Sign in with
Sign up | Sign in
Your question

64bit Only processors vs Hybrid

Last response: in CPUs
Share
September 29, 2003 10:56:16 AM

I would imagine that 64bit only processors will run better than their hybrid counterparts. Also, If you had a 64bit CPU, could the OS not have a 32bit compatibility mode to run them in (and they would still be faster/better?)

I'm happy about the idea of having a hybrid system - but would the 64bit only processor not be able to be more tweaked and run more efficient? The hybrid idea is definately a good chioce for migration purposes, but I imagine that new AMD / Intel chips will hit the market as 64bit only, which will probably outperform the hybrid ones.
Anonymous
a b à CPUs
September 29, 2003 12:33:16 PM

point is not wether it is "hybrid" or not.. after all, our current 32 bit cpu's are hybird 8/16/32 bit machines. They sill run CP/M or DOS. this backwards compatibility only costs a tiny ammount of die space, hardly an issue at all.

A valid question is wether or not a new ISA (64 bit or not) would be faster than extending a 15 year old one like x86. In theory, I would say yes, most likely. In reality, hardly.

x86 cpu's benefit from such an enormous economy of scale, that they can be produced far cheaper and on leading edge process technologies, with fine grained binsplits 64 bit RISC cpu manufacturees can only dream off.

So, in theory a SPARC, Itanium, Alpha or Power based cpu's ought to run circles around x86 products, in reality, opteron and xeons are near the top of the chart in absolute performance, and lightyears ahead in performance/$ or performance/mm². I don't see this changing either.. Get used to it, x86 is here to stay.

= The views stated herein are my personal views, and not necessarily the views of my wife. =
September 30, 2003 7:24:28 AM

The tides are certainly changing. A look at the LV Itanium 2's should tell the tale. During its first generation, IA-64 products were slow (Merced), bulky (Merced and McKinley) and chewed up power (again both Merced and McKinley). The shift to .13 micron copper interconnect has brought it very comparable with x86 offerings.
Consider the 1 GHz Itanium 2 with 1.5 MB of L3 cache for $750. Then consider the Xeon 3.06 with 1 MB of L3 cache for $725 and then look at the SpecFP scores (not to mention POVRay scores), and on top of that, look at the 50-60W of the Itanium 2 1 GHz compared to the 80-90W TPC of the 3.06 Xeon, and all of a sudden IA-64 based chips are looking a lot better. And that's *without* mass market penetration, economy of scale and pushing yields to accomplish higher clockspeeds.
The only real thing holding back IA-64 (as with all previous chips) is backwards compatibility with x86 software. Without a good method of doing that, nothing will replace x86. Actual technical abilities to outperform x86 have been available for a long time.

"We are Microsoft, resistance is futile." - Bill Gates, 2015.
Related resources
Anonymous
a b à CPUs
September 30, 2003 9:06:44 AM

>Consider the 1 GHz Itanium 2 with 1.5 MB of L3 cache for
>$750. Then consider the Xeon 3.06 with 1 MB of L3 cache for
>$725 and then look at the SpecFP scores

Do you have numbers for the small cache 1 GHz Itanium 2 ? I know the 1.5 GHz 6 MB scores 2119. If I extrapolate to 1 GHz I get 1412 for a theoretical 6 MB 1GHz model, the 1 MB model will score considerable less. Now, lets look at Athlon FX: with just 1 MB it scores 1423.

And this is just SPEC FP, a bench that makes IA64 shine, with its extremely powerfull floating point performance. Most business apps, however require INT performance. Lets look at SPEC INT:
Itanium 2 1.5 GHz 6 MB:1322
Itanium 2 1 GHz extrap:881
Athlon FX 51 1 MB :1447

The itanium 1.5 is roughly twice as big, an order of magnitude more expensive an yet it doesnt beat a 1 MB bread and butter desktop chip. I stand by my point: x86 sweeps the floor with "high end" 64 bit cpu's in performance /$$$ and performance /mm². Deerfield isnt changing squat in this regard; the only area where it *may* have a lead, is FP performance / Watt.

= The views stated herein are my personal views, and not necessarily the views of my wife. =
September 30, 2003 8:48:30 PM

Quote:
Do you have numbers for the small cache 1 GHz Itanium 2 ? I know the 1.5 GHz 6 MB scores 2119. If I extrapolate to 1 GHz I get 1412 for a theoretical 6 MB 1GHz model, the 1 MB model will score considerable less.

Spec.org has listed a 900MHz Itanium 2 with 1.5 MB of L3 cache. Assuming linear scalability, that puts a 1 GHz Itanium 2 with 1.5 MB of L3 cache at around 1278 for SpecFP. Compare this with the highest submission for the 1 GHz 3 MB L3 cache model of 1451 and we have roughly a 11.97% difference that the extra cache makes.
Compare even this extrapolated score with the 3.06 Xeon at 1243 and consider that the Itanium 2 consumes 50-60W vs 80-90W.

Quote:
Now, lets look at Athlon FX: with just 1 MB it scores 1423.

With "just" 1 MB of full-speed, low-latency on-die L2 cache. And consider how long the chip has been released (is it even selling?) and the advantage of mass-market not to mention the profit margins gained from Opteron it has. Include that with the fact that it's on an SOI manufacturing process and consumes ~70W compared to the 50-60W UV Itanium 2.

All of a sudden, all of the advantages of x86's "mass market" is becomming less and less of an issue. The only issue I'd say, would be backwards compatibility.

Quote:
And this is just SPEC FP, a bench that makes IA64 shine, with its extremely powerfull floating point performance. Most business apps, however require INT performance. Lets look at SPEC INT:
Itanium 2 1.5 GHz 6 MB:1322
Itanium 2 1 GHz extrap:881
Athlon FX 51 1 MB :1447

Most performance-dependence (i.e database services, webservers, etc.) business applications rely more on SpecInt_rate and SpecFP_rate as they usually depend more on MP scalability than on single-processor performance. While Opteron has great scalability due to its NUMA-like Hypertransport network, the implementation is completely up to the manufacturer who makes the machine and not on the processor itself. A prime example is the scaling performance of SGI's Altix compared to say, HP's SuperDome.
With lack of desktop and workstation benchmarks (with the exception of POVRay, which Itanium 2 excels in), I could only go with SpecFP scores. You can feel free to look at the TPC-C or SpecWeb_SSL results for a more realistic business-class performance measurement.

However, we were (at least that's the impression I got) talking about the consumer impact of Itanium 2. As we are speaking of replacing x86 (which is primarily the consumer and workstation market). Both of which rely much more on FP performance nowadays than integer performance.

Quote:
The itanium 1.5 is roughly twice as big, an order of magnitude more expensive an yet it doesnt beat a 1 MB bread and butter desktop chip. I stand by my point: x86 sweeps the floor with "high end" 64 bit cpu's in performance /$$$ and performance /mm². Deerfield isnt changing squat in this regard; the only area where it *may* have a lead, is FP performance / Watt.

You seem to regard integer performance as "performance" and everything else as "specialized". It's the converse I think nowadays. FP performance for most of x86's main markets is the "standard".

I find your argument of only using the 1.5 GHz 6MB L3 cache Itanium 2 to be a strawman argument. You've picked probably the *worst* case performance/die and performance/watt case for Itanium 2 and the *best* case in both of those in terms of x86.
I could easily go and pick the worst case x86 (the 3.2 P4EE) and compare it to the best case Itanium 2 (1.4 GHz LV 1.5 MB L3 cache Itanium 2) and show differently.

Comparing best cases in both performance/die and performance/watt, x86 chips certainly aren't "sweeping the floor" with "high-end" RISC (well, VLIW) chips (although the 1.4 GHz LV Itanium 2 isn't exactly "high end", it's a workstation chip priced at around $1100). At least, not for the "bread and butter" desktop market (in which FP performance is crucial).
The Athlon64 FX has a die size of 192mm^2.
The 1.4 GHz Itanium 2 1.5MB L3 cache has a die size of 180 mm^2.
The Athlon64FX at 2.0 GHz has a TDP of <A HREF="http://www.amd.com/us-en/assets/content_type/white_pape..." target="_new">89W</A>
The Itanium 2 1.4 GHz 1.5MB has a TDP of <A HREF="http://www.intel.com/design/itanium2/datashts/25379501...." target="_new">91W</A>
Considering that the AthlonFX has 1MB of full-speed, high-bandwidth, low-latency on-die L2 cache, it's hardly as "lowly" compared to IA-64 offerings as you seem to imply through connotations.

Are IA-64 based products sweeping the floor with its x86 counterparts? No. Is it being "humiliated" like you've painted? Hardly. With the recent move to a more mature manufacturing process, IA-64 products are looking a lot better and very competitive with its x86 counterparts for the desktop/workstation market. Despite the fact that it doesn't have economy of scale and the latest manufacturing techniques (like SOI or Strained Silicon) to help it along.
The only real barrier, like I said, is backwards compatibility.

"We are Microsoft, resistance is futile." - Bill Gates, 2015.
!