I attended a lecture on March 2, 1999 by Fred Weber, AMD's K7 
Product Line Director. He spoke as part of MIT's 
Microsystems Technology Laboratories' 1999 VLSI Seminar Series, 
a series of weekly lectures that includes speakers from many of 
the top companies and research labs in electrical engineering in 
the world. 

-------------snip------------------------------------------------

Production, Technology, Scaling 

The first silicon K7 was produced approximately 9 months ago. It 
has already been thoroughly tested in both simulations and 
real silicon, and there don't appear to be any showstopping 
bugs. There are very small volumes sampling to various groups, 
but you probably won't be able to pick one up for some time. 
The first chipset's design is done, but I am not certain how 
far along it is in the production stage. 

The K7 die is about 180 mm2 in size. This is about normal for a 
new processor generation on introduction. After one process 
shrink, sizes typically go down to about 100 mm2, and the K7 in 
0.18 micron technology should be no exception. High-end 
processors for workstations (Alpha, HP, Sparc, etc.) are often in 
the range of 300-400 mm2, which makes their production costs too 
high for the PC market in most cases. 

Both the K6-III and K7 are intended to go through a process shrink 
to 0.18 micron later this year. Unfortunately for AMD, Intel's 
0.18 micron technology is superior to AMD's. The gates in it 
are smaller, and it is several months ahead. The K6-III is 
expected to scale up to about 700 MHz after this die shrink but 
no further; the K6 core will have reached its limits. It sounds a 
bit early, but remember that this same core started at 166 MHz! 

There are no plans for a K6-4 of any sort. 

The K7 will have taken a total of about 18 months to go from 
concept to final shipping product. As microprocessors go, 
that's pretty good. In fact, a large portion of the lecture 
was devoted to discussion of how the K7 was designed, but 
discussing that would put me squarely off topic, and you'd have 
to know a fair amount about circuits for it to make any sense at 
all. 

June 1999 still looks like a valid target date. AMD could just 
pull this one off!

------------------- snip -----------------------------------------

Performance 

I don't have a real K7 chip to test. A shame, but I'll survive. 

However, the K7 has been tested extensively in-house by AMD. 
Long before a single chip had been fabricated, engineers 
were simulating it on benchmarks and code samples of all 
kinds, ranging from Winstones to FPUmarks to SPECs to 
Quake. 

The results? I don't have any real numbers, but the K7 
will definitely be faster than the Pentium III at 
essentially everything at the same clock speed. The only way that 
the P3 could win out would be if software was SSE optimized but 
not 3DNow! optimized, or probably in some degenerate cases. 

3DNow! has been enhanced slightly since the K6-2 and K6-III. For 
a matrix-times-vector transform loop, essential for all 3D 
graphics, the K6-2 was coming in around 20 clock cycles per vector 
to transform, and the K7 is at about 14. The P3 (with 
SSE optimizations) can get to about 18 cycles. The total 
maximum throughput will remain 4 operations per cycle, 
though, placing the K7's peak FP calculation rate at 2 
gigaflops. 

All that remains is for AMD to get cracking on ISVs to optimize 
their products. Hopefully more applications will start to use 
the DX6.1 T&L engine, which will give them automatic optimization 
for SSE and 3DNow!, but it's a fact that Intel has more money 
than AMD, and more money means more optimized programs. There d
oes not seem to be any significant technical reason to adopt SSE 
over 3DNow!, and 3DNow! has a larger installed base. 

----------------- snip -----------------------------------------------

Conclusion 

The K7 is looking like a big-time winner. I would predict at 
this point that as of August 1999 (leaving a bit of room for 
delays), AMD will have the fastest x86 microprocessor you can 
buy. The shrink to 0.18 microns will be interesting, since it 
may give Intel a comparative advantage, but the superiority of the 
K7 core should be enough to keep AMD ahead in performance, if not 
in clock speed.