"Under the hood - Why is the K7 better than the P3" The K7, or more appropriately, Athlon, is Advanced 
Micro Design's (AMD) latest venture intothe microprocessor market and yet another 
attempt at tackling that microprocessing giant that is Intel. 


AMD made 2 attempts at the lower end 
segment of the industry with their K5 and later on the 
K6 processors, both CPUs however were not capable of 
going head to head with the Intel offerings, especially 
in FPU intensive tasks, like games. 


AMD made an attempt to overcome that weakness by introducing 
SIMD instructions (Single Instruction Multiple Data) to their 
K6 line of processors and calling those CPUs K6-2. Those instructions 
were not new, MMX was SIMD, introduced by Intel in their 
Pentium line, but 3DNow!, as AMD called their SIMD instructions, 
were the first to support floating poing instructions, 
where as MMX only handled integer instructions.
A bold move it was, and as benchmarks showed, the performance 
gained by optimizing game code and 3D card drivers to support 
3DNow! was substantial. Still, the K6-2 was hindered by it's weak FPU and never really caught on as a viable alternative to 
the Pentium 2, and when Intel launched their Celeron line of 
CPUs that really hurt AMD's market share at the low end segment where the K6 and K6-2 CPUs were selling quite well.
Since then, AMD has decided to drop Intel's architectural designs 
as with all their previous CPUs, and design an architecture that 
is unique, hopefully beat Intel at it's own terf sort of speak. So, whereas the K5, K6 and K6-2 all used Socket 7 or 
Super7 for the K6-2, this time around the K7/Athlon is going 
to use a deviant of Intel's Slot 1 architecture, dubbed Slot A. 

 Early reports on the Athlon were starting to appear, an overwhelming 
amount of which were positive touting the Athlon as a viable alternative 
to Intel's offerings. That much we expected, but when AMD finally 
started shipping final versions of the Athlon, everyone was blown away. I'll explain.

Once AMD's weakpoint, the FPU, is now with the Athlon, it's 
strong point, finally beating Intel's CPUs. Before I get into 
that though, I'll give an overview of Athlon's features, compared to the P3's.
The Athlon was said to support a 200Mhz bus, something not seen 
as acheivable with current RAM yields, since current RAM could 
only be rated up to 133Mhz. This may hold true with Intel's CPUs, since their bus must be synchronised with the 
system memory. 


So a P3 that has a 100Mhz bus would always have memory running 
at 100Mhz as proposed by the GTL+ bus used by Intel's CPUs. AMD, 
however, chose to use Alpha's EV6 bus, which can have it's bus 
asynchronous to system memory. So it is possible for AMD, to have 
a 200Mhz bus from the chipset to the CPU(s) while system memory 
can run at any speed the chipset sets it to, scaling up of course, 
as memory yields increase. 


Before I wrote CPU(s) because with the Athlon, AMD is finally 
supporting SMP(Symmetric MultiProcessing), meaning multiple 
CPUs in one system. In fact the EV6 bus enables each CPU to 
have it's own, seperate pipeline to the chipset(200Mhz bus) whileIntel's CPUs have to share that 100Mhz bus. 


Obviously AMD has the advantage here. With the K6-2, AMD had 
on-die cache running at full clock speed, very similar to Intel's Celeron A' range of CPUs. The Athlon, like the P2/P3 has, for the 
moment, a 512KB L2 cache running at 1/2 the clock rate of the CPU, 
with plans to have on-die cache, and as much as 8MB(!) of L2 cache, 
running at either 1/2, 1/3 or 1/4 the clock rate. On top of that 
the Athlon has 4x the L1 cache the P3 has, meaning 128KB L1 cache, 
while the P3 has 32KB. 


On the FPU department, the P3 has 2 FPUs(Floating Point Units), one 
of which is fully pipelined while the other is partially pipelined. 
Athlon on the other hand not only has fully pipelined FPUs, but has 
3 of them! Dramatic increase indeed. AMD pulled all the stops, 
increasing the capability of their 3DNow! 


SIMD instructions with features like integer operation support 
and DSP instructions, mainly useful for decoding/encodingsound(eg MP3) streams and video among others going from 29 3DNow! 
instructions on the K6-2 to 45 3DNow! instructions on the Athlon. 
Of course Intel's P3 line has similar capabilitiesin the form of SSE(Streaming SIMD Extensions). Which better remains 
to be seen, but it appears that Intel has gained more developer 
support with SSE than AMD has with 3DNow!
A combination of the Athlon's large 128KB L1 cache and 9-issue 
Superscaler instruction pipeline(meaning that it can use multiple 
execution units to process certain operations in parallel) results 
in pipelines that are constantly fed with instructions, almost always busyand not wasting any CPU cycles doing nothing. That 9-issue pipeline 
is consisted of the 3 FPUs I mentioned above, along with 3 pipeliness 
for integer operations and 3 addressing units. The P3 has a 12 stage, 
2 issue pipeline which means that each execution unit is split into 12 
stages, each stage processing a different instruction throughout the 
pipeline. 


Again remember that one FPU is only partially pipelined. Those 12 
stages the P3's execution units split into, are a viable source 
of latency, so bad in fact that a single multiply takes 2 CPU cycles to execute(of course the CPU tries to keep the pipelines filled most 
of the time with other instructions so as to make use of those 
2 cycles that is has to spend anyway). 


The Athlon is not hindered by such latency problems, as it's 
execution units are not split into stages, whilebeing a 9-issue pipeline it can handle plenty of instructions 
per clock, still being able to execute a single instruction 
without taking the performance hit caused by latency like with the P3.
All these architectural improvement show up in real world benchmarks. 
e.g. a single Athlon at 600Mhz outperforms dual P3 550s in Quake3Test!
 So, undoubtably, the Athlon is, clock per clock, faster than the P3. 
Apart from that though, the question arises whether AMD can keep it's 
timeschedule and ship Athlons, in bulk. AMD's new fab plant is being 
used to pump out 0.18micron Athlons, while the only Athlons availiable now are 0.25micron ones, and not in retail, only by manufacters. 


The Athlon's retail release date has already slipped from the 
mid-July/early August that was promised at first. When willwe be able to go guy one at our local computer hardware store 
is yet unknown. Let's hope their fab can keep up with the demand, 
on top of that, lets hope AMD can keep up with Intel's staggering 
price cuts on the P3 line(40% off on the p3 550!!) not to mention 
win the all important clock rate war. 


I mean, how many of you would buy an Athlon 650 when a P3 700 is availiable? Most people just look at the raw clock rate, even 
if an Athlon at 650 would just barely outperform a P3 700(judging 
by a comparison I saw of an Athlon 600 and a P3 650). I'm as bold to say that most probably the Athlon would outperform even 
the awaited Coppermine CPU, a new revision of the P3 with a bigger 
L1 cache and at 0.18microns among other things. 


Of course Intel's next big thing is the Merced, due to be released 
sometime next year(see EPIC: What's It All About).
Price wars and market share wars only translate to better, faster technology 
availiable to us, sooner than expected and cheaper too! :) Good news for 
us. Long live the competition then :)