Speed No More

The Sunnyvale, California-based Advanced Micro Devices (AMD) does not want to
see or hear the word megahertz (MHz) anymore. Instead, it wants to educate the
market. It’s maxim–processor clock-speed is no measure of actual
ground-level performance.

Not that the company has much choice, for it was rejoicing just sometimes
back. And the cause for the celebration was its victory over chip giant Intel in
the GHz race. AMD had not only managed to hit the 1-GHz landmark before Intel in
March 2000, it also unveiled, barely two months later, the full-speed cache ‘Thunderbird’
for prime time. Then at the end of October came its double data rate
memory-capable chipset, the 760. The highlight of AMD’s performance, last
year, however, was the Athlon processor. As a result AMD’s global share zoomed
some 7% to touch 22.5% figure in 2000.

Things, however, came back full circle in September this year. Intel hit back
at AMD in the speed game and launched the 2-GHz Pentium 4 with a i845 chipset
and DDRAM technology. AMD’s response–it launched in October the Athlon XP
(extra performance) series of processors. Technology-wise, Athlon XP isn’t the
jump from the original Athlon that the Pentium 4 was from Pentium III, but test
reports flooding the Internet suggest that it still beats off its main
competitor. According to a EuroGamer report, ‘‘Athlon XP is likely to appeal
to the average consumer for several reasons. For starters, it is much faster
than its rivals, beating every other desktop processor in the market hands
down." And that’s where the real problem lies.

Pros Improved performance Lower power consumption Fits into socket A motherboards Supports SSE instructions Cons No improvement in floating point performance (FPP) Motherboards with thermal diode support are currently scarce Offers only 266 MHz frontside bus support

For its part, AMD so far has had little problem convincing experts that
despite being 30% slower in clock-speed and $200 cheaper, the 1.53 GHz XP
version can part performs well in excess of Intel’s 2 GHZ Pentium 4. And while
most of them agree that it has beaten Intel in the laboratory, standing in the
way of AMD’s success is a very tricky marketing dilemma. How to change the
mindset of the average consumer who doesn’t understand latency, pipelines and
several other factors that determine the speed of a CPU. Most consumer cling to
the principle that just doesn’t work anymore–higher clock speed is faster
and better. If one chip runs at 1530MHz, and another runs at 1800MHz, chances
are that most of them will pick the higher clocked CPU. This has put AMD at a
disadvantage.

So how does it plan to address this issue? The answer lies in its attempt to
develop new processor performance metric–the PR ratting. No wonder then,
things like processor architecture, benchmark numbers and overclocking did not
hog the limelight at Athlon XP’s global launch on October 9 (October 23 in
India). Instead, the company talked more about why megahertz was no longer the
true performance indicator anymore.

What is PR rating?

Public relation it definitely is for AMD but let us call it ‘performance
rating’, the way microprocessor industry likes to call it. However, AMD is not
the first to tread this path.

A few years ago, Cyrix, now part of VIA, released a line of CPUs called the
6x86. Cyrix claimed that these CPUs were equivalent to Intel CPUs of a higher
clock, and labeled them as such with ‘performance ratings’–PR ratings for
short. The first time PR ratings were used, by Cyrix, they were somewhat fair
but the company allegedly started misusing it by giving incorrect ratings and
was severely criticized for this. Between. AMD also used it for a brief period
before deciding to discontinue it because Cyrix’s style had reduced it to a
mere marketing gimmick. Nevertheless, Apple has been using it successfully for
quite some time now. It has been organizing demonstrations where Macs running
CPUs at half the speed, routinely beat PCs running on Intel processors.

This time AMD has decided to call its attempt as true performance initiative
(TPI). According to Sanjeev Keskar, country manager, AMD Far East (India),
‘‘to the end-user, the ultimate benefit of processor performance is how fast
their applications run. Performance to them, simply put, is the amount of time
taken to perform a given task.’’

What this means is that the processor that performs a given task in the least
amount of time should be considered the best performer. The microprocessor
industry, while comparing the performance of processors that execute the same
instruction set, such as x86 instruction sets in PC, defines it as the work done
by the processor in each clock cycle (instructions per clock or IPC) times the
number of clock cycles (frequency). In English, performance is the amount of
work done per clock cycle multiplied by the total number of cycles per second.

PR ratings are the company’s means to communicate to the customers how much
more work the XP range of processor can do compared to other processors and even
that of Athlon range. It’s a megahertz to real performance comparison.

For this reason, AMD has run a large number of standard benchmarks, 35 to be
precise, on Athlon and Athlon XP processors and named the new products according
to the clock speed the existing Athlon would require to match the new Athlon XP.

In other words, a Thunderbird Athlon would need to run at 1.8 GHz to achieve
the benchmark results delivered by the XP 1800+, despite the fact that the XP
1800+ actually uses a 1.53 GHz clock. It is also using the same method to name
its four products of the XP range as 1500+ (1.33 GHz), 1600+ (1.40 GHz), 1700+
(1.47 GHz) and 1800+ (1.53 GHz).

But how can this happen, anyone who has been talking just the speed, may
think.

What makes up XP?

This extra performance flows from a number of internal changes, which AMD
calls the QuantiSpeed architecture. According to Keskar, this is the processors
ability to execute nine things at one go as against Intel’s six, using new
super scalar micro-archetecture, a super scalar fully pipelined floating point
unit and the ability to get in data faster by using translation lookaside
buffers and hardware data pre-fetch. The Athlon XP also introduces 3Dnow!
Professional, 52 new instructions that help to keep things faster than a Pentium
4, with full SSE compatibility (but not the Pentium 4’ SSE2). A general
optimization of the core design has also been done. Besides boosting processor
performance, this also results in a 20% reduction in power consumption compared
to the Thunderbird design and a resulting drop in operating temperature, which
was Athlon’s weakest point as compared to Pentium 4.

The new Athlon XPs also have support for internal thermal diodes, allowing
for temperature monitoring and shut down. This, however, requires motherboard
support and compliant motherboards are expected to be in shortage at least for
some time. The good news is that AMD has resisted the temptation to change
socket design, so that XP can any Socket A motherboard. Unlike Thunderbird,
Athlon, however, XPs will only operate with a 266 MHz frrontside bus.

As far as the internal difference between the Athlon XP and Thunderbird
1.4GHz is concerned, experts suggest that they are the same as the differences
between the Duron Morgan and Spitfire cores. The data prefetch hardware that has
been added to XP enables the processor to attempt to figure out what data it
will need next, and grab it from memory, so the data is already waiting in the
chip’s faster cache memory.

There is another reason for getting out of the ‘mad race for megahertz’
too. It reflects AMD’s desperation match its performance of year 2000 and its
desire to mop extra market share by end 2001. The company has also realized that
despite beating Intel in the megahertz game once, it cannot fight it on
clock-speed to clock-speed basis.

And as PCQuest’s executive editor Krishna Kumar says in his editorial,
‘‘The point that megahertz is not an accurate measure is secondary. Instead,
like any other marketing company AMD is trying to convey that its products are
better. Let us not be under any illusion that AMD has woken up to reality only
now and is out to set the record straight. Remember that they also played the
clock-speed game as long as it suited them.’’

Another question that Keskar and his team at AMD have so far managed to evade
answering is how they plan to bridge the gap between AMD’s less than 2%
marketshare and Intel’s 97% (according to MAIT) in India.

SHUBHENDU PARTH In New Delhi

Interview: Jayant Murty, general manager (marketing), Intel Asia

How
does Intel define processor performance?

The performance of a processor depends on combination of factors–the processor
architecture, the CPU clock-speed, the type and organization of the software
application, and the speed and efficiency of system components. Overall
performance is measured by the time it takes for an application to run. In terms
of run-time, performance is governed by the equation: Run time =
(clocks/instructions) X (number of instructions/sec) X number of instructions.

On the role of internal design and architecture on processor
performance...

Well-designed processor architecture determines how efficient and powerful the
system is. The Pentium 4 processor is the world’s highest performance desktop
microprocessor. Design innovations in the Pentium 4 processor’s NetBurst
micro-architecture, such as the Execution Trace Cache and Advanced Dynamic
Execution, and deep pipelines help amplify the frequency benefits on multimedia
applications, while maintaining Intel’s performance leadership on productivity
applications. The processor has been designed with enormous headroom for the
future. For example, when the PIII 500 co-existed with the PII 450 in 1999,
skeptics and analysts alike said that the only difference between the two
processors was 50 Mhz. Two years later, when audio and video on the Internet
became a reality, the PIII 500 was 73% faster than the PII 450 on audio encoding
benchmarks and 41% faster on video encoding benchmarks.

On IPC’s role in overall performance...

IPC is not a unique number for any processor, and must always be discussed in
terms of a specific application or benchmark. There are some applications for
which the IPC might go up in case of some processors and might go down in other
instances. However, a lower IPC need not translate to lower performance. An
application that can be executed with fewer instructions at high speed can help
significantly reduce the run-time. For instance the SSE2 or Streaming Single
Instruction, Multiple Data (SIMD) Extensions 2 instructions developed for the
Pentium 4 processor do the equivalent work of several simple integer
instructions on floating point and multimedia workloads. So even though the
number of instructions required may be less, the overall performance delivered
is much higher. The Intel Pentium 4 processor was designed to deliver
performance where end-users need it the most.

On Intel’s reaction to the PR rating issue...

Intel has always maintained that performance depends on a combination of
processor architecture, CPU clock-speed, type and organization of the software
application, and speed and efficiency of system components. We don’t react on
others’ programs or strategies. We believe it is best to consider whether a
new program is designed to clarify and educate, or possibly to confuse
consumers. We support efforts to clarify and educate the consumer. We believe
consumers use a number of criteria to determine which is the best computer for
them. They look at system performance, configuration and price. They look at the
performance, quality and reliability of the key components of the system.