PCI Express will revolutionize workstations;
CAD is the big winner
Workstation users, especially high-end workstation users,
have an insatiable appetite for high performance in their workstations,
and historically they have never gotten all the power they have needed.
The situation is about to improve as computer architecture is due to
undergo a major overhaul.
Performance is measured in terms of processing horsepower: BOPS—billions
of operations per second (in the CPU and the GPU), TPS—triangles
per second, and MPPS—millions of pixels per second. From a user’s
perspective the size and number of screens can also be a measure of
performance.
The performance or power available to a user from his or her workstation
is gated by components within the system, i.e., amount of memory and
its speed, the CPU and its speed, the graphic or video processor unit
(GPU/VPU) and its speed, and the North Bridge, the traffic manager between
these components. The three primary performance components are interconnected
via buses and port interfaces to the North Bridge. The efficiency or
speed of these interfaces gates the actual performance of the primary
engines (CPU and GPU/VPU) and creates bottlenecks within the system.
The primary engines are capable of moving more data than they do typically
but are held back by the data transfer speed of the interface communications,
i.e., the buses and ports.
Basic PC Architecture
All that changes this year. The computer industry is going to experience
a step function in the level of performance of the inter-processor interfaces
and enter into a new era of workstation architecture—one that will
no longer be backward-compatible.
The North Bridge connects to the CPU via the CPU’s front side bus (FSB),
to the memory via the memory bus, to a peripheral controller known as
the South Bridge via a proprietary bus, and to the GPU/VPU via the Advanced
Graphics Port (AGP). The key to a computer’s performance is just as
much influenced by the links between these components—the front
side bus, the connection between North Bridge and South Bridge, the
connection to the GPU/VPU—as by the components themselves. Now
all these connections can be PCI Express and the advantage in performance
will be significant.
The numbers game
PC graphics data transfer speeds.
The speed of the FSB is now at 800 MHz and 32 bits wide, making the
data transfer speed 3.2 GBytes/second. The 32-bit CPU now runs at 3.5
GHz and new processors increase in speed about 20% every 6 months. Today’s
32-bit CPU is capable of generating 14 GBytes of data per second, but
is limited by the speed of the FSB and the memory’s speed. The memory
bus runs at a 400 MHz clock today in a 2X clock mode known as double
data rate (DDR) and is 64 bits wide, giving an effective data rate of
6.4 GBytes/second. Memory speed increases about 30% every 18 months.
Today’s GPU and VPUs can operate at amazing speeds and have a local
memory (frame buffer) data transfer rate of 8 GBytes/second. Yet all
that power is gated by the current technology linking it to the North
Bridge, the AGP, which limits the data transfer rate to 2 GBytes/second
between the GPU/VPU and the North Bridge.
Today’s GPU/VPUs can process over 150 million triangles per second
and can fill polygons at over 5 billion anti-aliased pixels a second.
In a typically modern complex model, of, say, an automobile, there will
be 4 million or more triangles. A triangle used in a CAD program typically
consists of 36 bytes, thereby creating a need for a data transfer rate
(from the North Bridge) of 4.32 GBytes/second, for a 30 frame/second
(fps) display, which means the AGP8X port is keeping the GPU/VPU from
processing all the work it is capable of and it’s keeping the user from
seeing his or her CAD models in a timely way. AGP8X limits the frame
rate to 14 fps, or less, which is very jerky and constrains productivity.
Since the introduction of the AGP in 1997, the bandwidth of the AGP
bus has steadily improved and the current version, 8X (version 3.0),
is capable of running at 2.1 GBytes/second, while maintaining backward-compatibility
with AGP4X systems and graphics cards.
PC buses for graphics have changed about every 10 years as Figure 2
indicates, and PCI Express is the next step, offering transfer rates
of 4.0 GBytes per second over 16 PCI Express lanes for graphics controllers.
PCI Express has the scalability and the bandwidth to accommodate faster
processors. In addition to being able to support processor speeds in
excess of 10 GHz, PCI Express also supports faster memory and new serial
connections such as Serial ATA, SATA. It also accommodates 1-Gbit and
10-Gbit LAN, 1394b, and USB 2.0, to name several other advances in computer
architecture that together will signal a significant change in the design
of computers.
It’s a new system
The architecture looks deceptively similar to the original PC, as shown
in Figure 3 (above). However, the physical and electrical characteristics
are completely different. What that means is users will not be able
to move the graphics add-in board (AIB) from their previous workstation
to the new one if it has PCI Express in it. Some workstations will be
built using what’s known as a bridge. The bridge will convert the PCI
Express signals into AGP8X signals and such systems will actually have
two AIB slots, but only one will be usable at a time.
PCI Express will be scalable also, just as AGP was, and in a year or
two you can expect to see the bandwidth specifications pretty much follow
Moore’s Law.
All of the leading manufacturers will offer PCI Express AIBs by the
end of 2005. The traditional high-end companies like 3Dlabs, ATI, and
Nvidia will be first and systems will appear in the first half of 2004.
Others will follow as the population of PCI Express systems enters the
mid-range user base.
No pain, all gain
Other than not being able to move one’s AIB from an older system to
a newer one, there really is no penalty to the user in making the transition
to a PCI Express-based system, and there’s a lot of gain. There will
be a doubling in the graphics performance, and subsequently greater
productivity. These gains will first appear in high-end, high-performance
systems like workstations and in a few years be commonplace on all machines.
The last major change in the PC’s architecture was in 1997 when graphics
got a private connection, the AGP. That byte-wide architecture with
its common clock and inherent limitations is now going to be replaced
by 16 independent high-speed, scalable communications lines and with
it we will enter a new phase in the organization and operation of the
computer.
Platform transitions are never easy, but there are several reasons
why the transition to PCI Express could happen faster and more smoothly
than the transition from AGP (or the coming transition to 64-bit technologies
in Windows workstations). First and foremost, the transition to PCI
Express represents a major advance in terms of simpler designs with
more efficiency. Customers in fields requiring high throughput including
visualization, CAD, data mining, etc., have already reached the limits
of their hardware and they are ready for more. In addition, hardware
providers Intel, Nvidia, ATI, VIA, and others are eager to reap the
benefits of a broad transition to new architecture. The year 2004 represents
the start of a major shift to new infrastructures that is just being
kicked off by PCI Express. Look for radical new form factor designs,
brand new opportunities for graphics—even external graphic—and
system efficiencies.
It’s All About the Pixel
Expressing PCI Express and other oddities
For a standard that won’t be officially released until
Intel says so, which is gated by their new chipset Grantsdale, there’s
an awful lot of talk about PCI Express—who has it, who doesn’t,
and what comes with it. We’ve discussed the silliness of harping about
bridges—who has them, who doesn’t—elsewhere in this issue,
so we don’t need to retread that ground.
The questions we hear from interested parties (competitors, investors,
journalists, and users) are: who needs it, why is it a good or important
thing, has Intel finalized which silicon they will use yet, are there
still PCI Express issues that people are wrestling with? Also, we’re
asked about the stories coming out of Taiwan that the migration toward
the Grantsdale chipset has significantly slowed. Any reason to think
any of this is true? The answers, which would make for a great white
paper, are:
Power users need it—no one else.
It is a good thing for future scaling, and for marketing—the PC
suppliers need something new to rave about.
A few people are still having some yield issues and characterization
problems, but they will all be resolved by the time Intel announces
the brave new world.
Taiwan is finding answer number one to be true and having some difficulty
getting OEMs to sign up for it (the exception being Dell, of course).
I find all that rather odd. It’s odd in the sense that it’s not like
the move from ISA/EISA/VLbus to PCI, or from PCI to AGP, when new buses
and subsequent architectural changes were met with general enthusiasm
(except by the purveyors of the displaced and obsolesced technologies).
But, in those heady days the PC market was still expanding at double
digits, the level of cynicism wasn’t nearly as high, and in general
we all felt like we were participating in something new and exciting.
Today there is still a depressing malingering skepticism about the
PC industry’s vitality (with the notable exception of the cheerleaders,
our fathers who art in Redmond and Santa Clara, hallowed be their name
(Wintel), and one or two or a few of their most loyal lockstep followers
(Dell, HP, BenQ, etc.).
I personally don’t think that skepticism is called for, that it is
late and misplaced. Where was this skepticism when it was most needed
in the late nineties and 2000? The PC industry is still running and
growing at a rate greater than the GNP of the world or any country,
so it’s hard to criticize the industry as a whole—not many other
(legal) industries are showing that kind of growth today.
However, I have detected a weariness on the part of several suppliers
in the PC industry who feel there is a lack of opportunity for innovation
and real competition due to the controlling hand of the Wintel juggernaut.
Most if not all of the people I’ve spoken to seem to feel they have
to look for crumbs and niches to work in, and they talk about staying
out of the headlights of Intel and/or Microsoft. There are some who
have expressed the opinion, misguided of course, that perhaps PCI Express
wasn’t the godsend it’s being purported to be, and that (blasphemy)
perhaps other approaches might be better—and if not, at least it
would have been nice to have had some participation in the establishment
of the goals and specifications. “Does Intel think they are the
only ones who have engineers and can write standards?” more than
one senior person in the PC industry has said to me. The answer is,
Yes, silly; why do you ask now?
The same sort of comments come from the software developers, who look
with glee at the E.U. Microsoft trails and snicker about some of Microsoft’s
defense premises. (Would XP really stop being able to deliver multimedia
without WM9? Gosh, who ever thought it was so vulnerable?) And yet,
as laughable and blatantly absurd as Microsoft’s defense about bundling
is, no one thinks any court has the stamina, smarts, or political will
to really go against the software giant—so they just shut up and
stay out of their headlights. But the Wintel cartel has so many headlights
and so many initiatives and is so bright it’s hard to find a dark space
to work in these days, and that, more than anything is the real depressant
in the industry.
Truthfully, there is little consensus about what is right for the industry.
There are decidedly mixed emotions regarding leadership in the industry,
fear of domination by the few, and the alternative of having too many
voices suggesting the way to go and grow. No other company with the
possible exception of IBM could provide the leadership and back it up
with massive marketing to sell initiatives except the Wintel consortium,
we are told, and maybe that is so. However, other industries seem to
be able to support a half dozen or so substantial companies and successfully
market their own initiatives and ideas. I’m thinking of the pharmaceutical,
banking, automotive, and insurance industries, to name but a few.
No one inside or out of the PC industry would argue that Intel and
Microsoft do not own it, lead it, and control it. No one can name companies
in other industries with comparable power and influence. One of the
differences is those other industries, especially the ones just named,
have government oversight agencies associated with them, often more
than one. And of course those industries complain bitterly about how
the government holds back innovation and stifles growth. Yet if you
look at ROI and investor payback, those other regulated industries are
giving a damn good return and leading the PC industry in several points.
The PC industry, one investor told me, is hooked on hypergrowth and
doesn’t know how to make a profit on normal growth—the PEs are
too high and they can’t let them slip so they hype and look for other
markets.
Hooked on growth. Odd, isn’t it? Here we are at Cebit and the most
common comment is, “Not as many people this year.” When the
final count comes in I’m sure it will report there are actually more
than last year—but it doesn’t feel as frantic. Even with all the
initiatives, Wi-Fi, Wi-Max, PCI Express, HD DVD, media centers and servers,
H.264, PixelShader 3.0, and wireless everything, we, the hooked, are
feeling let down somehow. We need a better drug, something more powerful,
and PCI Express just isn’t enough. But can anything be enough anymore?.