Slashdot Mirror
AMD's Roadmap revealed
NoPants writes "It looks like the aces at Anandtech were able to get their hands on some of AMD's internal roadmaps. Anand has some interesting information including the new upcoming Socket 939 CPU standard as well as AMD's predicted release dates for Athlon 64 4000+ processors. Hopefully this will shed some light on what AMD is trying to do with all the different socket types..."
"Why in the world introduce an AthlonFX based on Socket 940, especially at the outrageous price, when you're moving to socket 939 imminently?"
Simple really. AMD feared that Intel was about to release the next revision of the P4 aka. Prescott. The 940-pin FX was an attempt to get something out the door ASAP.
Unfortunately that means that some people might be caught at a loose end when it comes to upgrade time, but that is not clear cut at this stage to my knowledge.
i dunno whart you're talking about, most AMD chips put out between 60-75 watts, and the intels are running up to 90-100 watts. people must drop the misconception that amd runs hotter because they really dont.
Unless you are a gamer or looking to compute Pi to the last digit
Or are wanting to do things like rip MP3s (trivial) or burn DVDs (non-trivial; technically it's the MPEG2 mastering, not the burning, that takes the CPU time). Developers, graphics artists, and most engineering can also use as much CPU as is available. For just a plain old file server you do very well by using the cheapest (in terms of purchase and run cost) that you can get. A webserver probably needs more juice. A database server definitely does. Trivial home use excluded of course. I'm not talking about trivial usages -- they can always be solved easily.
Wake me when a cheap "build your own system" RISC alternative hits the market.
When you realize that the core ISA of all x86 chips is RISC let me know. Not to mention that most of the classic "RISC" designs have deviated far from the "reduced" portion of that moniker. Looked at the Power or PowerPC ISA recently? RISC was created not because a reduced instruction set is inherently better, but because it allowed for a number of technologies such as pipelining, branch prediction, caching, and so forth to be implemented. Every single one of those is in x86 architecture now. Sure, the ISA is still a mess, but it's a better price/performance than anything else out there. All the naysayers have been disproven, time and time again. And yes, when I was a little college student I was horrified at the design of x86. Then I grew up.
Where's the boards I can cram 32 or more into?
Here.
Well, the Athlon FX line seems to have been aborted shortly after birth...
Actually, read the roadmap. AthlonFX and Athlon64 will both be Socket939 and dual-channel memory controller products by the end of 2004. The chips will be the same, except A64 will have 512k cache and AFX will have a full meg.
This actually makes a lot of sense, and should have been the way it was done from the start. No socket 754, no socket 940 (except Opterons), just a "mainstream" and a "performance" product.
the Performance Ratings are a very poor way to describe a processor...
...
I agree it's all marketing, but
As a data-point, I recently upgraded from an Athlon 2400 to the 3200 -- a 30% improvement in "PR," and saw my benchmarks go up almost exactly 30%. By benchmarks here I'm talking about a CPU bound computational mechanics code I wrote for my thesis. About as useful to everyone else as SPEC, but very relevant to me.
As usual, YMMV.
Running at full CPU load, an Athlon 64 3200+ uses less power than an Intel P4 3.2GHz. Furthermore, with AMD's Cool and Quiet power management enabled the Athlon 64 CPU slows down to 800MHz and drops to 1.275v when you don't need much CPU performence, ie, while I'm typing this message. ASUS has a nifty little program that displays the current CPU speed and core voltage on my desktop.
.
AMD CPU power requirements are expected to drop substantially when they switch to 90nm in the second half of this year. OTOH, Intel's prototype 90nm Tejas CPU burns up 150 watts
AMD chips haven't used more electricity than Intel chips for years. Pay attention.
BTW, Athlon 64 notebooks are out. $1,550 for a widescreen 64-bit notebook! I'm going to stick with my Athlon 64 desktop, at least until I come up with an excuse to buy a portable. Really, I am...
Correct me if I'm wrong, but I was told the PR xxxx+ was in comparison of what a Athlon Thunderbird clock speed would be.
>* By the way, just try a 3 Ghz processor for a while (maybe a week or two). Then go back your 700 Mhz system. You'll see the light.
:-)
Disagree. I use a 2.4Ghz at work and a 533Mhz at home and can't really tell a difference, except when (un)zipping files, or installing software(maybe 1% of my total use of the machine). My home machine can play music, games, surf the web, edit docs, etc. just fine.
> Faster boot times, faster archive extracing, faster application start times.
I think the faster disks in newer machines are probably the main reason this is faster, except the archive extraction, which is CPU intensive.
>Also, most developers need all the juice we can get.
You must be a Java programmer.
-Comedian
If you need better performance it may be another reason to switch to linux
Only Microsoft didn't catch up, but who cares ;-)
Regards
Unless you want to factor in things like power consumption....
Feel free to. x86 is still cheaper. Equivalent speed systems don't use vastly less power. Best number I've seen is ~75W, which is 3/4 of what a P4 or Athlon64 uses. That's not an abundance in savings.
To put it clearly -- 25W saves you 219 kWh/year (assuming it's on 24 hours a day year round (365.25 days/year)). If electricity costs $0.10 kWh then that's a savings of $22. Wow.
And that, of course, is assuming that the CPU is fully loaded the entire year. If you aren't using 100% CPU power then the CPU will reduce its power draw -- not massively (these aren't laptop CPUs), but somewhat.
You can point to other CPUs that draw less power than a P4/AMD64/970, but they cost vastly more to purchase. The PowerPC 970 is already overly expensive -- particularly given that AMD has already cut prices on the Athlon64 and Opteron lines since introduction.
But the PowerPC line has superior signal processing capabilities these days
According to who? And why? SSE2 provides plenty of instructions for signal processing, and SSE3/PNI will fill in a couple of the last remaining holes. AMD64 also doubles the number of registers for SSE2 as well as the general purpose registers.
Exactly what advantage does the PowerPC have for signal processing? They do have a nifty multiply-add instruction that is missing on x86/SSE2, but on the flip side their vector processing engine (Altivec) seems to be limited to single precision floating point math unless I missed something.
is much easier for compilers to cater to
Ohh, so THAT'S why there are WAY more x86 compilers out there than PPC ones.
And runs cooler and more efficient than the X86 offerings
The PPC 970, used in Apple's Powermacs is listed at a "typical" power of ~48W at 1.8GHz. Maximum power is probably a good 20-40% higher. In other words, power consumption is pretty much identical to AMD's Athlon64/Opteron line and pretty darn close to Intel's P4 line. It also has 52 million transistors, nearly the same as the 55 million transistors of the P4. The IBM Power4+ consumes a ton of power, somewhere up around 100W.
AFAIK, x86 units since the i686 have all used a RISC-like core that runs x86 ops by breaking them down into micro-ops and reconstituting them. It -works- but whay do that when the real thing is available?
Why do it? Becuase hardware is cheap, software is expensive. They also do it because it doesn't really make a difference, ALL chips start by decoding their instructions, executing them and then putting it all back together again. It's not just an x86 thing.
The main downsides to x86 are as follows:
- Too few general purpose registers and restrictions on what registers can be used for which instructions
- 32-bit pointers
- Stack based FPU
- Somewhat more complex instruction decoder
AMD has addressed the first two points with AMD64. They've extended the general purpose and FPU (SSE) registers from 8 to 16 and removed the restrictions on what registers can be used for which instruction. Not quite the 32 registers of some other architectures, but still pretty good.
The stack-based FPU is in the process of being replaced by SSE2. While SSE may have started out just as a vector SIMD engine, with SSE2 it's really a full fledged floating point unit to replace the old x87 unit.
The more complex decoder has also been largely addressed by Intel in their P4 design with their trace cache. This stores already decoded instructions, so about 90% of the time the decoder is not needed (and therefore can be made much simpler).
I don't know much about processor design, but one problem of the IBM POWER series is that they are NOT very streamlined. They in fact use more transistors to do the same job as other designs (I think the Alpha is a good counter example). This is because IBM uses automated design techniques. The tradeoff for the less efficiency is faster development. see for example,
o ck .html
e si gn_automation/ddh.pdf
http://www.research.ibm.com/journal/rd/461/warn
http://pix.cs.olemiss.edu/csci523/64bit.html
http://www.research.ibm.com/compsci/spotlight/d
I'm sorry i can't relocate the original article, but maybe this helps.
The marketing of the Athlon64 FX has become a bit confusing. It kind of made sense for the initial launch to combat a precieved weakness of the design compared to Intel's P4 though. With the P4 you get up to 6.4GB/s of memory bandwidth, while the first Athlon64 would only have 3.2GB/s of memory bandwidth. Now, it turns out that the extra bandwidth doesn't actually buy you much on most applications, but this was seen as a weakness, hence the Athlon64 FX. With Intel bringing out the P4EE to compete with the FX, now AMD might need to keep the chip, even if it isn't a worthwhile product (The P4EE isn't a worthwhile product either).
*Is it really a good idea to have the memory controller on the CPU?
Yes, yes it is a good idea. A VERY good idea in fact. Memory latency has only improved by about one order of magnitude in the past 15 years. Meanwhile everything else in the system has gone up by at least two orders of magnitude. Virtually everything that is being done in CPU design these days is to hide memory latency (larger caches, out-of-order executation, branch prediction, even SMT).Integrating the memory controller reduces latency by 20-30%. At 2.0GHz this makes a BIG difference (this is the main reason why a 2.0GHz Athlon64 is faster than a 2.2GHz AthlonXP), at 4 or 5GHz the difference will be huge.
but it hasn't lowered mainboard costs
You can buy new Athlon64 motherboards for only $100, only 3 months after the chips release. It took ages for Athlon or P4 motherboards to reach that price point. What's perhaps even more impressive is the dual-processor boards that are only $200. In short, it HAS reduced motherboard costs, whether you've noticed or not. It also means that ALL Athlon64's support ECC, chipkill and a few other nifty reliability features, regardless of how badly VIA screws up their chipset design.
*Why in the world introduce an AthlonFX based on Socket 940, especially at the outrageous price, when you're moving to socket 939 imminently?
The Athlon64 FX was a bit of a last minute decision I believe. They found a marketing weakness and wanted the quickest and easiest solution they could find. The answer? Sell your server chip as an "enthusiast" chip. Intel did exactly the same thing for the same reason with the P4EE.
Also, it's actually VERY normal to switch sockets soon after releasing a new processor. Intel's upcoming Prescott will use Socket 478 for only about 6 months before switching to socket 775. The original P4 used Socket 423 for a very short time before switching to socket 478. The original Athlon used Slot A for a year or so before switching to Socket A. The PIII came out in Slot 1 form but then switched to Socket 370 about a year later. The Celeron followed the same path a couple years before.
I think it would have been more of a slam-dunk as a platform and a "brand" to release Athlon64 as all dual-channel, all Socket 939 (or some standard), and left Opteron as the high-end platform. Any other takers?
In retrospec that might seem like a good idea, hindsight is 20-20 after all. However the original split of ALL Athlon64 chips being socket 754 and ALL Opteron's being socket 940 seemed to make the most sense when AMD was desigining them. It wasn't until market conditions changed and a new perceived weakness was discovered that AMD felt they need a consumer chip with a 128-bit wide memory bus. By that time the chip was already late to market and designing a new socket would have added more delay to the equation.
There's also the question of budget chips. AMD hopes to move their entire product line to the Athlon64/Opteron platform by the end of 2004. That means they need a budget chip, and socket 939 with it's 128-bit wide memory bus is problematic for that. Hence the continued existance of Socket 754 and the AthlonXP for that platform.
justified by a simpler motherboard; the 940 uses a six layer motherboard, while the 939 only uses a 4 layer one.