AMD Talks About Internal Benchmarks for Opterons

ggruschow writes "AMD's CTO says their 2.0-Ghz Opteron (aka Hammer) beat a 2.8-Ghz Xeon (P4) on both SPECint2000 and SPECfp2000 tests, but was mixed against an Intel 1-Ghz Itanium 2 (details at ExtremeTech). IBM predicted "conservative" 1.8-Ghz PowerPC 970 scores, which fall in the middle of the pack (sweet for OS X). It's probably not a coincidence that AMD's news comes so soon after Gartner said x86-64 would fail. Even if Intel loses the performance crown again, their upcoming mobile processor is looking pretty spiff with its recently announced 1MB of cache. Sounds like next year might finally bring a worthy upgrade for my 486dx4-160."

Comments

486dx4-160?

[acehole]

You're weak my friend ;)

You've got no holding power... hell i've still got my Commodore 64 with accoustic coupler modem, and i'll hold onto it until I see something worth spending money on...

Yawn, wake me when it ships.

[Gldm]

Benchmarks are nice and all, but I'm getting kinda tired about hearing how great a CPU benches for about 6 months before I could even buy one with a sack full of money.

Not that I'm not excited about 64bit CPUs on the desktop, I could really find a use for one (I've got something interesting that likes to malloc more than 4GB sometimes).

Re:*sigh*

[tcdk]

For straight CPU intensive tasks it matters.

But for 99% of normal peoples taskes 10% whont matter.

But it's the edge and it has to be somewhere and it has to move.

My rule is that I upgrade when I can get a cpu that is twice as fast as my old one for about 1000dkr (130$/).

Thats possible right now (I've a 850Mh celeron), but I need a new motherboard, which kind of changes the rules.

How much is adequate?

[Deton8]

Oviously there is a market for super-fast processors to those of us on /., but aren't we at a point where currently available processors are fast enough for more and more user segments? What I mean is, people who do Word and Excel were happy along about 800 MHz and ordinary CAD people like me don't need more than about 2 gig. There are only two guys in my organization (running VHDL simulations day in and day out) who have any need for faster processors. Will we soon get to a point where the total market size of gamers and /. people will not pay for another processor spin?

Re:How much is adequate?

[ottffssent]

No-one NEEDs more than a P100 tops

Yeah, but only in the way than no-one NEEDs modern medicine, central heating, or citrus fruit during the winter.

On the other hand, I NEED faster than a Duron/600 for:
sending messages in ICQ (yup, sending a message is O(n) or O(n^2) - not sure which) with n the number of messages in your scrollback
Encoding MP3s - I spent over 2 hours this afternoon switching CDs every 10-15 minutes.
Recording TV - I can only record to divx at quarter VGA or less
Using Mozilla the way I want (with 20-50 tabs open at a time and 128M of RAM cache)
Using an encrypted filesystem (unless win2k's implementation is just horribly inefficient)
Opening / manipulating 500M images

Sure, I could plop an XP2200+ in here, but I spent $50 on the original CPU and I'm unwilling to spend more on another until Hammer comes out. A dual Clawhammer should be about 10-20x as fast as my current machine depending on app - a most satisfying upgrade.

C64 ?

[stud9920]

hell i've still got my Commodore 64 with accoustic coupler modem

You overcomfortable rich kid ! A C64 is just a toy with loads of eye candy. I am still doing it with my Difference Engine 2.0 and IPoAC (IP over avian carriers). More would just be superfluous luxury. Besides, shouldn't you have typed your message in all caps ?

Re:C64 ?

[roundand]

IPoAC (IP over avian carriers)

That would be RFC1149, right?

Benchmarks...

[e8johan]

Benchmarks are as bad as statistics. They measure nothing but how much you can tweak your CPU and compiler to fit that specific benchmark.


I would say that AMD may have an advantage for being more backwards compatible than Itanium, but I also feel that it is time for a change!


All major CPU manufacturers make proper RISC CPU already so why don't we find them in our ordinary computers? It is because the Windows codebase cannot simply be recompiled for a new target but has to be ported function by function (painful assignment, to say the least). Perhaps they can reuse 3/4 of the code, but still, there is a whole lot or rewriting and verification to do.

I have worked in a Tru64 environment (running Alpha CPUs) and I was surprised of how easy it was to get 95% of the Linux apps to properly compile and run. I didn't try to get Linux it self running but I had gcc running and that was enough.

What I'm trying to say is that the open source movement has proven that one can write portable code successfully and that it is time to make a hardware change. The serial ATA and AGP solutions from the PC are good enough, so is the PCI bus (lots of peripihals available) so I wouldn't change that, but simply make the standard computer run multiple RISC CPUs and a proper multi-threaded OS that can take advantage of that and then you'll have a performance boost that would make P4 look like a bicycle compared to a F1 car (ok, perhaps a Porche, but still, an F1 does 0-200kph in
While I'm at the subject. As we have bochs, it would still be possible to run Windows in a VM, no matter what platform we use, so all M$ users could be happy, or do as ACorn did (does), have a PC as a extension card, i.e. run a PC natively in a window, just used the *fast* RISC CPU for any real work.

Clawhammer

[Perdo]

Clawhammer (Athlon) has a single 16 bit wide hyper transport bus.

The workstation Sledgehammer (Opteron) has two 16 bit busses

The server Sledgehammer (Opteron) has three 16 bit busses

The spec results are as follows:

Spec_int

PIII1G 426
G4 1ghz 306
G5 937 (IBM PowerPC 970)
2.8Ghz p4 1010
XP 2800 933
Itanium 1Ghz 810
Power4 1300 804
Clawhammer 2.0 Ghz 1202

Spec_fp

PII 1Ghz 426
G4 1Ghz 187
2.8 Ghz p4 947
XP 2800 782
Itanium 1Ghz 1356
Power4 1300 1169
Clawhammer 2.0Ghz 1170

Opteron??? Higher than clawhammer considering the multiple hyper transport busses 1/2 mb L2 (compared to clawhammer's 256/512 l2) and dual on chip DDR memory controllers compared to Clawhammers single memory controller

Bootleg Powerpoint Presentation:

http://130.236.229.26/download/misc/AMD-Opteron. pp t

and

http://a26.lambo.student.liu.se/download/misc/AM D- Opteron.ppt

Read the Show notes! AMD failed to edit them out

Filename is AMD-Opteron.ppt google search it.

Includes a system that is an Opteron workstation dualed with a clawhammer that still presents itself as a single proc system. The clawhammer acts as a math co-processer :)

Re:486 160 mhz? (History lane)

[zensonic]

Your konwledge isn't that good. The fastest 486 in terms of Mhz was the Amd 5x86 - 133Mhz (4*33Mhz) chip. That chip easily overclocked to 160Mhz (4*40Mhz). In terms of pentium performance (integer wise) it was equivalent of a P75 at 133Mhz and of a P90 at 160Mhz (give or take a few percent)

In terms of performance the fastest chip that fitted in a socket 3 was the Cyrix 5x86 120Mhz, which (again speaking of integer performance) was equivalent of a P100.

I hope Hammer will fix the rc5 crippled speed!!

[Anonymous Coward]

I hope THIS mask rev of Opteron (Hammer) chip will be faster than January 2002 PowerPC G4 chips.

Currently, according to the RC5 benchmarks AMD is far slower than dual cpu macintoshes (half as fast). (source available for cor rc5 loops for most processors). RC5 was silently completed in June or so but a bug went unnoticed for a couple months, but the contest is over. They measured performance in units of "Mac poerbooks" in their press releases.

The Mac Dual 1 Ghz g4 is faster than all existing dual AMD motherboards in RC5 benchmark by almost 100%.

21,129,654 RC5 keyrate for dual 1 Ghz g4 system ! And Now apple sells dual 1.25 Ghz stock which would be even faster.

A dual 1800+ AMD MP gets only HALF as many as a Mac! 10,807,034 rc5 keys !

Funny "Mhz myth" there showing itself I guess... Apple now is selling even FASTER machines but with smaller caches and less fast read-write ram (it now uses DDR on newest boxes).

And the macs are using low power g4 chips meant for microcontroller usages with very little predictive branching and a simple 7 stage RISC pipeline depth. (macs complete many many instructions per cycle though, unlike Pentiums).

The mac I mentioned uses a 2 MB L3 cache and no AMD MP dual cpu boards I know about have any L3 cache at all, so maybe that is whay some common macs are over twice as fast, its not just altivec meager tweaks to rc5. AMS have similar , but less mazing vector ops.

Another reason the mac might be over twice as fast as an amd dual mp board is not just the 2MB l3 cache but the fact that mac can read and write to a cold page of memory simulatneously FASTER than any AMD MP designs which are biased for linear access and streaming. Many memory scatter benchmarks show this too. Appels newest DDR-RAM machines might not offer this feature though.

So basically, will the new Hammer systems be able to get close to speed for RC5 and other crypto tasks as the RISC based Powerpcs?

I really want to know. And I am so sad to see Slashdot reduced to fanboys modding down anything discussing tech subjects like this as "flames" all the damned time. This post is all informatinve and factual and my reason for asking is genuine.

http://www.research.ibm.com/journal/rd46-1.html has 5 LARGE technical articles on how the POWER4 chip was designed... in PDF form too. Even if you do not appreciate the Power4 (which apple is using a dual-core version of in many months) you might want to read these PDFs because they are all about chip design.

They put the floating point on the corners of the chip die to help spread heat, etc. Hundreds of interesting facts and pictures on at that site.

Top500.org lists Power3 dominating the cluster speeds of the top 500 computer clusters for memory+float speed. Power4 will soon start appearing in that list as well as the "lite" version with only 2 MB of cache instead of 4,6, and 16 MB.

Plus the new chip apple will start using announced yesterday, will have SIMD "VMX" or Velocity Engine added (Moto calls theirs"altivec").... only 90% of altivecs hundreds of opcodes will be offerred though.

With Pricewatch showing cheapest 800Mhz Itanium bare cpu at almost 8 THOUSAND dollars, and 3.5 thousand for the old itanium 700 Mhz, it does not take a financial genius to see why apple's workstations are selling so well nowadays.

Re:I hope Hammer will fix the rc5 crippled speed!!

[Colin Bayer]

Are you going decide which CPU to get next based on bit rotate performance?

On the PC, all of my work was so slow. That repeated multiplication and division by powers of two took forever. That's why I got a Mac, which has great shift left / shift right performance! Now I have more time to ogle the secretaries from the water cooler.

I'm Colin Bayer, and I'm an accountant at Arthur Andersen.

Re:I hope Hammer will fix the rc5 crippled speed!!

[acidblood]

I suggest you read the distributed.net Slashnet forum, where I explain why the G4 performs faster than x86 processors. Summarizing:

• RC5 is completely parallelizable, so you could theoretically do as many simultaneous operations as you have execution units on your processor, as long as there's enough registers to mask memory load latency. Obviously, there's many more registers on PowerPC architectures than on x86.
  
• The distributed.net core uses the Altivec SIMD extension on the G4, which has a useless rotate instruction, which serves absolutely no purpose that I know of on anything other than RC5 encryption. So I see Intel's point in not including a rotate instruction in SSE2: bit rotation is a completely useless operation except for RC5. Did I make my point clear enough? However, that makes it difficult to use SSE2, given the limited amount of registers available, coupled with the need to emulate a rotate instruction by means of shifts, ORs and an additional temporary register.
  

It must be clear that, if Intel had included an SSE2 rotate op, the P4 would easily beat a G4, not at the same clock speed, but given that a G4 can't scale as well as a P4 it wouldn't matter anyway.

Hammer can't get any better on RC5 without an instruction set overhaul. Athlons already do pipelined scalar integers rotates in 1 clock cycle, it's impossible to beat that.

Also, please do not generalize G4's distributed.net RC5 speed to a ``PowerPC superiority in crypto tasks,'' because it makes me want to laugh really hard at your cluelessness. SIMD is completely useless in real-world crypto applications: when you use a cypher in Output Feedback mode, which is how stuff is done in the real world when you're encrypting data instead of trying to break keys, you need to know the output of the last crypto operation to mix in the next operation. It should be obvious that you can't do operations in parallel now, so SIMD becomes useless and the Athlon goes back to being faster than the G4 at the same clock rate, and of course much faster on commercially available speed rates.

Oh, and the larger cache you mentioned has absolutely ZERO effect over RC5 performance. RC5 memory usage for each key being encrypted/decrypted is:

• number of bits in key rounded to the next 32-bit multiple (64 bits in RC5-64, 96 bits in RC5-72)
  
• number of cyphers round plus one, times 8 bytes (12 rounds in the RSA Secret Key challenge equals 104 bytes)
  
• 8 bytes for two temporary variables, which hold the plaintext before encryption and the cyphertext after encryption, or the cyphertext before decryption and the plaintext after decryption.
  

As you can see, even if you take into account loop control variables and whatever else, it boils down to less than 150 bytes per key. You could probably fit a 60-wide superscalar core on the P4's measly 8 KB L1 cache.

Re:*sigh*

[tconnors]

For straight CPU intensive tasks it matters.

But for 99% of normal peoples taskes 10% whont matter.

10% never matters. We regularly run simulations here that take a month. What is 10% on top of a month? 3 days. If you have already been waiting 30 days, what does another 3 matter? It probably corresponds to the weekend anyway.....

Re:*sigh*

[tcdk]

How many people are "we"?

If you are ten people, one of them could be fired, by your argument, without anybody noticing.

Let me turn it around - how many procent do you need before it matters? 12? 15?

But I agree, one can't upgrade everytime theres a 10% speed increase. One has to do the cost/benefit thing carefully first (and then ignore the c/b and just spend, spend, spend - the only way to get the economy back on track ;)

Windows XP

[droyad]

I hear that people are saying it would be difficult to port Windows XP to RISC chips (and new 64bit arch). This infact is not true. In the Windows NT family there are 2 features that make it easy:

1) It's mostly written in c/c++
2) The HAL (Harware Abstraction Layer) contains most of the platform specific code. As I understand it the kernel does not actually handle the hardware directly

Ofcourse I can see it going like this:
1) Apple, Intel, AMD and Moterola put forward new Chip designs
2) They ask MS to support it with their OS
3) MS picks Intel

--

$vi any_article_on_iraq
:s/iraq/microsoft/gi
:s/Weapons of mass destruction/Windows/gi
:s/Axis of evil/Redmond/gi
:s/In this post september 11 climate/Service Pack 1/gi
:s/Bush/Linux/gi
:wq

486dx4-160

[clinko]

486dx4-160? No wonder you crazy linux folks hate windows. You haven't bought a computer since 1995.

Old hardware, old software and efficiency

[XNormal]

At work I've got a 49000 line Microsoft Visual C++ project that compiles in 5.5 minutes on a 1700 MHz Pentium 4. That's right, about 150 lines per second.

Turbo Pascal used to compile at thousands of lines per second on machines with a clock nearly two orders of magnitude slower that tool several cycles per instruction instead of running several instructions per cycle.

Before you say something like "hey, but moderns compilers have optimizations yadda yadda" perhaps I should mention that this compilation time was with no optimizations and features like updating browser files disabled. With optimization it's even slower.

We're talking about four orders of magnitude difference in efficiency here. It's not all the compiler's fault, of course. The libraries and code use complex templates and multiple levels of definitions that make the compiler work much harder.

At each one of these layers someone probably said "It's OK if this is 10 times slower. It's easier to write and maintain, I'm more productive (or lazy) and the CPU is fast enough". Each one of these decisions may be justified *in itself* but they add up (or rather multiply up) to a 1/10000 difference in efficiency. Slowing the edit/compile/debug cycle reduces programmer productivity and code quality. Reduced code quality to more code bloat and even slower edit/compile/debug cycle and so on.

Damn, it's depressing.

Re:Old hardware, old software and efficiency

[Junks Jerzey]

Turbo Pascal used to compile at thousands of lines per second on machines with a clock nearly two orders of magnitude slower that tool several cycles per instruction instead of running several instructions per cycle.

Object Pascal (Delphi) still compiles that fast, only now it does include optimization (maybe not as hardcore as some C compilers, but still pretty good). Borland used to advertise speeds of 800,000 lines per minute, back in the day when a 266MHz Pentium II was a hot machine. For most projects, the compilation speed is *zero*. For medium sized projects, it's in the "barely perceptible" range (as in maybe 1/30 second). Very, very impressive.

Why is it so fast? There are a variety of reasons, in rough order of importance:

1. There are no header files. All exported identifiers are in the "interface" section of the main source file.
2. Interface information is always precompiled into a lean format, so there's no need to #include giant files (kind of like having all headers always be precompiled).
3. There's no preprocessor.
4. "Object" files are stored in a lean "almost linked" intermediate format, rather than traditional, bulky object formats. This makes the linker a very simple and fast affair, but linking can be the slowest part of building a C++ project.
5. The compiler, linker, and build manager are all in one executable, so there's no loading programs during compilation (typically for C++, make is loaded first, the compiler is loaded for each source file, then the linker is loaded at the end; yes, disk caching helps here).
6. Object Pascal is generally a cleaner language than C and C++, so parsing and optimization are easier.

And another ten, and another ten...

[Kjella]

I don't pretend to feel the difference between 2.0GHz and 2.1GHz. I don't "feel the difference" when going from a HD with 3x20gb platters to 2x30gb platters. I don't feel the difference between PC3200 and PC2700.

But I do feel it when I upgrade from an outdated system to a new one. And to know what kind of performance I could get for a reasonable* (*as defined by me ;) ) price, I do need to know what the state of the art is.

Maybe that isn't relevant to you, maybe your 486 / Pentium / Duron / Space heater does what you want it to when you check your email and type up your word document, but not for all of us. I know a few tasks where I'd like 4gb+ of memory, solid-state SATA drive and a multi-GHz proc+, or a dual, for that matter.

Large strides are best made one small step at a time. This is just another one of them.

Kjella

Everyone, look AWAY from the clock speed.

[Neil Watson]

I think the industry has to stop being blinded by clock speed. Before you can improve the speed of the chip there are still bottle necks on the motherboards (e.g. PCI bus, Disk controllers). Also, there is the problem of power consumption and heat.

I think a better approach for the future are smaller less power hungry modular CPUs. We've all seen the evidence of the clusters that makeup super computers. What if all standard computers came with 4 CPUs that used the same power as the P4 today? What if, instead of buying a newer faster computer, you could add CPUs like expansion cards but, at a reasonable price?

paradigm shift...

[john_uy]

i think the new release of hammer lines will be very difficult for amd. intel is one step ahead. if you see right now, they are already announcing next generation product lines in all fronts. like banias in cpu, ultra low voltage and integrated chips for small devices, extremely high speed chips for network devices.

i believe intel has shifted its focus in the battle of the desktop cpus. while amd is just playing catch up, intel now is already looking at what consumers will benefit from. maybe intel has realized that the speed today is an overkill for majority of today's needs. they are just speeding up their chips to keep up with moore's law.

but look at their products, right now, they are focusing on making things smaller, lightweight, ultra low power consumption, low heat devices, integration. the future is not on desktop computers requiring very high speed cpu but mobile devices such as phones, pda, tablets, etc. intel will be a clear winner (if only i have humongous money so i can buy intel stocks at discount.)

they have good engineers that produce good results. right now, they are already producing better chipsets for their server product lines, maybe a few years, they will no longer rely on broadcom's serverworks.

they are also picking up on their storage chips. from all the raid controllers in the market, i hardly see a card that does not have an intel 960 i2o processor or their new ixp processors.

their network and communication is very dynamic. like introducing 10gigabit products today (even with the downturn of telecoms.) enabling encryption and decription at 10gb/s is no joke. maybe a few years from now, we will see intel as chips in those network gear from cisco, et al.

they are now focusing on wireless integration. few years from now, capacitors and resistors will be in a silicon chip. it is the future, and they are very lucky to realize that. when the economy recovers, intel will clearly be a winner.

and for the server, i would want to say this. i believe amd will produce good cpu. but that is just half of the story, amd is not emphasizing any good chipsets/system to come with it including support pci-x at 133mhz with hotplug slots, interleaved memory with chipkill(tm), good server management, good integration.

(as one who decides what to purchase in a server,) amd must make a lot of effort before i will take them seriously. their cpu is not enough for me to get their system, yet.

let's just wait and see, but i see that intel will always be a step ahead. now for amd, the challenge is to be at par or even be ahead of intel.

PPC is not a great example of RISC

[be-fan]

I hate it whenever Mac-heads point to PPC and show how its such a great example of RISC that runs "all you're programs 2x as fast as the fastest Pentium4!" In all reality, the PowerPC line (not necessarily the POWER line) are very unimpressive. These days, a 1.25 GHz Alpha can still hold its own against a 2.5 GHz P4 in terms of floating point power. Yes, the same Alpha that has been neglected for the last half-decaded whose design has stagnated since the 21264 and whose process technology is antique compared to AMD's and Intel's. But the Alpha still keeps kicking x86 in the head. Yet, the PowerPC, running at the same 1.25 GHz, backed by the dual giants Motorola and IBM, built with leading edge copper fab technology, the second most common desktop RISC architecture (after x86 :) shipping in every single Apple computer isn't even competitive with the P4. Damn you DEC! Damn you to all hell!