Sign in with
Sign up | Sign in
Your question

Prescott... Too many transistors!!! (hidden tech?)

Last response: in CPUs
Share
January 14, 2004 6:48:20 PM

This french site, <A HREF="http://www.x86-secret.com" target="_new">http://www.x86-secret.com&lt;/A>, has recently posted this very strange and strikingly simple calculation... About how transistor count is distributed in execution core and cache. So here it goes:

<b>Willamette</b>: 42 Million transistors. Spread across:
<b>13.8M</b> cache + <b>28.2M</b> execution core.

<b>Northwood</b>: 55 Million transistors. Spread across:
<b>27.6M</b> cache + <b>27.4M</b> execution core.

<b>Gallatin</b>: (P4EE) 169 Million transistors. Spread across:
<b>135M</b> cache + <b>34M</b> execution core (few thousand more for cache control)

These calculations consider that it takes 6 transistors for every bit of cache, and they work. As far as I know, these numbers are reasonable.

Now enter Scotty...

This beast has 125 Million transistors, according to Intel. 1MB of cache has to account for something like 54 Million transistors... And an improved Northwood execution core couldn't exceed 40 million transistors, even counting SSE3, and possible Yamhill technology...

But this execution core weighs in at 71 Million transistors! This might just be enough to reproduce a second entire northwood execution core!... This could also explain Intel's thermal issues....

Either way, there's too much smoke here... There has to be a fire around... and burning fiercely...

Come on! What can all those tens of millions of transistors possibly be doing there, Intel? That's way too much!...

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 14, 2004 6:55:37 PM

Quote:
God knows

Well... God and the Intel CEO... You're damned right! It'll be interesting to find out.

What is almost for sure is that there is something in there about which Intel hasn't told us anything...

Maybe Prescott is actually a dual core Northwood? This would make it unbeatable... Quite simply. Even A64 doesn't really stand a chance. I think this would be really scary... it's far more likely that it's just Yamhill technology... (taking up a pathetically huge amount of transistors, mind you... and AMD said it'd only take 2~3 million transistors...)

:evil:  <font color=red><b>M</b></font color=red>ephistopheles<P ID="edit"><FONT SIZE=-1><EM>Edited by Mephistopheles on 01/14/04 05:08 PM.</EM></FONT></P>
January 14, 2004 7:16:20 PM

Dont think there will already be dual core in the prescott.
The benchmarks we have seen so far indicate it is as slow as the P4 now.
Could be ofcourse, but the engeneers did a hell of a bad job with it, much heat and no performence, i dunno...
January 14, 2004 7:21:36 PM

Of course, whatever is in it is presumed to be disabled right now, coop... Much like HT was in Willamette already, but didn't get activated until the 3.06Ghz...

Hm... even so, if it has been disabled, then it shouldn't be dissipating this amount of heat, or should it?...

Oh well, maybe we should just wait and see...

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 14, 2004 8:11:03 PM

Hmmm... according to SoDNighthawk's New Math, which states that nVidia cards have many more transistors than comparable ATi cards; therefore nVidia is the faster card... The P4EE should kick major ass when it comes to the AthlonXP or Athlon64. Northwood should also kick major ass when it comes to the AthlonXP... yet SoD claims the opposite is true.

Care to clarify your stance on a few things, SoD? I mean the rest of us KNOW you're wrong... but it'll be fun to watch you talk your head out of your ass on this one.

<font color=red> If you design software that is fool-proof, only a fool will want to use it. </font color=red>
January 14, 2004 8:22:04 PM

Prescott has a dedicated SSE execution unit, a dedicated integer multiply unit, and probably some more...

It's like I've said many times before: Hyper-Threading is the future. Instead of searching for independent instructions in one thread, go look in several at a time. Current Pentium 4 Hyper-Threading technology just tries to keep every execution unit as busy as possible. But you can 'easily' go further with it by adding more execution units and allowing more threads to run in parallel!

What's the difference with dual-core? Well, with Hyper-Threading it can share things like the decoder, jump predictor, etc. So the number of extra transistors required can be relatively low, while still nearly doubling performance. It also means latency is of lesser importance, so they could have added more pipeline stages as well.
January 14, 2004 8:29:15 PM

Just some quick numbers

Barton 54m -> Athlon64 106m
Northwood 55m -> Prescott 125m

Give or take a memory controller weighed against SSE3 and 64bit Hyperthreading and you’re at least in the same ballpark may it be at opposite sides of the bleachers.

Dichromatic for your viewing plesure...
January 14, 2004 8:52:34 PM

Quote:
64bit Hyperthreading

That's what I said... There's something hidden in there, and most likely, it's 64 bit Yamhill technology... This is what I think.

Also, A64 has 106m transistors, but 54 of those are cache. So the execution core itself is about 52 million transistors. Plus, the memory controller isn't a simple matter; it requires lots of transistors.

Even so, Scotty's core still has about 20 million transistors more than A64, and more than twice the transistors in the northwood execution core. It suggests that there's a little more than added SSE3 in the main core, in comparison to northwood. Yamhill, maybe. Or maybe this is all too far-fetched. Only time (maybe several months) will tell... I guess...

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 14, 2004 8:57:53 PM

Quote:
The P4EE should kick major ass when it comes to the AthlonXP or Athlon64. Northwood should also kick major ass when it comes to the AthlonXP... yet SoD claims the opposite is true.

Actually, zoron, while I agree that you can't judge a processor based on transistors, I'm afraid this case is not at all like what you've said there. When comparing Barton's to Northwood's transistor counts, and when comparing ATi GPUs' to VPUs' transistor counts, you're comparing transistor counts between two very different architectures. What we're doing here is comparing two processors from the same company - Intel: Northwood and Prescott... So Prescott should be just a slightly revamped Northwood core, right? Because it's the logical evolution thereof...
So, what we're saying here is that the transistor count in the execution core itself suggests more architectural differences than we'd have thought, that's all.

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 14, 2004 9:03:26 PM

For your information, SSE3 is not going to make any significant difference in the number of transistors. It's just a small number of extra instructions. To implement them they just have to add a few more connections and a handfull of transistors in the already existing execution units.

What will make a difference is that the whole of SSE1-3 will have its own execution units, appart from the legacy FPU. They are supposed to do four floating-point operations totally in parallel.
January 14, 2004 9:22:40 PM

I was trying to point out that transistor count isn't a good way to guage performance... whether it be GPUs or CPUs. One of SoD's arguments for nVidia being superior to ATi is that nVidia had more transistors, so therefore it MUST be faster than ATi. We both know this isn't the case. So... using that logic on processors, Intel should have the fastest processors by a considerable margin... and we both know that isn't the case either.

I realize CPUs and GPUs do different things. But damnit, transistor count doesn't mean ANYTHING when measuring performance... I don't care what kind of chip it is.

<font color=red> If you design software that is fool-proof, only a fool will want to use it. </font color=red>
January 14, 2004 9:38:13 PM

I understand what you mean there, zoron.
Quote:
I was trying to point out that transistor count isn't a good way to guage performance...

I was trying to point out that we were not gauging performance. We were merely imagining how Northwood's architecture compares to Prescott's, if it's supposed to be the same execution core. Why is it that Intel's execution core is up to 71 Million transistors, from Northwood's 34, if it's the exact same architecture, with just a few minor tweaks? That's the point. I'm not comparing performance - which we actually don't even know yet - I'm comparing the architectures themselves. If Intel figured a way to get Northwood to the performance it has now with 34 million transistors in its execution core, and tweaked a bit, then how come they ended up with an execution core with more than twice the transistors?...

I'm not at all saying that Prescott has more transistors, therefore it must be faster. Not at all. I'm saying that these extra transistors have to be there for some reason, and it looks strange to have so many extra ones for a smooth and relatively small evolutionary step.

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 14, 2004 11:10:09 PM

Quote:

The P4EE should kick major ass when it comes to the AthlonXP or Athlon64. Northwood should also kick major ass when it comes to the AthlonXP... yet SoD claims the opposite is true.

Maybe the P4 has a lot of things disabled ? and they will enable it when they think its the good time, a few years ore so ?
January 15, 2004 1:56:51 AM

I havn't seen any thing on dual core Scotty. But have seen it for Tejas.
January 15, 2004 1:59:20 AM

ROFL We are all waiting on that.
January 15, 2004 4:00:08 AM

hmm, interesting

i didn't realize how much bigger Prescott's execution core is than the Northwood. I know that Prescott has potential 64-bit registers, dual 32-bit integer cores. And Advanced/enhanced hyperthreading. The reason for prescott being hot probably.

-------
<b><font color=red>Aquamark3 score: 28,208</b></font color=red> :smile:
<font color=blue><b>512mbDDR | Intel P4 2.4ghz 533fsb | ATI Radeon 9500 pro</b></font color=blue>
January 15, 2004 5:01:31 AM

Quote:
I was trying to point out that we were not gauging performance.


I REALIZE that... heh. Again, I was trying to point out SoD's stupidity. According to his math, P4 should beat Athlon... P4EE should beat Athlon64... all because of greater transistor count. He cited this as a reason nVidia beats ATi.

Did I explain it well enough yet? :tongue:

<font color=red> If you design software that is fool-proof, only a fool will want to use it. </font color=red>
January 15, 2004 8:53:01 AM

We also have to consider the possibility that the trace cache and L1 data cache has grown significally. That might be a big portion of that 71 million. It would again be in great benefit of Hyper-Threading...
January 15, 2004 10:00:17 AM

No, that wouldn't increase transistor that much. Not by tens of millions of transistors!...

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
January 15, 2004 10:01:12 AM

I got it alright, Zoron. :smile:

:evil:  <font color=red><b>M</b></font color=red>ephistopheles
!