EPIC - What's it all about?
Zef Zerfer
zefrer@hotmail.com
EPIC, computing, INTELExplicitly Parallel Instruction Computing 
or EPIC is the basis uppon IA-64(Intel Architecture 64(bit)) 
is built. But before I go over what EPIC really is, I must first 
tell you about IA-64.

Intel, untill now, has not built a 64bit processor, something 
that companies like Digital, Sun Microsystem and IBM have been 
doing for years, and of course, these are the same companies 
that have dominance in the server market and really high end 
computers. Intel's first 64bit processor is scheduled to be 
launched mid-2000, codenamed Merced is not just evolutionary, 
as the 64bit architecture suggests, it is also revolutionary 
and here is where EPIC steps in.


EPIC is mainly based on Explicit Parallelism, as you might 
have guessed, which enables it to execute many instructions 
at once(in a bundle) while also executing many bundles per 
clock. To do this of course the processor must have lots of 
execution units and a very large data path which I'll discuss 
later on. 


To aid in parallelism the EPIC architecture has two features, 
predication and speculation, that are used to minimize stalls. 
A big change is that features like predication 
and speculation are now not in the proccessors control but 
rather in the compiler's. The compiler decides how and when 
to use features like these so that the processor doesn't have 
to. How these two features actually work is what I'll be 
discussing now.


Speculation:        

A large culprit of execution stalls are memory latencies. 
No matter how fast your memory is, CPUs will always be faster. 
What speculation actually does is that it minimises the effect 
memory latencies will have on execution time thus enabling the 
processor to sustain it's parallel operation at all times(or 
most of the time :)).


There are two types of speculation, namely control and data. 
In a control speculation a memory load occurs even before an 
operation is executed. Now if an invalid value(like an 
undefined value) exists? In a normal program this would can a 
general protection fault and a crash but in EPIC systems it 
remembers that that data had been speculated and simply check 
for exceptions(invalid values) when it actually needs the data. 


If the data has already been loaded before the operation has 
been determined(executed) then there would be no stall due to 
memory latency...

Data speculation on the other hand is when a memory load 
occurs before a memory store that is directly related to 
the load. 

For example you have a program that stores data in memory and 
then loads some other data that is directly related to what  it 
just stored. Relationships between memory addresses are 
determined at compile time by a proccess called disambiguation.


Again, before the program actually uses the data it loaded via 
data speculation it must check if it is indeed the data it's 
supposed to be(ie, no memory addresses overlapped). If addresses 
did overlap different data might be present at the address theprogram 
loaded before the store than they would be if the memory 
load had occured after the store.


Predication:   
Hope you're still with me :). Next big feature of EPIC, is 
predication. It's main use is to eliminate branches. How? Say you
have:
 	if(a<b) 	jmp2	else jmp3	

2:	callfunction1; callfunction2;	jmp4	
3:	callfunction3; callfunction 4; 	
4:      If the above had been predicated, callfunction1,2 
	and callfunction3,4 


would had already been executed before the comparisonhad 
been determined. Predication thus minimises branches(calls 
to functions, above jmp 2,3,4). Once the comparison has 
beendetermined predication then selects the corresponding 
function's result to display.

Since the processor can execute many instructions at a time no 
performance hit occurs by executing one more function. The compiler 
determines which instructions should be predicated, 
and sets the predication to true(out of 64 registers availiable).
For all this speculation and predication to actually have a positive 
effect, parallelism is a given. To accomplish this, EPIC processors 
must have a lot of registers, Merced for one has(or will have) 128 
all-purprose registers, 128 floating point registers and the 64 
predication registers I mentioned above. 


As you can see explicit parallelism is great, but to be able to sustain 
it most of the time, speculation and predication are used, which in 
turn require parallelism to have a positive effect.

The news are good all around, and pretty soon we'll be seeing Merced 
processors make their debute. They won't make a big impact at once 
though. The EPIC architecture relies on compilers a lot, which means 
good compilers are a must if the processors are to be of any use. 


Plus there's the transition period from IA-32 code(current software) 
and EPIC/IA-64 code. All EPIC processors will be backwords compatible 
to IA-32 code but whether they'll run it as fast as current IA-32 
processors remains to be seen. Still, after years of "core evolving" 
by Intel, in the way of just adding features like MMX or SSE, or 
coreclock increases on existing cores, like what they've been doing since 
the original Pentium, I can't wait to see their first 64bitprocessor, 
not to mention the first revolutionary step from them since 
the Pentium.