Sun also gets it
Just got back from a meeting with Sunacle [slides of my presentation as well as an abbreviated write-up of the talk can be found here], and I have to say, I couldn’t be happier. The near-term future is going to rock.
A bit of history: in the late 1980s and throughout the 1990s, Sun had system performance that was very impressive, despite having microprocessor performance that lagged behind the rest of the crowd. For example, throughout the 1990s the battle for top of the heap, as determined by SPECmarks, was fought by Hewlett-Packard (via the PA-RISC processor) and DEC (via the Alpha processor) … and occasionally IBM would get a POWER system in there.
Didn’t know that HP ever made anything besides printers? Never heard of Digital Equipment Corp before? No worries, the point is that the SPEC suite of benchmarks (now termed “SPEC CPU” to differentiate it from other benchmarks they’ve created since then) is designed to test CPU performance and not system performance: it intentionally tries to disengage the processor from the memory system, the IO system, the network, etc., and so only tests the performance of the CPU.
“Why on earth would that be valuable?” you ask.
Good question. I have no good answer, other than “it makes evaluation simpler and puts things on an evener playing field.”
“Yes, but,” you counter, “how can that be ‘even’ if you can’t run a CPU without the surrounding system? Wouldn’t the numbers be more realistic and thus more valuable if you intentionally include the effects of the surrounding system?”
Yeah, see this is the difference between science and marketing.
So, back to the story. As measured by SPECmarks, the SPARC family of processors (SuperSPARC, UltraSPARC, etc.) all fell way behind the Alpha, the PA-RISC, the POWER, the various MIPS processors, etc. For those of you just joining us, there used to be a lot of general-purpose CPUs to choose from. Then Intel ate everyone’s lunch. Anyway.
The interesting thing is that Sun systems (i.e. the boxes containing the CPU, the memory, the I/O subsystem, and running the OS) had excellent performance, right up there with the rest of the crowd. They could get excellent system performance without needing top-of-the-line processing performance. Think balance; system-level design is a balancing act, and these guys were masters at it. Still are, from what I can tell. Another data point for comparison: IBM’s BlueGene is a supercomputer built of processors that are several generations old — effectively embedded processors with low performance and low power requirements. Though IBM also has its ultra-high-performance POWER line, they don’t use those CPUs in BlueGene. Don’t have to. BlueGene systems are some of the world’s highest performing computers, while simultaneously being some of the most energy-efficient.
It’s all about balance.
So I met with the Sun guys the other day, with a bunch of Oracle people in the room as well. Sun has recently developed some amazing interconnect technology with the potential to transform the way systems are built — again, getting to the heart of the problem, which in today’s systems is the interconnect fabric: interconnect from CPU to DRAM and between CPUs.
Some links to their stuff:
- Proximity Communication (for chip-to-chip communications)
- Silicon Photonics (for system-level communications)
(apologies in advance: the second link requires access to IEEE’s library)
At the meeting I ranted about a handful of important open problems in computer memory systems today, giving the fundamental reasons for those problems — and interconnect is at the heart of most of them. So pretty much all of the problems pointed to limitations that would go away were either of Sun’s technologies to succeed. Woo!
Bottom line: the Sun research guys are attacking precisely the right problems. Expect really cool things out of Sunacle in the next few years.
