Sign in with
Sign up | Sign in
Your question

Webcast of 64 bit hammer

Last response: in CPUs
Share
February 20, 2002 11:26:21 AM

About two years ago (if I read the date right) an AMD engineer gave a speech to students at Stanford University. The speech was about SledgeHammer, and it was fairly detailed. The speech can be viewed as streaming video. I highly recommend it.
<A HREF="http://murl.microsoft.com/videos/stanford/ee380b/000927... " target="_new">http://murl.microsoft.com/videos/stanford/ee380b/000927... </A>





Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.

More about : webcast bit hammer

February 20, 2002 7:01:49 PM

Great stuff !

Thanks .

<font color=purple>~* K6-2 @ 333MHz *~
I don't need a 'Gigahertz' chip to surf the web just yet ;-)</font color=purple>
February 21, 2002 4:42:36 AM

Btw I forgot to mention it is an hour long. So grab a snack. So anyone else think it was good stuff?

Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.
Related resources
Can't find your answer ? Ask !
February 21, 2002 7:11:48 AM

I stopped half way through, gonna finish tomorow, nice indeed.

"The Cash Left In My Pocket,The BEST Benchmark"
No Overclock+stock hsf=GOOD!
Anonymous
a b à CPUs
February 21, 2002 9:43:11 AM

interesting indeed.. one thing struck me, the guy keeps talking about SSE, but not SSE2. Was it not planned yet back then ?

= The views stated herein are my personal views, and not necessarily the views of my wife. =
February 21, 2002 11:05:52 AM

SSE2 was probably considered proprietary at that time. So he couldnt mention it. Besides, I think his main focus was architecture and not "bells-n-whistles" like SSE2.
If i heard him ocrrectly, he said the switch between 32 bit and 64 bit applications would take 40 clock cycles. That sounds like one hell of a hit. Did anyone else catch that part?

Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.
February 21, 2002 11:42:00 AM

Well...going on my own knowledge here, which may be a little outdated...

It's not all that major. IIRC, on a 486, a task switch between two 32-bit apps took a hundred or so clock cycles. This is mainly because of all the work required to save and restore the state of processes--reloading GPRs, instruction pointers, possible descriptor table switches, flushing CPU caches, etc. etc. A task switch is a lot of work, even from 32-bit to 32-bit. To complete it in 40 cycles, going from 32-bit to 64-bit or vice versa, is actually rather impressive IMHO.

Compared to simple integer instructions that only take one or two clock cycles (or less than one cycle, in some cases), 40 cycles sounds like a lot. But then again, it helps you appreciate the "creamy goodness" of multiple CPUs. :wink:

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?
February 21, 2002 11:56:42 AM

Your probably right Kelledin, 40 clock cycles is probably a blink of an eye ( how long is 40 clocks in seconds?)

BTW Did anyone else here catch the guy's reference to licensing: He said that AMD has licensed DDR AND RAMBUS.
Think there is any chance AMD will use RAMBUS? Or did they license it to hedge their bets against intel?

Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.
February 21, 2002 12:08:22 PM

Probably just to hedge their bets. The Hammer's current roadmap is to use DDR and DDRII memory controllers, and that's no easy thing to change when it's built into the CPU.

<i>If a server crashes in a server farm and no one pings it, does it still cost four figures to fix?
February 21, 2002 12:20:42 PM

2Ghz is 2 Bilion clock cycles a second! so 40 is like... not alot

hell... evrey branch-miss-predicton (which happens ever so often) on P4 takes 20 clock cycles (less if the micro-op could be loded from trace-cache) until the processor fully recovers (all stages are filled).

switches between 64bit to 32bit shouldnt be very often... even if you do run 64bit and 32bit apps together... even if the OS will switch between 32 to 64 evry say 0.01 second then its 40 Cycles hit evry 20 million Cycles...

to put it in Perspective....

This post is best viewed with common sense enabled<P ID="edit"><FONT SIZE=-1><EM>Edited by iib on 02/21/02 04:36 PM.</EM></FONT></P>
Anonymous
a b à CPUs
February 21, 2002 12:24:12 PM

>how long is 40 clocks in seconds?

Not much ;-) Lets assume Hammer launches at 2 ghz. That is 2000 Megahertz or 2 billion clock cycles per second (right ?). So 1 clock cycle takes 1/2 billionth of a second. So a 32 bit / 64 bit task switch would take 40/2 Billion the of a second. Put it this way, in one second, Hammer would be able to switch almost 50 million times from 64 bit to 32 bit mode.

Not that this means a lot you know.. God only knows how often such a switch would actually oocur. I have no idea, but having a 64 bit OS running 32 bit apps, this might be *very* often. Perhaps Raystonn could give us some insight (regardless of intel/amd/hammer/P4 differences)

= The views stated herein are my personal views, and not necessarily the views of my wife. =
February 21, 2002 5:26:39 PM

I only managed to watch the first 15 mins. I have got a really slow modem connection and the transmission keeps on cutting out :( 

Is there any way to download the file ? or do you know if there is an audio version of the lecture anywhere ?

<font color=purple>~* K6-2 @ 333MHz *~
I don't need a 'Gigahertz' chip to surf the web just yet ;-)</font color=purple>
February 22, 2002 7:21:32 AM

I found the link at the forums on Anandtech or xbitlabs, i forget which.
I guess I would play around with the URL. The speech was given at stanford. But if ya look at the URL it is from microsoft. So maybe you can dig thru microsofts wensite.

Benchmarks are like sex, everybody loves doing it, everybody thinks they are good at it.
February 22, 2002 8:27:07 AM

40 clock cycles at 2ghz is not that much.

"The Cash Left In My Pocket,The BEST Benchmark"
No Overclock+stock hsf=GOOD!
February 22, 2002 8:29:35 AM

Quote:
Your probably right Kelledin, 40 clock cycles is probably a blink of an eye ( how long is 40 clocks in seconds?)




40 clocks at 2ghz would be.


0.0000012 seconds.

(divided 60 by 2billion then multiplied that result by 40)


I think I got the math right, anyone wanna double check?



"The Cash Left In My Pocket,The BEST Benchmark"
No Overclock+stock hsf=GOOD!
February 22, 2002 1:46:26 PM

40 cycles would take 0.00000002 seconds on a 2GHz machine...

40 / 2.000.000.000 = 0.00000002 sec


/* The more you know, the more you realize how little you know */
February 22, 2002 2:21:33 PM

I think you have your math wrong parma.

The delay would cost 40 clockcycles, so you would not divide anything by 40, you would multiply the time 1 cycle would take by 40, as I did.


I believe my math is correct here.

"The Cash Left In My Pocket,The BEST Benchmark"
No Overclock+stock hsf=GOOD!
!