Slashdot Mirror
The Gigahertz Race is Back On
An anonymous reader writes "When CPU manufacturers ran up against the power wall in their designs, they announced that 'the Gigahertz race is over; future products will run at slower clock speeds and gain performance through the use of multiple cores and other techniques that won't improve single-threaded application performance.' Well, it seems that the gigahertz race is back on — a CNET story talks about how AMD has boosted the speed of their new Opterons to 3GHz. Of course, the new chips also consume better than 20% more power than their last batch. 'The 2222 SE, for dual-processor systems, costs $873 in quantities of 1,000, according to the Web site, and the 8222 SE, for systems with four or eight processors costs $2,149 for quantities of 1,000. For comparison, the 2.8GHz 2220 SE and 8220 SE cost $698 and $1,514 in that quantity. AMD spokesman Phil Hughes confirmed that the company has begun shipping the new chips. The company will officially launch the products Monday, he said.'"
This is a desperation move. AMD is back on their heels and their recovery plan is too far off in the future. In hopes of saving face they are pulling the only lever they have, clock speed.
Not so. AMD never said that they wouldn't increase clock speed on their CPUs. In fact, that's pretty much standard practice to get higher performance. So now their manufacturing process is capable of producing 3 GHz CPUs in sufficient volumes to sell, and they're selling them. As the process is refined there may be faster CPUs.
Intel does the same thing. As the manufacturing process is refined they are able to produce more and more CPUs at higher clock speeds. It's not a sign of anything other than business as usual.
Funny, Intel was chumped by AMD just like this a couple of years ago, why did AMD let themselves get tagged back? Intel woke up in a major way. Can AMD? Doesn't look too good...
AMD has more than just clock speed coming, Barcelona (aka K10) is supposed to be shipping in the next month or two. That's generally expected to take back the performance crown from Intel, and even if it doesn't it should at least eliminate the performance gap. For purposes of historical reference, AMD pretty much bitchslapped Intel when they released the Athlon 64. It took Intel 4 years to finally catch up to AMD and pass them with the Core 2 architecture, and even today the Opterons are still higher performers on 4 and 8 processor systems. If Barcelona turns out to be as fast as or faster than Core 2 (and by all rights, it should be) then it will have taken them only 1 year to catch up. Conroe was "previewed" at Spring IDF in 2006, but didn't ship until several months later.
As for why it's taken AMD a year to catch up, it takes quite a long time to design, layout, test, and debug a new CPU. Once all that is done the manufacturing process has to be designed and tested too. Then the CPUs have to actually be produced, and once production has started it takes almost 2 months to go from silicon wafers to functioning CPUs. However, something to keep in mind is that Intel is a much, much larger company than AMD and that Intel runs severals CPU design teams concurrently, while AMD doesn't. Intel has several times the number of designers, engineers, and fabs that AMD does. Because of their resources, Intel is able to completely scrap a CPU project and switch to something else if they need to. AMD can't, or at least not without seriously hurting the company. The fact that AMD is even competitive with Intel says quite a lot about the talent they have in-house.
The thing that I find most interesting was that last year when Intel was on the ropes, they offered the IDF preview to select web sites in order to generate buzz and FUD regarding Intel vs. AMD. And it worked too, because for 3 months everybody was talking about how Intel was king again even though they still hadn't shipped any Conroe CPUs. This year they're doing the same thing with their new Penryn architecture, and they don't appear to be on the ropes. Why would you tip your hand early if you don't have to? That indicates to me that Intel is concerned about something, and I suspect that something is Barcelona.
Even more interesting is that none of the previews compare Conroe with Penryn at the same clock speed. Most of the benchmarks that I have seen show a roughly 20% performance advantage for Penryn. But the Penryn CPU was running at about 14% higher clock speed, a 25% higher FSB, and with 50% more L2 cache onboard. Now who's playing the Gigahertz Game? I suspect that if you overclocked a Conroe and it's FSB to reach the same speeds, you probably would see little to no difference with Penryn. Which means that Intel's response to the all-new Barcelona is going to be...you guessed it...run up the clock speed and slap on some cache, because we're in for a bumpy ride.
You know what's funny about Beryl/Vista ? They're doing the same stuff the Mac was doing years ago on puny hardware. I mean really, how frickin' hard is it to draw a window as a texture on a pair of triangles ? Seriously!
Programmers are sloppy, because sloppy is all the industry wants to pay for. Way back in the day when CPU cycles were super expensive, programmers were paid better money and given the time to tweak the crap out of everything, because if they didn't, the app would run dog slow and people wouldn't buy it. The problem is that somehow, people now tolerate underperforming software. They see it as a reason to upgrade... good god, they actually fall for it! Gee I certainly remember surfing the web on a 486 with 8mb of Ram back in the day. Now my OS needs a good 50-60mb to itself, and that's after I ripped out all the cruft. Normally it would be 100mb just for sitting idle with a background image and a neon-colored task bar. Gee uh, where'd all my system resources go ? Does it really require 7.3 million bytes to house a TCP/IP stack when some embedded devices pull it off with oh, 6kb or so ?
The truth however, is that if we were to write code as tightly and meticulously as we did in the 80's and 90's, software would perform, on average, at least 5 to 10 times faster than today, excluding hard bottlenecks like disk access and network bandwidth. It would also take 50 times longer to write the software, and I'd say less than 1% of people who call themselves "programmers" are even able to write such finely tuned code. Everyone doing VB ? Out. Everyone doing RAD ? Out. All you Ruby on Rails weenies ? follow me to this dark alley *BLAM*
I remember spending hours on little loops, with a CPU reference manual and a calculator. Sometimes I did little time sketches to figure out the best way to stagger memory accesses so as to not starve the execution pipes. Often times that meant weaving two disparate functions together, one being memory-hungry, the other CPU hungry. Together they filled each other's latency pockets, and my routine ran nearly thrice faster as a result. No C compiler I've ever seen could do such kinky things. Heck one time I even wrote a little assembler demo whose code executed twice: forward, then backward. The opcodes and data were carefully selected to represent valid instructions when reversed. It was more than a nerdy trick, it allowed my routine to fit entirely in the CPU's on-die cache, which gave it a huge speed boost but more importantly, it enabled a lowly 486 to mix 48 sound channels in real-time. Today's Cubase can't even handle a couple dozen channels without stuttering and/or crashing, on computers over 100 times faster than a 486.
-Billco, Fnarg.com