Intel’s New Weapon – The Coppermine – THG.RU

Introduction

Not long ago the impossible happened. Intel, the largest processor-maker in the world, was pushed from its throne of the provider of the fastest x86-CPU. Small and struggling AMD had provided a product that’s not only just about faster than Intel’s flagship, but leaving Intel’s Pentium III-series, including Xeon, in the dust rather badly. As if this wouldn’t be bad enough, AMD had beaten Intel where it really hurt, in Intel’s old domain, the floating-point area. While this new AMD Athlon-CPU was ahead of Pentium III in the integer-arena quite considerably, it left all Intel-processors far behind at ‘number-crunching’, the area that used to be pronounced so overly important by Intel’s marketing in the past. It was a nasty slap in the face of the self-confident Intel-management and everyone in this company dealt with it in a different way. Many OEMs and system integrators will remember Intel road shows, where Chipzilla’s marketing guys tried to make fun of Athlon or decided that the Earth was flat, water flows uphill and Athlon was an inferior product to the wonderful Intel-processors. Others – God bless them – just kept quiet as you should do when you’re defeated, but the majority faced this ‘ridiculous’ condition with the well-known ‘Intel-arrogance’ and tried to just ignore the existence of Athlon.

Stress Creates Mistakes

A lot more happened behind the scenes. While the guys in chipset sales were instructed to make motherboard makers feel a bit queasy about producing Athlon-platforms, the engineers were pushed to give 200%. The first reaction was the early release of Pentium III 600, a CPU that acted unstable in the very beginning, but that was saved by one of Intel’s many microcode-updates (which always decrease performance a bit, but who would care about that?), so that the majority of users never noticed it. Then the development of the next processor in Intel’s long roadmap was pushed. This product by the code name ‘Coppermine’ was released today and we’ll hear a lot more about it later. However, the release of Coppermine was supposed to be a lot earlier. Coppermine and the infamous ‘Camino’- or i820-chipset were meant to be introduced on September 5, 1999. Unfortunately the engineers weren’t quite able to sustain the stress that was put onto them. Used to making products that were way ahead of the competition at the time of their release, they now had to face a situation where they had to catch up with this bold competitor AMD and its Athlon. This resulted into making mistakes. Coppermine’s release was pushed to October 25, 1999 and nobody really knows if and when i820 will ever see it’s official launch-day.

No i820-Release, but i820-Benchmark Data … ?

Now we’ve finally reached the introduction of ‘Coppermine’ and you will see that Intel’s processor-engineers did rather respectable work. Everything would be fine for Intel, if there wasn’t the delay of i820. This chipset by the code name of ‘Camino’ should have provided the platform for the new ‘Coppermine’-CPU. Coppermine runs at 133 MHz front side bus clock and has a lot more great features. Intel’s history was always to supply a new advanced platform for a new advanced processor. This time Intel is in rather bad shape. i820 might be released (hopefully?) sometimes by the end of this quarter and until then you’ll either have to plug this high-end and high-cost CPU into a low-end / low-cost platform with the i810e-chipset or you’ll have to choose a platform based on the Apollo Pro133+-chipset from Intel’s Taiwanese ‘enemy’ VIA. Intel is hurting so badly that they did something they’ve never done before. In the official Coppermine presentation you’ll find benchmark-data of Coppermine-CPUs scored on a yet unreleased i820-platform! That’s rather ridiculous, because Intel was the number one company that got upset if any of their competitors used anything unreleased in their benchmarks and marketing documents. Now they did it themselves because they didn’t have another choice. How would it look if Intel provided Coppermine-performance-data based on the mediocre i810e-platform? It’s rather obvious that nobody would be impressed. The alternative would have been to market VIA’s Apollo Pro 133+-chipset in their presentations by showing Coppermine-numbers produced on platforms with this chipset. All that was unacceptable to the guys in Intel’s management, so that they decided to simply ‘bend the rules’ a little bit and promise their OEM’s that ‘eventually i820 will be available’. “Just be a little bit patient! After all we are almighty Intel and you better take us for real!”

So far about the policy-background of Intel’s new Coppermine-processor, let’s now get down to the technical side.

Coppermine’s New Enhancements

What’s the new stuff about Coppermine? Well, first of all it’s important to note that Coppermine is not a completely new design. Rather than that it’s the probably last enhancement of a CPU that was launched some 5 years ago as ‘Pentium Pro’. Coppermine is an improved Pentium III-processor and most of it is based on its predecessor. I’d also like to note that the code name ‘Coppermine’ does NOT imply that this new processor is built using IBM’s new copper-technology. The metal layers used in Coppermine are still aluminum (or ‘aluminium’ for all Europeans). The new stuff is the following:

256 kB of integrated, on-die second level (L2) cache, running at core clock

• Effect A:
  
  This fact alone makes the new Pentium III a ‘one-chip’-solution, so that the good old ‘single edge cartridge’, filled with CPU-core, L2-cache controller and L2-cache chip, is technically no longer needed anymore. You will still find Coppermine for Slot1 because Intel wants it so, but the new Coppermine-based Pentium III processors can reside just as well in a housing for Socket370. Intel calls those chips FC-PGA370, where ‘FC’ stands for ‘flip-chip. Today’s release includes only Pentium III 500E and Pentium III 550E in this kind of form factor, but in the future Slot1 will go away and the majority of Coppermines will reside in the FC-PGA packaging.

• Effect B:
  
  The L2-cache of Coppermine may be only half the size of the L2-cache of its predecessor, but it’s running at double the speed and is thus rather faster than slower than before.

Manufactured in .18µ-technology

• Effect A:
  
  The Coppermine silicon-chip is very small, although it hosts no less than 28.1 million transistors. As a matter of fact it’s with 106 mm2 even smaller than the Katmai-chip, which only hosts a third of the amount of transistors on 128 mm2. This means that it takes less silicon to produce Coppermine and thus the productions costs go down too.

• Effect B:
  
  Coppermine requires less power and can be clocked higher. It can run at a voltage between 1.1-1.7 V and it needs less than a third of the power consumed by AMD’s Athlon processor. Thus Coppermine is not only a ‘pretty cool’ CPU, it is also perfect for mobile computing, where power consumption is very important. Coppermine’s highest clock speed that Intel releases today is 733 MHz and thus higher than AMD’s highest clocked Athlon at 700 MHz.

Coppermine’s New Enhancements, Continued

‘Advanced Transfer Cache’

Behind this new name, invented by Intel’s marketing department, you find the real reason why Coppermine is performing so much better than its predecessor. The integrated 256 kB-L2 cache is not only running at clock speed now, it became a lot of nice little enhancements that really make the difference.

• Effect A:
  
  First of all, the L2-cache is now connected to the core via a 256-bit wide data path, which is 4 times as wide as Katmai’s 64-bit wide data path to its external half-speed L2-cache. Therefore Coppermine can transfer 32 byte L2-cache data every two clocks, resulting in a whopping 11.2 GB/s data bandwidth between the core and the L2-cache.

• Effect B:
  
  The L2-cache associativity was increased to 8-way, which is supposed to improve performance by 3-6%.

• Effect C:
  
  The L2-cache latency was reduced to a quarter of what it used to be in Katmai (code name of the previous Pentium III). The reduced latency is lowering the penalty of L1-misses (that’s when the core has to look up the data in the L2-cache or in main memory) and it makes the impact of snoops (that’s when the core checks if the L2-cache data has been altered) less time consuming.

‘Advanced System Buffering’

This is another marketing term for some more enhancements in Coppermine.

• Effect:
  
  6 fill buffers vs. former 4 fill buffers, 8 bus queue entries vs. former 4 and 4 write-back buffers vs. former one make sure that Coppermine can take better advantage of the 133 MHz front side bus clock. More buffers allow more outstanding FSB operations and they again reduce latency of those operations.

The SpeedStep Technology for mobile systems

Behind this new name you’ll find Intel’s ‘Geyserville’-technology.

• Effect:
  
  SpeedStep runs the mobile Coppermine at full speed and full voltage when the notebook is plugged into an external power source and reduces clock speed and voltage for a lower power consumption (but also lower performance) when the notebook operates from its battery.

Coppermine’s New Enhancements, Continued

Improved Compiler

Intel will release a new compiler that can take full advantage of Coppermine’s faster cache and the SSE-extensions by the beginning of next year. This will result in software being more optimized for Coppermine while not optimized for Athlon. You can imagine that this will make Coppermine look better against Athlon and you can’t even blame Intel or the software developers for it, because nobody keeps AMD from supplying a compiler that optimizes for Athlon as well. If they would only do that.

You can see that the core was hardly changed, but Coppermine’s cache operations were optimized close to a theoretical maximum. You will be surprised to see how much impact those optimizations have.

Here are the three contestants. Athlon on the top, then Coppermine and underneath Katmai. You can see that Coppermine’s chip is the smallest, although it has 256 kB integrated L2-cache. The other two require external L2-cache, you can spot the two chips on the Athlon and Katmai-PCB.

What Coppermines will be available?

There are a lot of different new Pentium III processors based on the Coppermine-core released today, here’s the rather confusing list:

You can see that Intel is afraid of overclockers. The Coppermines at 500 and 550 MHz share the same multiplier as the Coppermines at 667 and 733 MHz. Thus you could easily overclock the 500 to 667 and the 550 to 733 MHz by simply running them at 133 MHz front side bus clock. To make that a bit tougher Intel released the two low-end Coppermines as PGA-versions only. This won’t scare off any real overclocker though, since you can solve this problem by either getting a VIA Apollo Pro 133+-platform with Socket370, as e.g. the Tyan board we used or by using a Socket370 to Slot1 converter card.

This is what Coppermine in FC-PGA package for Socket370 looks like.

The Coppermines for notebooks are released at 500, 450 and 400 MHz, all using 100 MHz front side bus now. Intel added a couple of small new packages for the new mobile coppermine, as you can see in the picture below:

The upper of the two can be removed, since it simply plugs into a socket, the lower of the two has to be soldered onto the notebook’s PCB.

The Benchmark Results

You remember me mentioning it, Intel was not able to release the i820-chipset along with Coppermine. Thus i820-motherboards are not really the right platform for testing Coppermine. We decided to include results of Coppermine running on a preliminary i820-platform as well, but they don’t really count right now. Thus we ran Coppermine with a FSB-clock of 100 MHz on a BX-platform and Coppermine with 133 MHz FSB on a platform with the only official 133 MHz FSB chipset, VIA’s Apollo Pro 133+.

Benchmark Setup – System & Environment Settings

This is Tyan’s S1864 Trinity motherboard with VIA’s Apollo Pro 133+ chipset.

Direct Comparison of preliminary Intel i820 and VIA Apollo Pro 133+-platform

How much ‘worse’ is VIA’s Apollo Pro running with PC133 virtual channel SDRAM than Intel’s great i820 with 400 MHz direct RDRAM really? We ran Coppermine on both:

You can see that the differences aren’t big. In office applications i820 has a slight edge of up to 3% over the Apollo Pro 133+, but in 3D-games the VIA chipset scores even better. Thus there’ shouldn’t be too much to worry about using an Apollo Pro 133+-platform with Coppermine.

Direct Comparison of Katmai and Coppermine Core

Let’s see how much faster Coppermine’s new enhancements are over the ‘good old’ Katmai core. I decided to have some mercy with Intel and used the results scored on a preliminary and unavailable i820-platform for the 133 MHz FSB scores.

Coppermine scores a respectable 10% more in office applications and for some reason in Q3Test as well. I fully understand the gain in Sysmark98, but the only explanation I’ve got for the high Q3Test-scores is GeForce’s driver. I wouldn’t be too surprised if NVIDIA already used an early version of Intel’s upcoming compiler for their drivers. It could be Id’s software too though.

Overall Processor Comparison

Now it’s about time to see how Coppermine sticks up against its archenemy Athlon. It should be a tight match between the two, as we could see from the results above. In office as well as some game applications Coppermine is a lot faster than the old Pentium III and then there is Coppermine’s clock speed advantage. AMD sells its fastest Athlon at 700 MHz, the fastest Coppermine runs at 733 MHz.

Office Application Performance under Windows98SE

Under Windows 98 Athlon still rules SYSmark98, but the margin has shrunk. You may be surprised to see that Coppermine at 133 MHz FSB scores less than at 100 MHz FSB. The reason for this is that we ran Coppermine in a legally available Apollo Pro 133+-platform, that scores worse than an i820-platform would. If i820 was ready and available, Coppermine would get closer to Athlon. Let’s face the facts though, Intel screwed it up and Coppermine is still behind Athlon.

Office Application Performance under Windows NT4

Coppermine gets a lot closer under Windows NT and the i820-scores we produced were even almost identical to the Athlon-scores. However, i820 is not available, so that Coppermine is slightly slower than Athlon in office applications under Windows NT.

Floating Point Calculation Performance

Not much has changed in the pure floating point performance from Katmai to Coppermine. Athlon is still way ahead of its pursuers from Intel. Coppermine hasn’t got the slightest chance. If you need to run a lot of rendering or other floating point intensive software, Athlon is the only sensible choice.

Gaming Performance Quake3

I mentioned my surprise already before, Coppermine scores incredibly well under Q3Test. This is the one benchmark where Athlon is looking bad, but I wonder if this hasn’t got something to do with GeForce’s driver. I am sure that Athlon could look a lot better with some software optimizing as well.

Gaming Performance Descent3 – OpenGL

Gaming Performance Descent3 – DirectX

Coppermine is scoring slightly ahead of Athlon, particularly under DirectX7.

Gaming Performance Dagoth Moor Zoological Gardens

This benchmark is obviously limited by GeForce’s T&L-engine. Athlon is edging it out at 550 MHz already, while Coppermine is a bit slower and reaches the ceiling at 650 Mhz.

Conclusion

I have to congratulate Intel for finally getting the best out of the good old P6-design. Coppermine has got very close to Athlon, it even surpasses it due to SSE and other enhancements in several 3D-games, particularly in Quake3. Coppermine suffers a bit from the delay of i820, but it is still able to show its teeth already. What we shouldn’t forget however is that there’s a new platform for Athlon due soon as well, which should give this AMD CPU some speed boost too. For workstation users the answer should still be clear, Athlon is the way better choice due to its far superior floating point performance and it will get even more obvious once there are SMP-platforms available for Athlon. Office application users shouldn’t look at those CPUs in the first place really, since every office application runs just fine on K6-2, K6-3 or Celeron platforms. The story has changed a bit for gamers. Coppermine is at least as fast as Athlon in games and until the world changes significantly, game developers will help out Intel by optimizing their games for SSE rather than putting any major effort into 3DNow!-enhancements. This is partly AMD’s fault, since it would be their job to supply developers with a compiler that optimizes for Athlon. The final decision will be made when you look at the pricing though.

*E = to differentiate 0.18-micron from 0.25-micron processors at the same frequency

*B = to differentiate 133 MHz front side bus processors from 100 MHz front side bus processors at the same speed

This is the latest pricing of AMD’s Athlon for comparison.

Intel hasn’t changed much and takes a lot more money for Coppermine than AMD takes for Athlon. AMD will have to get their .18-micron-process going and integrate their L2-cache on-die as well. Then we’ll see if Intel has only caught up or if it will overtake AMD again.