MHz	1,400-2,200 MHz
FSB	100/133 MHz (DDR)	133 MHz (DDR)	133/166 MHz (DDR)	166/200 MHz (DDR)
FPU	built-in	built-in	built-in	built-in
SIMD	MMX, Enhanced 3DNow!	MMX, Enhanced 3DNow!, SSE	MMX, Enhanced 3DNow!, SSE	MMX, Enhanced 3DNow!, SSE
Fabrication process	180 nm	180 nm	130 nm	130 nm
Number of transistors	37 million	37.5 million	37.2 million	54.3 million
Power consumption	38-72 W	46-72 W	49-68 W	60-76 W
Voltage	1.7-1.75 V	1.75 V	1.5-1.65 V	1.65 V
Die surface area	120 mm²	129.26 mm²	84.66 mm²	100.99 mm²
Connector	Socket A	Socket A	Socket A	Socket A

We should mention that AMD also produced versions for servers (Athlon MP) and for laptops (Athlon 4, Athlon XP Mobile), as well as the Geode NX (130 nm and a 256 KB cache). AMD marketed the Thorton (130 nm, 512 KB of cache, 256 KB of which was disabled) and planned on Trinidad, an Athlon using a 90 nm process. There were more PR oddities: the Athlon XP 2600+ was clocked at 1,900, 1,917, 2,000, 2,083, or 2,133 MHz depending on the version, for instance

Intel’s 15 Most Unforgettable x86 CPUs-Part 1

8086: The First PC processor

The 8086 was the first x86 processor—Intel had already released the 4004, the 8008, the 8080 and the 8085. This 16-bit processor could manage 1 MB of memory using an external 20-bit address bus. The clock frequency chosen by IBM (4.77 MHz) was fairly low, though the processor was running at 10 MHz by the end of its career.

The first PCs used a derivative of this processor, the 8088, which had only an 8-bit (external) data bus. An interesting aside is that the control systems in the US space shuttles use 8086 processors and NASA was forced to buy some from eBay in 2002 since Intel could no longer supply them.

Intel 8086

Code name	N/A
Date released	1979
Architecture	16 bits
Data bus	16 bits
Address bus	20 bits
Maximum memory	1 MB
L1 cache	no
L2 cache	no
Clock frequency	4.77-10 MHz
FSB	same as clock frequency
FPU	8087
SIMD	no
Fabrication process	3,000 nm
Number of transistors	29,000
Power consumption	N/A
Voltage	5 V
Die surface area	16 mm²
Connector	40-pin

80286: 16 MB Of Memory, But Still 16 Bits

Released in 1982, the 80286 was 3.6 times faster than the 8086 at the same frequency. It could manage up to 16 MB of memory, but the 286 was still a 16-bit processor. It was the first x86 equipped with a memory management unit (MMU), allowing it to manage virtual memory. Like the 8086, it did not have a floating-point unit (FPU), but could use a x87 co-processor chip (80287). Intel offered these processors at a maximum frequency of 12.5 MHz, whereas their competitors reached 25 MHz.

Intel 80286

Code name	N/A
Date released	1982
Architecture	16 bits
Data bus	16 bits
Address bus	24 bits
Maximum memory	16 MB
L1 cache	No
L2 cache	No
Clock frequency	6–12 MHz
FSB	same as clock frequency
FPU	80287
SIMD	No
Fabrication process	1,500 nm
Number of transistors	134,000
Power consumption	N/A
Voltage	5 V
Die surface area	49 mm²
Connector	68-pin

386: 32-Bit and Cache Memory

Intel’s 80836 was the first x86 with a 32-bit architecture. Several versions of this processor were offered. The two best known are the 386 SX (Single-word eXternal), which had a 16-bit data bus, and the 386 DX (Double-word eXternal) with a 32-bit data bus. Two other versions are worth noting, though: the SL, which was the first x86 to offer management of a cache (external) and the 386EX, used in the space program (the Hubble telescope uses this processor).

Intel 80386 DX

Code name	P3
Date released	1985
Architecture	32 bits
Data bus	32 bits
Address bus	32 bits
Maximum memory	4096 MB
L1 cache	0 KB (controller sometimes present)
L2 cache	no
Clock frequency	16-33 MHz
FSB	same as clock frequency
FPU	80387
SIMD	no
Fabrication process	1,500-1,000 nm
Number of transistors	275,000
Power consumption	2 W @ 33 MHz
Voltage	5 V
Die surface area	42 mm² @ 1µ
Connector	132 pins

The 486: An FPU And Multipliers Too

The 486 is emblematic of a certain generation who were first discovering computers. In fact, the very famous 486 DX2/66 was long considered the minimum configuration for gamers. This processor, released in 1989, ushered in several interesting new features, like an on-chip FPU, data cache, and the first clock multiplier. The former consisted of an x87 coprocessor built into the 486 DX (not SX) series. An 8 KB Level 1 cache was built into the processor (write-through type, then write-back with slightly better performance). There was also the possibility of a Level 2 cache on the motherboard (at the bus frequency).

The second generation of 486s had a CPU multiplier, since the processor operated faster than the FSB, with DX2 (2x multiplier) and DX4 (3x multiplier) versions. Another anecdote: the “487SX” sold as an FPU for the 486SX was actually a full 486DX that disabled and took the place of the first processor.

Intel 80486 DX

Code name	P4, P24, P24C
Date released	1989
Architecture	32 bits
Data bus	32 bits
Address bus	32 bits
Maximum memory	4096 MB
L1 cache	8 KB
L2 cache	Motherboard (FSB frequency)
Clock frequency	16-100 MHz
FSB	16-50 MHz
FPU	On chip
SIMD	No
Fabrication process	1,000–800 nm
Number of transistors	1,185,000
Power consumption	N/A
Voltage	5 V–3.3 V
Die surface area	81 – 67 mm²
Connector	168 pins

The DX4 had a 16 KB cache and a few more transistors: 1.6 million. This processor, using a 600 nm process and measuring 76 mm², consumed less power than the original 486 (at a voltage of 3.3 V).

Intel Pentium: A Bothersome Bug

The Pentium, introduced in 1993, was interesting for more than one reason. It was the first x86 to drop the traditional model number for a more attractive name, since Intel wasn’t allowed to trademark a name made up of numbers only. It’s also famous because of a bug it contained. On the first generations of Pentiums, certain division operations produced an incorrect result. Intel replaced the processors, but the damage was done. A very rare error gave rise to the first big IT media buzz.

The Pentium was sold in three different versions, the first without a CPU multiplier, the second with a multiplier (including the very familiar Pentium 166), and the last with the SIMD instruction set for x86s, MMX. The Pentium MMX also increased the size of the Level 1 cache and brought in a few minor improvements. This was the first Intel x86 capable of executing two instructions in parallel. The L2 cache was on the motherboard with these processors (running at the frequency of the FSB).

Intel Pentium (MMX)

Code name	P5, P54	P55 (Pentium MMX)
Date released	1993	1997
Architecture	32 bits	32 bits
Data bus	64 bits	64 bits
Address bus	32 bits	32 bits
Maximum memory	4096 MB	4096 MB
L1 cache	8 KB + 8 KB	16 KB + 16 KB
L2 cache	Motherboard (FSB frequency)	Motherboard (FSB frequency)
Clock frequency	60-200 MHz	133-300 MHz
FSB	50-66 MHz	60-66 MHz
FPU	on chip	on chip
SIMD	no	MMX
Fabrication process	800-600-350 nm	350 nm
Number of transistors	3.1-3.3 million	4.5 million
Power consumption	8-16 W	4-17 W
Voltage	5 V-3.3 V	2.8 V
Die surface area	294-163-90 mm²	141 mm²
Connector	Socket 4, 5 or 7	Socket 7

Here’s a little explanation of the Pentium bug: certain calculations using the FPU resulted in erroneous results. This was fairly rare—though sources disagree about exactly how rare—and Intel replaced the defective processors free of charge. Here’s an example of a Pentium error:

4195835.0/3145727.0 = 1.333 820 449 136 241 002 (correct result) 4195835.0/3145727.0 = 1.333 739 068 902 037 589 (incorrect result on a defective Pentium)

Pentium Pro: The First To Handle Over 4 GB Of Memory

The Pentium Pro, released in 1995, was the first x86 CPU able to manage more than 4 GB of RAM using Physical Address Extension (PAE), 36-bit address size, and thus 64 GB. An interesting point is that this processor was also the first P6 (the architecture the Core 2 processors are loosely derived from) and also the first x86 to include a Level 2 cache on the processor instead of on the motherboard. In fact, between 256 KB and 1 MB of cache were placed next to the CPU, on the same socket, making the L2 cache on-package as opposed to on-chip, clocked at the same frequency as the CPU.

This processor also had a bit of a performance issue. It ran great in 32-bit applications, but was much slower with software still written in 16 bits (like Windows 95). The cause was simple: access to 16-bit registers caused problems with management of the (32-bit) registers, which canceled out the advantages of the Pentium Pro’s out-of-order architecture.

Intel Pentium Pro

Code name	P6
Date released	1995
Architecture	32 bits
Data bus	64 bits
Address bus	36 bits
Maximum memory	64 GB
L1 cache	8 KB + 8 KB
L2 cache	external, 256-1024 KB (CPU frequency)
Clock frequency	150-200 MHz
FSB	60-66 MHz
FPU	built-in
SIMD	N/A
Fabrication process	600-350 nm
Number of transistors	5,500,000 + cache
Power consumption	29-47 W
Voltage	3.3 V
Die surface area	306-196 mm² + cache
Connector	Socket 8

The cache measured 202 mm² (256 KB at 500 nm), 242 mm² (512 KB at 350 nm), or 484 mm² (1 MB at 350 nm). The number of transistors in the cache was 15.5 million (256 KB), 31 million (512 KB), or 62 million (1 MB).

Pentium II and III: Brothers

Released in 1997, the Pentium II was an adaptation of the Pentium Pro aimed at the general public. It was quite similar to the Pentium Pro, but the cache memory was different. Instead of using a cache at the same frequency as the processor (which is expensive), the 512 KB Level 2 cache operated at half-frequency. In addition, the Pentium II abandoned the classic socket for a cartridge containing the processor and the Level 2 cache, which was in the cartridge and not on the motherboard or in the processor itself.

New features compared to the Pentium Pro were essentially MMX (SIMD) support and a doubling of the Level 1 cache. The first Pentium III (Katmai) was very similar to the Pentium II. Released in 1999, its new feature was essentially support for SSE (SIMD instructions), but the rest was identical.

Intel Pentium II and III

Code name	Klamath (Pentium II 0.35µ), Deschutes (Pentium II 0.25µ), Katmai (Pentium III)
Date released	1997, 1998, 1999
Architecture	32 bits
Data bus	64 bits
Address bus	36 bits (32 bits on the P III)
Maximum memory	64 GB (4 GB on the P III)
L1 cache	16 KB + 16 KB
L2 cache	external, 512 KB (1/2 CPU frequency)
Clock frequency	233-300 MHz (Klamath), 300-450 MHz (Deschutes), 450-600 MHz (Klamath)
FSB	66-100-133 MHz
FPU	built-in
SIMD	MMX (SSE)
Fabrication process	350 nm (Klamath), 250 nm (Deschutes, Katmai)
Number of transistors	7,500,000 + cache (Pentium II), 9,500,000 + cache (Pentium III)
Power consumption	25-35 W
Voltage	2.8 V (0.35µ), 2 V (0.25µ)
Die surface area	204 mm² (0.35µ), 131 mm² (0.25µ), 128 mm² (PIII) + cache
Connector	Slot 1

The Pentium II and III had 512 KB of Level 2 cache (31 million transistors). One Pentium II actually had an on-chip 256 KB Level 2 cache—the Pentium II Mobile Dixon. Using a 180 nm fabrication process, this processor was significantly faster than the desktop versions

Celeron and Xeon: Intel Aims At The High/Low End

At the end of the 1990s, Intel launched two of its best-known processor brands: Celeron and Xeon. The former was aimed at the budget market and the latter at servers, and sometimes workstations. The first Celeron (Covington) was a Pentium II without a Level 2 cache, and suffered extremely poor performance, whereas the Pentium II Xeon had a large cache. Even now, both brands still exist—Celeron for the entry-level market (generally with a reduced cache and a slower FSB) and Xeon for servers (with a fast FSB, sometimes more cache, and high clock speeds).

Intel quickly added a cache to the Celeron with the Mendocino model (128 KB). The Celeron 300A is famous for its overclocking capacities, able to go 50% or more above its rated clock speed much of the time.

Intel Celeron and Intel Xeon

Code name	Covington, Mendocino	Drake
Date released	1998	1998
Architecture	32 bits	32 bits
Data bus	64 bits	64 bits
Address bus	32 bits	36 bits
Maximum memory	4 GB	64 GB
L1 cache	16 KB + 16 KB	16 KB + 16 KB
L2 cache	0 KB/128 KB (internal, CPU frequency)	external, 512 KB-2,408 KB (CPU frequency)
Clock frequency	266-300 MHz/300-533 MHz	400-450 MHz
FSB	66 MHz	100 MHz
FPU	built in	built in
SIMD	MMX	MMX
Fabrication process	250 nm	250 nm
Number of transistors	7,500,000/19,000,000	7,500,000 + cache
Power consumption	16–28 W	30-46 W
Voltage	2 V	2 V
Die surface area	131 mm²/154 mm²	131 mm² + cache
Connector	Slot1/Socket 370 PPGA	Slot 2

Like the Pentium II, Xeon had an external L2 cache inside the processor cartridge. Its capacity was between 512 KB and 2 MB, and the number of transistors between 31 million and 124 million.

The Pentium III Hits 1 GHz

The Pentium III Coppermine was the first commercial x86 processor from Intel to attain a clock speed of 1 GHz; a 1.13 GHz version was even released, but was quickly taken off the market because it was unstable. This new version of the Pentium III improved the Level 2 cache—now on-die. It was faster than the 512 KB external cache on the first model and was touted as a feature able to speed up the Internet experience. It was released in three versions: server (Xeon), entry-level (Celeron), and mobile (with the first version of SpeedStep).

Intel Pentium III

Code name	Coppermine
Date released	1999
Architecture	32 bits
Data bus	64 bits
Address bus	32 bits
Maximum memory	4 GB
L1 cache	16 KB + 16 KB
L2 cache	internal, 256 KB (CPU frequency)
Clock frequency	500–1,133 MHz
FSB	100-133 MHz
FPU	built in
SIMD	MMX (SSE)
Fabrication process	180 nm
Number of transistors	28.1 million
Power consumption	25-35 W
Voltage	1.6 V, 1.8 V
Die surface area	106 mm²
Connector	Slot 1-Socket 370 FCPGA

A slightly improved version (Tualatin), with more L2 cache (512 KB) and centering on a 130 nm process, was released in 2002. Essentially intended for servers (PIII-S) and mobile devices, it was less common in consumer-level machines.

The Pentium 4: A Lot Of Noise Over Very Little

In November 2000, Intel announced its new processor, the Pentium 4. With a higher clock speed (at least 1,400 MHz), this processor had a major drawback in that its performance wasn’t as good as competing models on a per-clock basis. AMD’s Athlon (and even the Pentium III) performed better at the same frequency. Complicating matters, Intel tried to shift to Rambus’ RDRAM memory (the only memory at the time capable of meeting the requirements of the CPU’s FSB), but failed. Expensive and hot, the Pentium 4 nonetheless managed, with many modifications, to more or less stay in the competition for a few years (by adding L3 cache and technologies like Hyper-Threading).

Intel Pentium 4 32-bit

Code name	Willamette	Northwood	Prescott
Date released	2000	2001	2004
Architecture	32 bits	32 bits	32 bits
Data bus	64 bits	64 bits	64 bits
Address bus	32 bits	32 bits	32 bits
Maximum memory	4 GB	4 GB	4 GB
L1 cache	8 KB + 12 Kµops	8 KB + 12 Kµops	16 KB + 12 Kµops
L2 cache	256 KB	512 KB	1,024 KB
Clock frequency	1.3-2 GHz	1.8–3.4 GHz	2.4–3.8 GHz
FSB	400 MHz	400, 533, 800 MHz	533, 800 MHz
SIMD	MMX, SSE, SSE2	MMX, SSE, SSE2	MMX, SSE, SSE2, SSE3
SMT/SMP	no	Hyper-Threading (certain versions)	Hyper-Threading
Fabrication process	180 nm	130 nm	90 nm
Number of transistors	42 million	55 million	125 million
Power consumption	66-100 W	54-137 W	94-151 W
Voltage	1.7 V	1.55 V	1.25–1.5 V
Die surface area	217 mm²	146 mm²	112 mm²
Connector	Socket 423/Socket 478	Socket 478	Socket 478/LGA775

Mobile versions (with a variable multiplier), Celeron versions (with a smaller L2 cache), and Xeon versions (with an L3 cache) of the Pentium 4 were sold. Hyper-Threading and the L3 cache are two technologies that first appeared on servers and were then adapted to standard processors (though L3 cache was available only on the expensive EE models).

We should also mention the FSB, which was clocked at a fourth of the nominal clock frequency, using what is called Quad Data Rate (QDR) technology—a 400 MHz bus is actually 100 MHz QDR, 533 MHz is 133 MHz QDR, etc. Finally, 64-bit versions of the Pentium 4 appeared in 2005, which we’ll talk about later on.

AMD ATI Comparison Table

With more and more graphics chips being released every day it became very complicated for the user who does not follow the video card market to know the differences among all ATI graphic chips in the market today. To facilitate knowing and understanding the difference among major ATI chips, we have compiled the following table.

It is important to notice that starting 2007 both ATI and nVidia started referring to the memory clock of their video cards with the real clock rate used. In the past manufacturers referred the memory clocks with double their real clock rate, because DDR and subsequent technologies (DDR2, GDDR3, etc) allow the memory chip to transfer two data per clock cycle. So a video card with a memory chip running at 500 MHz would be referred as having a 1 GHz memory. In order to keep the compatibility of our table, we are still referring the memory clocks with the DDR naming convention – i.e. double the real clock rate – on cards with memories based on DDR or subsequent technologies.

Chip	Core Clock	Memory Clock	Memory Interface	Memory Transfer Rate	Pixels per clock	DirectX
Radeon 9200	250 MHz	400 MHz	128-bit	6.4 GB/s	4	8.1
Radeon 9200 Pro	275 MHz	550 MHz	128-bit	8.8 GB/s	4	8.1
Radeon 9200 SE	200 MHz	333 MHz	64-bit	2.6 GB/s	4	8.1
Radeon 9250	240 MHz	400 MHz	128-bit	6.4 GB/s	4	8.1
Radeon 9250 SE	240 MHz	400 MHz	64-bit	3.2 GB/s	4	8.1
Radeon 9500	275 MHz	540 MHz	128-bit	8.6 GB/s	4	9.0
Radeon 9550	250 MHz	400 MHz	128-bit	6.4 GB/s	4	9.0
Radeon 9550 SE	250 MHz	400 MHz	64-bit	3.2 GB/s	4	9.0
Radeon 9500 Pro	275 MHz	540 MHz	128-bit	8.6 GB/s	8	9.0
Radeon 9600	325 MHz	400 MHz	128-bit	6.4 GB/s	4	9.0
Radeon 9600 Pro	400 MHz	600 MHz	128-bit	9.6 GB/s	4	9.0
Radeon 9600 SE	325 MHz	400 MHz	64-bit	3.2 GB/s	4	9.0
Radeon 9600 XT	500 MHz	600 MHz	128-bit	9.6 GB/s	4	9.0
Radeon 9700	275 MHz	540 MHz	256-bit	17.2 GB/s	8	9.0
Radeon 9700 Pro	325 MHz	620 MHz	256-bit	19.8 GB/s	8	9.0
Radeon 9800	325 MHz	580 MHz	256-bit	18.56 GB/s	8	9.0
Radeon 9800 Pro	380 MHz	680 MHz	256-bit	21.7 GB/s	8	9.0
Radeon 9800 SE	325 MHz	500 MHz	128-bit or 256-bit	8 GB/s or 16 GB/s	4	9.0
Radeon 9800 XT	412 MHz	730 MHz	256-bit	23.3 GB/s	8	9.0
Radeon X300 SE	325 MHz	400 MHz	64-bit	3.2 GB/s	4	9.0
Radeon X300	325 MHz	400 MHz	128-bit	6.4 GB/s	4	9.0
Radeon X550	400 MHz	500 MHz	128-bit or 64-bit	8 GB/s or 4 GB/s	4	9.0
Radeon X600 Pro	400 MHz	600 MHz	128-bit	9.6 GB/s	4	9.0
Radeon X600 XT	500 MHz	730 MHz	128-bit	11.68 GB/s	4	9.0
Radeon X700	400 MHz	600 MHz	128-bit	9.6 GB/s	8	9.0
Radeon X700 Pro	420 MHz	864 MHz	128-bit	13.8 GB/s	8	9.0
Radeon X700 XT	475 MHz	1.05 GHz	128-bit	16.8 GB/s	8	9.0
Radeon X800 SE	*	*	*	*	8	9.0
Radeon X800	400 MHz	700 MHz	256-bit	22.4 GB/s	12	9.0
Radeon X800 XL	400 MHz	1 GHz	256-bit	32 GB/s	16	9.0
Radeon X800 GT	475 MHz	**	128-bit or 256-bit	**	8	9.0
Radeon X800 GTO	400 MHz	1 GHz ***	256-bit	32 GB/s	12	9.0
Radeon X800 Pro	475 MHz	950 MHz	256-bit	30.4 GB/s	12	9.0
Radeon X800 XT	500 MHz	1 GHz	256-bit	32 GB/s	16	9.0
Radeon X800 XT PE	520 MHz	1.12 GHz	256-bit	35.8 GB/s	16	9.0
Radeon X850 Pro	520 MHz	1.08 GHz	256-bit	34.56 GB/s	12	9.0
Radeon X850 XT	520 MHz	1.08 GHz	256-bit	34.56 GB/s	16	9.0
Radeon X850 PE	540 MHz	1.18 GHz	256-bit	37.76 GB/s	16	9.0
Radeon X1050	****	****	****	****	4	9.0c
Radeon X1300 HM	450 MHz	1 GHz	128-bit or 64-bit or 32-bit	16 GB/s or 8 GB/s or 4 GB/s	4	9.0c
Radeon X1300	450 MHz	500 MHz	128-bit or 64-bit or 32-bit	8 GB/s or 4 GB/s or 2 GB/s	4	9.0c
Radeon X1300 Pro	600 MHz	800 MHz	128-bit or 64-bit or 32-bit	12.8 GB/s or 6.4 GB/s or 3.2 GB/s	4	9.0c
Radeon X1300 XT	500 MHz	800 MHz (DDR2) or 1 GHz (GDDR3)	128-bit	12.8 GB/s or 16 GB/s	12	9.0c
Radeon X1550	450 MHz or 550 MHz or 600 MHz	800 MHz	64-bit or 128-bit	6.4 GB/s or 12.8 GB/s	4	9.0c
Radeon X1600 Pro	500 MHz or 575 MHz	780 MHz	128-bit	12.48 GB/s	12	9.0c
Radeon X1600 XT	590 MHz	1.38 GHz	128-bit	22.08 GB/s	12	9.0c
Radeon X1650 Pro	600 MHz	1.40 GHz	128-bit	22.40 GB/s	12	9.0c
Radeon X1650 XT	575 MHz	1.35 GHz	128-bit	21.60 GB/s	24	9.0c
Radeon X1800 GTO	500 MHz	1 GHz	256-bit	32 GB/s	12	9.0c
Radeon X1800 XL	500 MHz	1 GHz	256-bit	32 GB/s	16	9.0c
Radeon X1800 XT	625 MHz	1.5 GHz	256-bit	48 GB/s	16	9.0c
Radeon X1900 GT	575 MHz	1.2 GHz	256-bit	38.4 GB/s	36	9.0c
Radeon X1900 XT	625 MHz	1.45 GHz	256-bit	46.4 GB/s	48	9.0c
Radeon X1900 XTX	650 MHz	1.55 GHz	256-bit	49.6 GB/s	48	9.0c
Radeon X1950 GT	500 MHz	1.2 GHz	256-bit	38.4 GB/s	36	9.0c
Radeon X1950 Pro	575 MHz	1.38 GHz	256-bit	44.16 GB/s	36	9.0c
Radeon X1950 XT	625 MHz	1.8 GHz	256-bit	57.6 GB/s	48	9.0c
Radeon X1950 XTX	650 MHz	2 GHz	256-bit	64 GB/s	48	9.0c
Radeon HD 2400 Pro	525 MHz	800 MHz	64-bit	6.4 GB/s	40 *****	10
Radeon HD 2400 XT	700 MHz	1.6 GHz	64-bit	12.8 GB/s	40 *****	10
Radeon HD 2600 Pro	600 MHz	800 MHz	128-bit	12.8 GB/s	120 *****	10
Radeon HD 2600 XT	800 MHz	1.6 GHz (GDDR3) or 2.2 GHz (GDDR4)	128-bit	25.6 GB/s (GDDR3) or 35.2 GB/s (GDDR4)	120 *****	10
Radeon HD 2900 GT	600 MHz	1.6 GHz	256-bit	51.2 GB/s	240 *****	10
Radeon HD 2900 Pro	600 MHz	1.85 GHz	512-bit	118.4 GB/s	320 *****	10
Radeon HD 2900 XT	740 MHz	1.65 GHz (GDDR3) or 2 GHz (GDDR4)	512-bit	105.6 GB/s (GDDR3) or 128 GB/s (GDDR4)	320 *****	10
Radeon HD 3450 ^	600 MHz	1 GHz	64-bit	8 GB/s	40 *****	10.1
Radeon HD 3470 ^	800 MHz	1.90 GHz	64-bit	15.2 GB/s	40 *****	10.1
Radeon HD 3650 ^	725 MHz	1 GHZ (DDR2) or 1.6 GHz (GDDR3)	128-bit	16 GB/s (DDR2) or 25.6 GB/s (GDDR3)	120 *****	10.1
Radeon HD 3690 ^	668 MHz	1,656 MHz	128-bit	26.5 GB/s	120 *****	10.1
Radeon HD 3850 ^	670 MHz	1.66 GHz	256-bit	53.12 GB/s	320 *****	10.1
Radeon HD 3870 ^	775 MHz	2.25 GHz	256-bit	72 GB/s	320 *****	10.1
Radeon HD 3870 X2 ^  +	825 MHz	1.8 GHz	256-bit	57.6 GB/s	320 *****	10.1
Radeon HD 4350 ^	600 MHz	1 GHz	64-bit	8 GB/s	80 *****	10.1
Radeon HD 4550 ^	800 MHz	1.6 GHz	64-bit	12.8 GB/s	80 *****	10.1
Radeon HD 4650 ^	600 MHz	1 GHz or 1.4 GHz	128-bit	16 GB/s or 22.4 GB/s	320 *****	10.1
Radeon HD 4670 ^	750 MHz	2 GHz (512 MB) or 1,746 MHz (1 GB)	128-bit	32 GB/s or 27.94 GB/s	320 *****	10.1
Radeon HD 4830 ^	575 MHz	1.8 GHz	256-bit	57.6 GB/s	640 *****	10.1
Radeon HD 4850 ^	625 MHz	2 GHz	256-bit	64 GB/s	800 *****	10.1
Radeon HD 4850 X2 ^ +	625 MHz	2 GHz	256-bit	64 GB/s	800 *****	10.1
Radeon HD 4870 ^	750 MHz	3.6 GHz	256-bit	115.2 GB/s	800 *****	10.1
Radeon HD 4870 X2 ^ +	750 MHz	2.6 GHz	256-bit	115.2 GB/s	800 *****	10.1

* ATI doesn’t set a default clock for Radeon X800 SE chip. The specs depend on the video card manufacturer. So you have to take care when comparing video cards using this chip.

** Depends on the model. There are boards based on Radeon X800 GT using DDR, DDR2 and GDDR3 memories running at different speeds. We’ve seen GDDR3 models running at 980 MHz and DDR models running at 700 MHz. You can calculate the memory transfer rate using the formula memory clock x number of bits / 8. A model with GDDR3 memory running at 980 MHz and 256-bit interface has a transfer rate of 31.36 GB/s.

*** There are models using DDR memories and running at lower clock rates.

**** There are three video card versions using this chip with very different specs, depending on the memory chips used. If they are 128 MB DDR, then the graphics chip runs at 400 MHz, the memory runs at 500 MHz, a 128-bit memory interface is used and the memory has a maximum theoretical transfer rate of 8 GB/s. If the card has 128 MB DDR2, then the graphics chip runs at 325 MHz, the memory runs at 666 MHz, a 64-bit memory interface is used and the memory has a maximum theoretical transfer rate of 5.3 GB/s. And finally if the card has 256 MB DDR2 then the graphics chip runs at 400 MHz, the memory runs at 666 MHz, a 128-bit memory interface us used and the memory has a maximum theoretical transfer rate of 10.6 GB/s.

***** The shader unit is unified, meaning that this chip doesn’t have separated pixel shader and vertex shader units. On video cards from Radeon HD 2400 and HD 2600 series the video card manufacturer can use a different clock for the memory (usually lower, thus achieveing a lower performance compared to the reference model); the clock rates published here are the official one.

^ Based on PCI Express 2.0, which doubles the available I/O bandwidth from 2.5 GB/s to 5 GB/s if a PCI Express 2.0 motherboard is used.

+ Radeon HD 3870 X2, Radeon HD 4850 X2 and HD 4870 X2 use two Radeon chips working in parallel (CrossFire). The specs published are for only one of the chips.

When you compare chips, you have to be very careful. Judging from the table, a Radeon 9800 may seem slower than a Radeon 9600 Pro, since its clock is inferior, and a Radeon X700 Pro seems faster than a Radeon X800 since it uses a higher clock rate.

However, Radeon 9800 accesses its memory using a 256-bit interface and processes eight pixels per clock pulse, while the Radeon 9600 Pro accesses its memory using a 128-bit interface and processes four pixels per clock pulse. This means that memory access and processing performance of the Radeon 9800 would the double of that of the Radeon 9600 Pro if they were working at the same clock. In other words, a Radeon 9600 Pro would have work at 650 MHz and access the memory at 1.360 MHz to have the same performance of the Radeon 9800.

The same idea goes for the Radeon X700 Pro example, it accesses memory using a 128-bit interface and processes data at 8 pixels per clock tick, while Radeon X800 accesses memory using a 256-bit interface and processes data at 12 pixels per clock tick.

Therefore, it is not correct to compare graphic chips only through their clocks. For the processing performance you will have to compare the clocks and the number of pixels per clock. As of the memory, the right way to compare its performance among different chips is through their memory transfer rate, which is calculated using the formula (clock x bits per clock)/ 8.

As you can see in the table, “SE” chips are the simplest and access the memory at only 64 bits per time. Another detail is that ATI uses the letters “XT” to indicate the fastest chip in a series, while its competitor, nVidia, uses the same letters to indicate the simplest chip in a series.

“PE” stands for “Platinum Edition” and are models even faster than the “XT” models, aimed to gamers with money.

As for the DirectX version, check the table below:

DirectX	Shader Model
7.0	No
8.1	1.4
9.0	2.0
9.0c	3.0
10	4.0
10.1	4.1

Nvidia Chips Comparison

Chip	Core Clock	Memory Clock	Memory Interface	Memory Transfer Rate	Pixels per clock	DirectX
GeForce 4 MX 440 AGP 8x	275 MHz	512 MHz	128-bit	8.1 GB/s	2	7
GeForce MX 4000	250 MHz	*	32-bit or 64-bit or 128-bit	*	2	7
GeForce FX 5200	250 MHz	400 MHz	64-bit or 128-bit	3.2 GB/s or 6.4 GB/s	4	9.0
GeForce FX 5200 Ultra	350 MHz	650 MHz	128-bit	10.4 GB/s	4	9.0
GeForce FX 5600	325 MHz	550 MHz	128-bit	8.8 GB/s	4	9.0
GeForce FX 5500	270 MHz	400 MHz	64-bit or 128-bit	3.2 GB/s or 6.4 GB/s	4	9.0
GeForce FX 5600 Ultra	500 MHz	800 MHz	128-bit	12.8 GB/s	4	9.0
GeForce FX 5700 LE	250 MHz	400 MHz	128-bit	6.4 GB/s	4	9.0
GeForce FX 5700	425 MHz	600 MHz	128-bit	9,6 GB/s	4	9.0
GeForce FX 5700 Ultra	475 MHz	900 MHz	128-bit	14.4 GB/s	4	9.0
GeForce FX 5800	400 MHz	900 MHz	128-bit	14.4 GB/s	8	9.0
GeForce FX 5800 Ultra	500 MHz	1 GHz	128-bit	16 GB/s	8	9.0
GeForce FX 5900 XT	390 MHz	680 MHz	256-bit	21.7 GB/s	8	9.0
GeForce FX 5900	400 MHz	850 MHz	256-bit	27.2 GB/s	8	9.0
GeForce FX 5900 Ultra	450 MHz	850 MHz	256-bit	27.2 GB/s	8	9.0
GeForce FX 5950 Ultra	475 MHz	950 MHz	256-bit	30.4 GB/s	8	9.0
GeForce PCX 5300	325 MHz	650 MHz	128-bit	10.4 GB/s	4	9.0
GeForce PCX 5750	475 MHz	900 MHz	128-bit	14.4 GB/s	4	9.0
GeForce PCX 5900	350 MHz	500 MHz	256-bit	17.6 GB/s	8	9.0
GeForce PCX 5950	475 MHz	900 MHz	256-bit	30.4 GB/s	8	9.0
GeForce 6200	300 MHz	550 MHz	128-bit	8.8 GB/s	4	9.0c
GeForce 6200 LE	350 MHz	550 MHz	64-bit	4.4 GB/s	2	9.0c
GeForce 6200 (TC)	350 MHz	666 MHz *	32-bit or 64-bit	2.66 GB/s or 5.32 GB/s *	4	9.0c
GeForce 6500 (TC)	400 MHz	666 MHz *	32-bit or 64-bit	2.66 GB/s or 5.32 GB/s *	4	9.0c
GeForce 6600	300 MHz	550 MHz *	64-bit or 128-bit	4.4 GB/s or 8.8 GB/s *	8	9.0c
GeForce 6600 DDR2	350 MHz	800 MHz *	128-bit	12.8 GB/s *	8	9.0c
GeForce 6600 LE	300 MHz	*	64-bit or 128-bit	*	4	9.0c
GeForce 6600 GT	500 MHz	1 GHz	128-bit	16 GB/s	8	9.0c
GeForce 6600 GT AGP	500 MHz	900 MHz	128-bit	14.4 GB/s	8	9.0c
GeForce 6800 LE	300 MHz	700 MHz	256-bit	22.4 GB/s	8	9.0c
GeForce 6800 XT	325 MHz	600 MHz	256 bits	19.2 GB/s	8	9.0c
GeForce 6800 XT AGP	325 MHz	700 MHz	256 bits	22.4 GB/s	8	9.0c
GeForce 6800	325 MHz	600 MHz	256-bit	19.2 GB/s	12	9.0c
GeForce 6800 AGP	325 MHz	700 MHz	256-bit	22.4 GB/s	12	9.0c
GeForce 6800 GS	425 MHz	1 GHz	256-bit	32 GB/s	12	9.0c
GeForce 6800 GS AGP	350 MHz	1 GHz	256-bit	32 GB/s	12	9.0c
GeForce 6800 GT	350 MHz	1 GHz	256-bit	32 GB/s	16	9.0c
GeForce 6800 Ultra	400 MHz	1.1 GHz	256-bit	35.2 GB/s	16	9.0c
GeForce 6800 Ultra Extreme	450 MHz	1.1 GHz	256-bit	35.2 GB/s	16	9.0c
GeForce 7100 GS (TC)	350 MHz	666 MHz *	64-bit	5.3 GB/s *	4	9.0c
GeForce 7200 GS (TC)	450 MHz	800 MHz *	64-bit	6.4 GB/s *	4	9.0c
GeForce 7300 SE (TC)	225 MHz	*	64-bit	*	4	9.0c
GeForce 7300 LE (TC)	450 MHz	648 MHz *	64-bit	5.2 GB/s *	4	9.0c
GeForce 7300 GS (TC)	550 MHz	810 MHz *	64-bit	6.5 GB/s *	4	9.0c
GeForce 7300 GT (TC)	350 MHz	667 MHz	128-bit	10.6 GB/s	8	9.0c
GeForce 7600 GS	400 MHz	800 MHz	128-bit	12.8 GB/s	12	9.0c
GeForce 7600 GT	560 MHz	1.4 GHz	128-bit	22.4 GB/s	12	9.0c
GeForce 7800 GS	375 MHz	1.2 GHz	256-bit	38.4 GB/s	16	9.0c
GeForce 7800 GT	400 MHz	1 GHz	256-bit	32 GB/s	20	9.0c
GeForce 7800 GTX	430 MHz	1.2 GHz	256-bit	38.4 GB/s	24	9.0c
GeForce 7800 GTX 512	550 MHz	1.7 GHz	256-bit	54.4 GB/s	24	9.0c
GeForce 7900 GS	450 MHz	1.32 GHz	256-bit	42.2 GB/s	20	9.0c
GeForce 7900 GT	450 MHz	1.32 GHz	256-bit	42.2 GB/s	24	9.0c
GeForce 7900 GTX	650 MHz	1.6 GHz	256-bit	51.2 GB/s	24	9.0c
GeForce 7950 GT	550 MHz	1.4 GHz	256-bit	44.8 GB/s	24	9.0c
GeForce 7950 GX2 **	500 MHz	1.2 GHz	256-bit	38.4 GB/s	24	9.0c
GeForce 8400 GS ***	450 MHz / 900 MHz	800 MHz	64-bit	6.4 GB/s	16	10
GeForce 8500 GT ***	450 MHz / 900 MHz	666 MHz or 800 MHz	128-bit	10.6 GB/s or 12.8 GB/s	16	10
GeForce 8600 GT DDR2 ***	540 MHz / 1.18 GHz	666 MHz or 800 MHz	128-bit	10.6 GB/s or 12.8 GB/s	32	10
GeForce 8600 GT GDDR3 ***	540 MHz / 1.18 GHz	1.4 GHz	128-bit	22.4 GB/s	32	10
GeForce 8600 GTS ***	675 MHz / 1.45 GHz	2 GHz	128-bit	32 GB/s	32	10
GeForce 8800 GS *** ^	550 MHz / 1,375 MHz	1.6 GHz	192-bit	38.4 GB/s	96	10
GeForce 8800 GT *** ^	600 MHz / 1.5 GHz	1.8 GHz	256-bit	57.6 GB/s	112	10
GeForce 8800 GTS ***	500 MHz / 1.2 GHz	1.6 GHz	320-bit	64 GB/s	96	10
GeForce 8800 GTS 512 *** ^	650 MHz / 1,625 MHz	1.94 GHz	256-bit	62.08 GB/s	128	10
GeForce 8800 GTX ***	575 MHz / 1.35 GHz	1.8 GHz	384-bit	86.4 GB/s	128	10
GeForce 8800 Ultra ***	612 MHz / 1.5 GHz	2.16 GHz	384-bit	103.6 GB/s	128	10
GeForce 9400 GT *** ^	550 MHz / 1.4 GHz	800 MHz	128-bit	12.8 GB/s	16	10
GeForce 9500 GT *** ^	550 MHz / 1.4 GHz	1 GHz (DDR2) or 1.6 GHz (GDDR3)	128-bit	16 GB/s (DDR2) or 25.6 GB/s (GDDR3)	32	10
GeForce 9600 GSO *** ^	550 MHz / 1.35 GHz	1.6 GHz	192-bit	38.4 GB/s	96	10
GeForce 9600 GT *** ^	650 MHz / 1,625 MHz	1.8 GHz	256-bit	57.6 GB/s	64	10
GeForce 9800 GT *** ^	600 MHz / 1.5 GHz	1.8 GHz	256-bit	57.6 GB/s	112	10
GeForce 9800 GTX *** ^	675 MHz / 1,688 MHz	2.2 GHz	256-bit	70.4 GB/s	128	10
GeForce 9800 GTX+ *** ^	738 MHz / 1,836 MHz	2.2 GHz	256-bit	70.4 GB/s	128	10
GeForce 9800 GX2 ** *** ^	600 MHz / 1.5 GHz	2 GHz	256-bit	64 GB/s	128	10
GeForce GTX 260 *** ^	576 MHz / 1,242 MHz	2 GHz	448-bit	112 GB/s	192	10
GeForce GTX 280 *** ^	602 MHz / 1,296 MHz	2.21 GHz	512-bit	141.7 GB/s	240	10

* The manufacturer can setup a different memory clock rate or interface, so pay attention because not all video cards based on this chip have this spec. The memory transfer rate will depend on the interface and clock rate used. See how to calculate below.

** GeForce 7950 GX2 and GeForce 9800 GX2 use two graphics processors in parallel (SLI mode). The specs published are for just one of the chips.

*** GeForce 8, 9 and 200 series use two clocks, the higher one is used by the shader unit and the lower one by the rest of the chip. The shader unit is unified, meaning that these chips don’t have separated pixel shader and vertex shader units.

^ Based on PCI Express 2.0, which doubles the available I/O bandwidth from 2.5 GB/s to 5 GB/s if a PCI Express 2.0 motherboard is used.

(TC) means TurboCache. TurboCache is a technology that allows the video card to simulate more video memory by using part of the main system RAM as video memory.

At first nVidia’s profusion of letters may seem confusing. The GeForce FX 5700 Ultra chip works at a higher clock than the GeForce FX 5900, GeForce FX 5900 Ultra and GeForce FX 5900 XT chips, and this may make you think that a GeForce FX 5700 Ultra is the faster than chips in the 5900 series.

But that is not really so. Chips from the GeForce FX 5900 series access the memory at 256 bits per time, while the memory is accessed at 128 bits in the FX 5700 series. That makes the 5900 series memory access performance twice as fast as those of the previous series. For instance, the GeForce FX 5700 Ultra would have to access its memory at 1,700 MHz – the double of the memory clock used – to reach the memory performance of the GeForce FX 5900 Ultra.

Another example. From the table you may think GeForce 6600 GT is faster than a GeForce 6800 because it has a higher clock rate (500 MHz against 325 MHz). But GeForce 6800 accesses memory 256 bits at a time while GeForce 6600 GT accesses memory 128 bits at a time, and also GeForce 6800 processes 12 pixels per clock tick, while GeForce 6600 GT processes eight pixels per clock.

The right way to compare the memory performance of different chips is through their memory transfer rate, which is calculated using the formula (clock x bits per clock ) / 8.

Another difference is the graphic processor of the FX 5900 series, which processes eight pixels per clock pulse, while the graphic chip only processes four pixels per clock in the other series. In other words, despite having a higher clock, the graphic processing performance of the GeForce FX 5700 Ultra is inferior than those of chips from the FX 5900 series, as they process the double of pixels when work at the same clock (simply put, the GeForce FX 5700 Ultra would have to work at twice its clock to have the same performance of the GeForce FX 5900 Ultra).

Therefore, it is not correct to compare graphic chips only through their clocks.

We must be careful with the GeForce FX 5900 XT, too. While ATI uses the letters “XT” to indicate high-end chips (for ex.: Radeon 9800 XT), nVidia uses the same letters to indicate the low-end chips of the series (see table).

You have to be very carefull with low-end video cards using nVidia chips, because they can use different clock rates and different memory interface from the table. For example, you can find GeForce FX 5200, GeForce FX 5500 and GeForce 6600 with 64-bit or 128-bit interface. We’ve seen GeForce FX 5200, GeForce FX 5500 and GeForce 6200 with 32-bit interface on the market!