AMD Athlon 64 Performance Preview

k-hell writes "It seems like X-bit Labs have gotten their hands on an 'engineering sample of the AMD Athlon 64 2800+ processor'. Damage at Tech Report is writing that 'This is really fun, but I am a little concerned about their memory latency numbers.'"

Comments

here's a mirror

[linuxbaby]

temporary mirror here

Context Context Context

[robbyjo]

You should include the full quote of Damage, because just quoting out of context can be misleading. Here's the full paragraph (emphasis is mine):

This is really fun, but I am a little concerned about their memory latency numbers. They haven't specified what units those numbers are in, but latency numbers come out of programs like cachemem in CPU cycles. Obviously, processors with higher clock speeds will see more clock cycles pass per second than processors with lower clock speeds. One must convert those numbers into comparable units, such as nanoseconds, in order to compare CPUs at different clock speeds. I do expect the Athlon 64 to have low memory access latencies because of its integrated memory controller, but I don't think the gap will be so great as the X-bit numbers would seem to indicate.

So, the worry is about the units the latency numbers are expressed in. And when you'd see the numbers below, you get an idea why it is so:

Athlon 64 2800+

• Mem read speed: 2610.2 MB/s
• Mem write speed: 1099 MB/s
• Mem copy speed: 1541.7 MB/s
• Latency: 96

Athlon XP 1.6GHz

• Mem read speed: 1747.8 MB/s
• Mem write speed: 1156.9 MB/s
• Mem copy speed: 1244.8 MB/s
• Latency: 165

Pentium 4 2.8C

• Mem read speed: 3193.5 MB/s
• Mem write speed: 1320.5 MB/s
• Mem copy speed: 2678.6 MB/s
• Latency: 260

See it for yourself.

Re:Context Context Context

[zenyu]

Athlon 64 2800+ (true clock 1.6Ghz)
latency: 96 cycles or 96/1.6 => 60 ns

Pentium 4 2.8C
latency: 260 cycles or 260/2.8 => 92 ns

So there is a 33% improvement, which is cool. (i.e. the P4 is 50% slower)

The SSE2 instructions were pretty much in equal to the P4 in throughput per cycle, that is as a SEE2 processor it performs like a 1.6Ghz P4... Hopefully they can push the clocks up as fast as Intel has with NetBurst.

Re:best processor/price?

[spike hay]

At $500, you have just limited yourself to a mid-ranged Athlon or P4.

More like very high range. A quick glance at Pricewatch will show you that currently, the fastest Intel, 3.06 ghz, costs $388. The fasted Athlon, the 3000+, costs $320. Even the 2.8 Xeon is $425.

Single page view

[TeknoHog]

Printable version here.

Re:best processor/price?

[atrus]

Note this is for Motherboard/Processor/RAM, so, if you get a $425 processor, you won't be finding a board and good ram for it for $75

For those who don't follow every AMD move.

[WoTG]

Remember that this "preview" probably violates one or more NDA's, and it is of a desktop x86-64 chip that is scheduled for September release. In the meantime, it's bigger brother, the Opteron, who has more memory bandwidth, (usually) more cache, and multiple processor support will be released in less than a week (Tuesday to be exact).

Now the reviews that out in 4 days time should be much more interesting reads. I expect to see someone do a solid x86-32 vs. x86-64 comparison using Linux, maybe other OS's too. And yes, probably even Quake frame rate results. =)

Re:Check out date on processor.

[Anonymous Coward]

To add more fuel to the "its only an engineering sample", check out the date on the

processor [xbitlabs.com] itself.

Imagine, with nearly two years of time to improve on this piece of silicon just what is in store for the Clawhammer. Personally, i'm waiting for it so I can finally upgrade my Athlon 600.

The 2001 date has nothing to do with when the CPU was manufactured. The brand new Barton 3000+ (and every Athlon) is dated 1999.

This Athlon 64 revision is only months old, not years.

Re:How old is this "'engineering sample"?

[Anonymous Coward]

The manufacture date shows as 0301 (jan 2003) and copyright applies to the text/picture on the chip not the tech inside.

Re:My Observations

[captaineo]

No, performance in general would go *down* if you simply re-compiled existing code from 32 to 64 bits. (this is because pointers would double in size, and executables would grow slightly, increasing the cache footprint).

The only ways to gain performance moving to 64 bits are:

1) re-write software that needs a >4GB working set (e.g. databases) to use 64-bit pointers rather than paged or segmented 32-bit addressing

2) re-write high-performance integer code to process values in wider chunks (although you can do this on today's hardware with MMX and SSE)

The great part about AMD's new chips is that you will be able to run a hybrid 32/64-bit system. 64 bits for the few programs that actually benefit from it, and 32 bits for everything else. Running an entire system at 64 bits, although cleaner, is a loss for most mundane programs (cat, grep, bash, etc).

The two ways of looking at a half-empty glass

[vlad_petric]

The tests are pretty much showing that an Athlon-64 does not outperform a 32 bit Athlon when running 32 bit apps.

But here's another way to look at it - Itanium also has an x86 layer, but because it's really just an emulation, its performance sucks.

So I view this as a huge success. Why ? Because an Athlon-64 will be able to run "legacy" 32 apps at the same speed, while 64 apps will run faster.

You'd probably wonder why this is the case. Well, IMNSHO it's not because of the wider registers/ALUs, etc, but because of other improvements to the Instruction Set Architecture, like the 8 extra registers (16 total). Because you only have 8 registers on a regular x86, compilers can register-allocate very little. Adding 8 more registers means that you can keep more stuff in the register file, and you don't have to go to the stack (data cache) every single time.

Re:Check out date on processor.

[rabidcow]

The one that I'm looking at is 0301 (2nd line of numbers/letters) which I will guess to be 1st week 2003.

Exactly, date codes on chips, which tell you the date of manufacture, are usually 4 digits: two digits for the year, then two digits for the week in that year.

If you RTFA (as opposed to just looking at the pretty pictures), they say, right under that image: "The production date in the next line of the marking indicates the beginning of this year."

This is pretty standard, I can pull out my old 8088 MB and read the date code off the processor: 8937 (1989, 37th week) You can find similar date codes on most chips and PCBs. (eg, that 8088 MB has 8945 printed on the back)

Re:3DNow!: Cause of Slow Clock Frequency in InnerC

[RzUpAnmsCwrds]

3DNow! is essentially a subset of SSE. Removing it would not save a signifigant number of transistors.

Re:No SMP??

[Slack3r78]

You mean this one coming out next week?

Re:UT2003...

[mabinogi]

You can't.

The UT2003 they tested was the normal 32 bit version.
And as they said, it already favoured the Athlon XP, and the Athlon64 is very similar to an XP.