Interview with Sun's Les Poltrack
Editor's note: The following is a lightly edited transcription of a telephone interview with Les Poltrack, Group manager desktop marketing at Sun Microsystems. The interviewer, Mark Cappel, had a 60-minute time limit.
We were thinking about creating a machine that would be the basis of our desktop product line in the latter half of the 90s. The overall architecture, and the things people could do, we tried to provide a platform that will let people do things that make sense in the latter half of the 1990s. Things like much more visual computing, the ability to handle the new data types, multimedia datatypes, visual data, graphical image data, the ability to make a more interactive network environment with the Fast Ethernet built and so forth.
We are trying to balance the new capabilities with a system that would work well with our customers' existing environments. One of the challenges when you have an installed base of a million seats is not breaking compatibility. That's one of the challenges -- balancing innovation. We're trying to do both of those things.
We provide high-performance desktops with better networking and visual computing.
The initial UltraSPARCs will be at the high-end of our desktop product lines. We anticipate continuing to sell models of our other desktop product. We just moved the (SPARCstation) 4 and 5 to 110-MHz standard. Those will continue to be available at lower prices and faster speeds.
We will reposition, at right around the same time as the UltraSPARC launch, the SPARCstation 20. It will be a little more aggressively priced. We'll offer 150-MHz version of the HyperSPARC. We'll be repositioning the products, but we won't get rid of any product lines.
We'll find these used as servers, but the majority will be used for single users. We think 10 to 15 percent will be used as servers, which is typical for our desktop machines. That will continue to be the case. They will be cost-effective. Some advanced sites are looking at (the UltraSPARC's) using groups of them for compute servers. That fits in with our business model.
I invented the desktop server at Sun with the Sun-/60 server in 1987. Since then, our desktop server business has become a multi-hundred million dollar business. The SPARCserver 1 was a fairly decent server. We learned some technical lessons from the Sun-3/60 about ways we could architect the desktop a little better to be a better server. The follow-on products -- the (SPARCstation) 2, the 10, and the 20 --were better desktop servers because of the way they were designed from an internal point of view. Faster caches, better ability to handle memory, and lots of processes and so forth. The majority will be used as desktops.
The analogy would be in the network world we've had 10 megabit Ethernet, which has been a good standard for more than a decade. In the network world people have gone from Ethernet hubs to Ethernet switches in large networks to carry one multiple conversations simultaneously. We've done something very similar inside the box. We can have multiple transactions taking place at a given time.
It also turns out with the switch, a central interconnect, you can make the physical interconnect lines shorter. You have less capacitance and inductance on a shorter line. So you can run the bus that goes into the interconnect at a faster speed. We can run it at a faster clock, we can do multiple things at once. We use a packet-switched protocol vs a circuit-switched protocol. That's sort of orthogonal to a cross-bar switch type of design. But nonetheless it gives you more effective bandwidth utilization.
In the next five years, you will see most high-performance desktops go to switch-based design. It's the logical way to market that high performance system with multiple processors, fast memory, and so forth.
The clock motherboard runs at either a half or a third of the CPU clock. So if you have a 167-MHz UltraSPARC it will run at 83-MHz. It runs at a fraction of the CPU clock. That's the way to get the most speed out of (the computer).
The higher-end, the Ultra-2 line, does have modular processors. The Ultra-1's CPU is on the motherboard primarily to make the Ultra-i the lowest-cost UltraSPARC computer we could build. Our goals is to push UltraSPARC down throughout the product in the next 18 months to two years. As we build more of these things, the first 100,000 or so, or more than that in the next year, and we get our cost down, we will be pushing these down further in the product line. We want a design that we could be at a similar point to the SPARCstation 5.
The UltraSPARC uses mounting technology called ball-grid array (BGA). It is the state-of-the-art way of putting processors on boards or modules. Initially there was a lot of concern about this being a manufacturing risk. We've had excellent experience and haven't experienced any failures with the BGA or processors related to the BGA mounting.
It is a design where you have to be smarter. There are things going on at any given time. As you design it, it is slightly more complex. These machines came up and ran the first time. It is not inordinately more complex, it is a more sophisticated design than a simple bus. In terms of cost, it is not greatly more than a bus design. With a bus, you have a set of wires on the motherboard. We implement the switch with a set of 18 small chips that are more interconnects than processors. But the cost is not materially different in terms of a high performance desktop computer. It's in the tens of dollars
It is a more sophisticated design, so there is more work to do. That hasn't led to failures or time-to-market issues. I think you will see more and more (competitors) building switch-based computers. (This) will be standard five or ten years from now.
The system design is all-Sun. We work with other vendors on our ASICs, but the system design is all Sun.
There are processor modules that do plug into the Ultra-2. They are a different form-factor than the (MBus), called the UPA, or UltraSPARC Port Architecture. One or two processors plug into slots. The boards mount orthogonally to the motherboard. A processor module with an edge-connector plugs into an edge connector on the motherboard.
Be had to because of the low-performance chips they were using. The basic architecture is MP-capable. The decision to make some of the systems uniprocessor, we look at cost/performance trade-offs. Today, about 15 to 20 percent of our SPARCstation 20s (are sold) as MP systems, or become MP systems. A considerable number, and that number's growing.
We anticipate 20 percent, and growing, of these machines will be Ultra-2s. For many applications a uniprocessor UltraSPARC is a nice match for what people are doing. There's nothing in the basic architecture that prohibits things from being MP. In the next couple years, 100 percent of users won't use MP. Maybe 25 to 35 percent will use MP.
Object environments take advantage of high performance. They take advantage of doing fast context switching. The UltraSPARC processor has an architecture designed to context-switch quickly, to get data in and out of registers quickly, it has shadow registers, things like that. It's certainly architected for environments like Neo in mind.
Neo isn't a necessary part of using these systems. But they will do things like Neo well. Neo is also networked, so building Fast Ethernet into the Ultra-1 V model and the Ultra-2 standard, and the ability to do fast context-switch in UltraSPARC are certainly part of the basic design.
64-bit applications will become important when, a) there's a large number of (64-bit) systems and b) when (64-bit) becomes attractive to ISVs as a business. We will be implementing 64-bit extensions to Solaris in future releases.
A consortium of companies, including HP, Sun, SGI, Intel, and others, is looking at a uniform way of implementing 64-bit data, pointers, and so forth. Our intent is to implement what we expect the other members of that group to implement so our software partners don't have to do a low-level rewrite of applications multiple different ways.
64-bit (computing) will become important when there are a lot of (64-bit) machines out there. The two companies that have 64-bit OSes (besides HAL) DEC and MIPS, no (customers) are running 64-bit applications. Look at DEC; its answer to everything now is NT, which is a 32-bit operating system. (Digital Unix) is not an interesting application environment for software vendors or users, given the number of Digital machines out there. Same case for (SGI's) Irix 6, which is SGI's 64-bit OS. Nobody is running it.
HAL Computer's argument is theoretical. They have to say something.
Well over 100,000 per year. We think we'll have that sort of run-rate for the Ultras as well.
We'll be shipping in volume in the calendar year. We'll ship thousands of machines this calendar year, and continue to ramp that up. We have today 200 or 300 machine outside of Sun at ISVs and beta sites in high quantity to see if there are any bugs or incompatibilities. It's the right thing to do to complete these tests to make sure your customers don't do your beta testing. We've not gotten a lot of issues with bugs. Primarily we've gotten feedback on the fact that people are pleased with the performance. In fact they are dropping existing binaries on the machines to see how much speed-up they see without changing (the software).
The head of the pack. Performance should be a problem for everyone but Sun.
DEC has its own set of problems.
We have not measured the final numbers on the Ultra-2 yet. We think that it's likely we'll have some good numbers. I think we'll be at the head of the performance pack.
Probably a combination. SPEC92s are better understood, but SPEC95s are more relevant due to the larger datasets, and so forth. The 95s are better in the long-term.
Other industry-standard benchmarks, including picture-level benchmarks. We'll do a suite.
We haven't finalized prices yet. (Editor's note: Recall this interview was conducted two weeks before launch date.)
We expect prices to start well under $20,000.
$8,995 with a 17-inch grayscale monitor. I priced it.
Initially, this will be at the high-end of our desktop product line.
There was a lot of good hard engineering work to make sure these things were in a good form-factor, and had good expansion capabilities.
After the SPARCstation-1 was done. Andy, Bill Joy, and others went to Lake Tahoe and tried to decide what architecture would be possible and necessary in latter half of the 1990s. They chose a switched architecture in 1991. The other ideas we see in the UltraSPARC computers have been percolating at Sun for some time. The design and roll-out has been very smooth because these guys have done this before, and the basic issues were ironed out early on, even before V9 was finalized. At the time, UltraSPARC's code-name was "spitfire."
The processor guys worked hand in hand with the OS and compiler group. So that when adding a capability, it would be taken advantage of by real applications. It was a joint effort by the chip designers, the hardware systems designers, the OS and compiler teams working on this thing for years.
The UltraSPARC design is not just a chip but what's inside the chip, how it's connected with the outside world. That was jointly done with the SPARC technology group. There's certainly architecture contribution from Bill Joy in Aspen. It was a distributed design.
Pieces like the 3D RAM technology was a joint design of Sun and Mitsubishi. The 3D RAM is based on a 16 megabit DRAM process. They certainly had DRAM and process expertise, and Sun had systems and graphics expertise. That's another basic part of the high-performance architecture. That was distributed among different companies. The UltraSPARC will be manufactured by TI. That's very much a joint collaborative process. You can't design a logical circuit without people who do the process and have an understanding how that will translate to real silicon. And how manufacturable it is.
In Texas and Japan. The exact process is that the process happens in one place and the packaging in another.
In Milpitas, CA and Linlithgow, Scotland. In fact they are being built in both locations today.
We have some white papers.
3D RAM provides the best aspects of DRAM, SRAM, and video RAM. It's based on a 16 megabit DRAM process. We take 6 megabits out of that leaving 10 megabits of memory on a single chip. And if you do the arithmetic, with four of those chips we get 32-bits-per-pixel frame buffer at 1280 by 1024. So it's a cost-effective way to make a true-color framebuffer. We take 12 of those, and make a double-buffer Z-buffer version of that. It's based on a DRAM chip and process for cost-effectiveness. It has ports like a video RAM, where the stuff going to the monitor and refreshing the monitor doesn't interfere with the processor or the system writing new data to the RAM, which what video RAM does. It also has an on-chip cache and ALU, so it has the best aspect of SRAM. What that lets it do, this is pretty cool.
It has a little processor on every chip itself. When we're Z-buffering and hidden surface removal, alpha-blending for transparency, anti-aliased lines, instead of having to read data out of memory, compare it to something in the processor, do an operation, and then write it back to memory. With 3D RAM that becomes a write-only operation. The system writes to 3D RAM, and the 3D RAM takes care of it. That gives us up to a 10X speedup, verses conventional video-RAM architecture for things like hidden-surface removal, transparency, or smooth lines. It's a different architecture, and these are the first systems to use it.
3D RAM is available on the open market from Mitsubishi. Anyone can use it. This is the first implementation. We did a joint announcement at Siggraph last year. With four (3D RAMs), we get a true-color framebuffer. 24-bits of color plus and 8-bit overlay.
It's just like the GX in the SPARCstation-1. It made 2D, wire-frame acceleration affordable. We think we'll bring 3D RAM and the creator graphics architecture to the same cost point. In the 1980s people migrated from monochrome systems to 8-bit color. In the latter half of the 1990s they will move from 8-bit color to true color. You can do better images, smooth shaded objects, you didn't have to do differing of this nonsense. Now, when you move your mouse from one window to another you have different color maps where you get a real funky looking display on the other windows. With 24-bit color that will go away. 3D RAM makes that possible from a cost point of view. The other hindrance of 24-bit true color was that you have to move three times the data around. 3D RAM is fast enough to handle that.
The system interconnect, the UPA switched architecture, is fast enough to move that much data around. The 24-bit windows in these systems are several times faster than the 8-bit windows for the SPARCstation-20 TGX.
True color is more appropriate for what people want to do.
We support our existing libraries, X11, XIL, and XGL. There will be a native port of OpenGL as well as a developer's release in the early part of the calendar year.
We put OpenGL on these systems several months ago and it came up and ran but we are tuning its performance, rewriting routines and libraries so that we offer good OpenGL performance.
Solaris 2.5 supports these platforms. 2.5 has the hooks for UltraSPARC. It's binary compatible. We'll package the CDE common desktop environment, which is very nice. We are preserving compatibility. 2.5 makes available NFS version 3, which gives gives a 3X performance increase over version 2. That benefits clients and servers.
We are not breaking compatibility.
The vast majority came up and ran. We've run into some subtle things where people, I don't want to be impolite, coded things wrong or didn't conform with standards. We worked with people to fix them so they don't hinder our customers. We have somewhere close to 100 machines at ISVs today, and have a while to iron out those issues. We've never done this large of a external ISV program.
In spring, we had an internal lab where we had a small number of ISVs come in and test their apps. Sun started sending machines to ISVs in July.
There are a number of businesses and applications where we are seeing phenomenal growth. An example is the Netra product line, which is putting application-specific servers in new environments. That's been well received and is a high-growth area. Our other software and hardware businesses have been strong.
If you have problems with this magazine, contact email@example.com
Last updated: 8 November 1995
If you have technical problems with this magazine, contact firstname.lastname@example.org