posted by Nicholas Blachford on Mon 9th Feb 2004 20:32 UTC
IconIn Part 1 I discussed how the software development world is about to be turned on it's head. Now in Part 2 I look at how the hardware world may be about to undergo even bigger changes and why it wont be a hardware manufacturer leading the way.

I'll get royally slagged off for this so I'll get this bit over with first...

PowerPC takes the lead from x86
The IBM PowerPC 970 (aka G5) is close to the x86 in performance with differences due to compiler and code quality currently holding it back. The new 90 nanometer 970FX will increase clock rates to well beyond 2.0GHz and IBMs compiler has already shown >30% gains over GCC on existing 970s. When updated CPUs ship and software using the IBM compiler gets into production I expect the PowerPC will finally again pull ahead of x86 after a long period in the doldrums. With 3GHz PowerPCs due later in the year long rumoured to be based on the more advanced POWER 5 architecture, I expect both Intel and AMD will remain behind.

I do expect that Intel could stay ahead in SPEC marks (or perhaps AMD [1]) but it'll become increasingly obvious that even the best intended benchmarks can be tamed by a sufficiently determined marketing department. Intel's newly announced Prescott core is proving to show little if any performance improvement over the existing Northwood core at the same clock speed yet SPEC scores are up markedly [2].

That said Prescott is only starting out and it has clearly been designed for high clock rates, the first CPUs shipping are not even beginning to show what this CPU is capable of. However, Intel are currently facing problems with high transistor leakage current in their 90 nm process and this increases power consumption dramatically, this could hold the Prescott core back from it's true potential. In any case we can expect top end Intel CPUs to climb to well past 100 Watts power consumption this year, Intel themselves have talked about water cooling and given the power consumption of early Tejas samples [3] that might not be such a bad idea.

On the other hand both IBM and AMD are using Silicon-On-Insulator technology which reduces leakage current, consequently the 970FX already has considerably lower power consumption than the 970. We can expect both the Opteron and Athlon64 to follow the same pattern. Intel will get it's power back to sensible levels when 65 nm manufacturing comes on line sometime around 2005-06.

What happens in 2005 is anyone's guess, but with a third player in the performance desktop CPU scene things will be very good for consumers. Many readers no doubt will have a different prediction for performance levels this year but one thing we can all be sure of is the performance levels of the different CPUs will be discussed, discussed and discussed...

Will the Itanic sink?
For many years the dominant computer system has been Windows running on an x86 CPU. Both Microsoft and Intel have both tried to change this to no effect, Windows NT initially ran on multiple platforms but this is no longer the case today. Intel wants us to all to switch to the Itanium but the initial performance was weak and even today it's promised performance gains over RISC CPUs have never materialised [4]. With Intel themselves now looking likely to produce a 64 bit x86 CPU, it looks like the Itanium and it's 13 year, multi-billion development cost could potentially turn out to be the biggest mistake in commercial history.

On the other hand Intel now have most of the team who developed the Alpha CPUs and they are working on the 3rd generation Itanium called "Tanglewood" due in 2005, perhaps Itanium has only had a difficult childhood and it may yet become blossom into a highly successful adult.

x86 will die
One of the x86 providers sell x86 chips which are not x86 chips. They do have some hardware to address the oddities of the x86 instruction set but nevertheless they cannot natively run x86 code. If you are into chips you'll know that it is Transmeta I am talking about.

The AMD Opteron / Athlon64 has more registers than the standard x86 instruction set defines but software has to be recompiled in order to use them. I'm curious to know what would happen if AMD was to include "code-morphing" software which recompiled the 32 bit x86 code to use these extra registers, would 32 bit performance be as good as it is in the hardware compatibility mode? Could performance even go up? If performance was comparable the 32 bit hardware mode could be largely removed, simplifying the CPU and making it both faster and cheaper.

What is more interesting is if such a technique worked on an AMD processor could it also work on something completely different? Well, yes, Transmeta have already proved this, but none of the Transmeta designs to date are not high performance CPUs. x86 code has ran on non-x86 performance platforms before and sometimes at very high performance [5].

Advantage Alpha
In the 1990s DEC produced Alpha processors and these ran Windows NT. However being incompatible with the x86 instruction set meant most of the existing software base would not run. In order to address this DEC produced a technology called "FX!32" which could emulate x86 CPUs at reasonable speed using a technique not unlike that Transmeta uses. In addition to this the Windows NT libraries were native Alpha code and a lot of programs spent a lot of time in them, this in combination with the Alpha's then huge speed advantage meant that x86 programs ran faster on the Alpha than they did on the x86.

Performance is no longer important for the vast majority of computing needs. As such, when CPUs can be made which can execute x86 code at high performance and / or low cost compared to native x86 designs there will be little point making them native any more. When x86 performance which is "good enough" and cost sufficiently low we could see the end of the native x86 processor.

Table of contents
  1. "Future of Computing Part 2, Page 1/3"
  2. "Future of Computing Part 2, Page 2/3"
  3. "Future of Computing Part 2, Page 3/3"
e p (0)    46 Comment(s)

Technology White Papers

See More