Slashdot Mirror
AMD Launches First 45nm Shanghai CPUs
arcticstoat writes "The wait for AMD's next-gen CPUs is finally over, as the company has now officially launched its first 45nm 'Shanghai' Opteron chips for servers and workstations. 'AMD's move to a 45nm process relies on immersion lithography, where a refractive fluid fills the gap between the lens and the wafer, which AMD says will result in 'dramatic performance and performance-per-watt gains.' It's also enabled AMD to increase the maximum clock speed of the Opterons from 2.3GHz with the Barcelona core to 2.7GHz with the Shanghai core. Shanghai chips also feature more cache than their predecessors, with 6MB of Level 3 cache bumping the total up to 8MB, and the chips share the same cache architecture as Barcelona CPUs, with a shared pool of Level 3 cache and an individual allocation of Level 2 cache for each core.'"
Does this mean that AMD chips are now competitive on price-performance with Intel's? I mean for a fairly high-end desktop or server; obviously different considerations apply in the embedded or netbook market.
-- Ed Avis ed@membled.com
"AMD Shanghai -- the perfect CPU for your newly-acquired botnet!"
Paleotechnologist and connoisseur of pretty shiny things.
The two companies take turns one-upping each other for the bleeding edge, but every time (10 years running) I've specced out a mid-range (home gamer, single CPU motherboard) to low-end (grandma's email/photo machine) machine, AMD's been the way to go. It's a lot like trying to decide which company's video boards to pick if you're trying to make a game machine without breaking the bank.
Some people are Intel partisans, some people AMD partisans. Benching them and looking at spec, I've consistently found that AMD's got faster chips (for the same $) up to the "sweet spot" in the curve where price starts shooting upwards during the times I've been buying, but I also know there were times I was not in the market when Intel had done a price cut and AMD hadn't caught up.
I'm not going to call someone an idiot for their CPU choice, as it's a long-term purchase decision that has to be balanced with other factors (motherboard choice, RAM, video board, power concerns, cooling solution, etc) anyways. In fact, I recommend consumers try to stay OFF the "bleeding edge" because they're basically throwing money away on it; even if you buy the latest, hottest chip right from the factory it's obsolete by the time you get it home. Your best bet is looking at the curve, because there's always a spot (usually between $150 and $250) where the price starts to jump up exponentially for only an incrementally "faster" product. Buy at the spot beyond which the relationship between price and performance fails to be linear and you'll turn out pretty happy.
If you're hinging your comments on the wafer size, you're blinded by Intel propaganda. Take a look at AMD's SPECjbb numbers, their cost per socket/core, and their threading for virtualization. Then perhaps you'll stop being an Intel shill.
---- Teach Peace. It's Cheaper Than War.
The first computer I ever worked on (as a data entry operator in the mid '70s) was an IBM S/360 mainframe with 64KB of "main" (physical) memory.
The first computer that I was a primary operator on, a S/360-135 plug-compatible 2Pi, had 768KB when it was delivered and was eventually bumped to 1.25MB shortly before I moved on to programming.
The computer upon which I wrote my first professional (COBOL) program was an IBM 3033 with a (for then) eye-popping 4MB of physical memory.
The first computer I ever owned was an RCA COSMAC with 4KB of memory.
The first DIY computer I ever assembled completely from parts (about 15 years ago) had 4MB of interleaved DRAM and a 256KB SRAM cache and was considered somewhat amazing by everyone who saw how fast it ran OS/2. I eventually boosted it up to 16MB
Now you get 8MB of on die cache with your four cores... And I still can't get a decent flying car.
It's about time... I mean, seriously. The CPUs coming out of AMD have stagnated in the last few years. The Phenoms are decent enough, I guess, if you have apps that can take advantage of the three or four cores, but they clock at slower than comparable X2s, and two cores is still the optimal point on the diminishing returns curve (on adding more cores).
I remember the 90s and early 00s when you were basically required to upgrade your processor every year or two or be hopelessly behind when the latest game came out. Now, I'm running the same machine I was back in '04, except with a new video card and an upgrade from a 3800+ (2.4Ghz) to a 4800+X2 (2.6Ghz) a year and a half ago.
I got curious how far I was behind these days, and found that as far as everything goes, a 4800X2 is still about as good a chip as anything AMD produces, only about 30% below the top chips AMD makes right now.
By contrast, Intel has the E8500 which is not only significantly faster, but is heavily, heavily OCable as well. I think Moore's Law has finally broken down for AMD.
How about AMD/ATI?
They have bee producing some very good chipsets for the desktop.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Of course it worked for Intel. Higher resolution lithography processes mean you can fit multiple cores in the same space as a single core from a decade ago. It means that the latency for critical paths is reduced, which means you can run the chips at a higher clockspeed. It means current consumed by transistor switching is reduced, so that chips can run at a lower power whilst maintaining or increasing throughput (thought interestingly leakage current increases as feature size shrinks).
Manufacturing process improvements are the number one driver of processor advances. It is obviously true that processor architecture changes, but mostly this is a response to new developments allowing more circuitry to fit in the same space. The latest Core processors have basically the same pipeline design as the original Pentium Pro. If you could go back and re-design the PP using our "new" architectural advances but older technology process, you would end up with a pretty similar design, since the process itself imposes such huge constraints on the architecture.
Don't know about you but try playing back a 1080p H.264 video file and watch it choke to death and then some.
Just an off-the-cuff calculation on my part shows power consumption dropped over %50 over Barcelona, clock-for-clock.
This is good news, because when AMD moved from 90nm to 65nm, their leakage was so bad that the power consumption only dropped around %10 clock-for-clock. Combine this with better cache architecture (larger, and faster), and AMD may have a winner in the server space.
I'm not sure if they're going to take back the desktop anytime soon. Intel doesn't have the FBDIMM downside on desktop systems, and I'm fairly sure that Shanghai didn't add major microarchitecure changes, so a quad-core Core2, let alone an i7, should continue to dominate the desktop.
However, it is nice to know that the market once again will have a choice in processors. AMD's 65nm offerings were spanked in terms of performance and power consumption by Intel's lineup, but Shanghai will at least compete on the power front, if not the performance front. We shall see what happens when AMD releases their desktop version.
Man is the animal that laughs.
And occasionally whores for Karma.
Comparing two CPUs based on clock speed alone is like comparing the speed of two cars by measuring only the RPMs of the tires. It won't get you anywhere ... you need to know the size of the tires as well!
Thus concludes my first /. car analogy. Thank you.
Despite any advances AMD makes in CPU's, they still have such a sub par selection of chipset vendors. I'm very happy that Intel makes both the best CPU's and chipsets at the moment. It makes the decision easy. Because if AMD came out with a killer CPU but I had to resort to an NVidia/AMD/VIA chipset to run it, that would be quite a conundrum.
Thriving on life. I've just lived there for four months and wish I could go back.
You haven't examined SPECjbb, then, have you? It's a Java-based business transaction kit that seems to have quite a bit of both fairness and repeatability; it's not easily manipulated, like others I've seen. After running more than several thousand runs with it, I find it pretty reasonable in terms of systems performance comparison, rather than motherboard/subsystem/peripheral benchmarks-- which shed light only on one specific characteristic of a machine, or are operating systems-specific. Admittedly, it doesn't do things with GPUs, network I/O, and the like.
But to call it a lie is specious.
---- Teach Peace. It's Cheaper Than War.
Or you can put the $500 you saved on the stock market, and by the time you need to upgrade you can use the money you saved, along with any capital gains and dividends to buy, um, a packet of waffles.
If you're buying "bleeding edge" you're not going to be satisfied with your purchase 2-3 years from now (about my replacement cycle for my personal box, and even then a lot of my components like hard drive / sound card / DVD drive tend to last through 3-4 iterations), unless your tastes suddenly radically change and you're no longer interested in the "bleeding edge" games you were trying to run.
Plus, consider the following two options:
#1 - "Bleeding edge" rig. Blow $900 on processor, $1200 on dual video boards, $400 on RAM, $800 or so on miscellaneous other components. Total system cost around $3000.
#2 - "Decent Gaming" rig, single $300 video board, $200 processor, etc. Total cost: $900 if you really push your luck.
I'll take my $2000, buy more games, take girlfriend to dinner, stick some in a rainy-day fund, etc. One of these years you need to run the numbers and then you'll figure out that the "savings" you claim are there from buying at bleeding edge aren't really there at all. Even if I spend $900 every 2 years upgrading my PC, it takes me 5-6 years to equal the cost of your rig, and I guarantee you're going to turn around and want to rebuild to get back to the bleeding edge because you'll be "disappointed" that your 2-3 year old "bleeding edge" machine is only getting 15 fps in the timedemo mode of CallOfUnrealCrysisDoomQuakeTournament 3: Yet Another Non-Scaling Tech Demo Masquerading As A Game.
So you're advocating Windows? I'd love to do H.264 decoding on the GPU in Linux, which driver and which video card do that for me?
I use XBMC, we're stuck with using the CPU for now but at least we can use both cores. So far the AMD CPUs haven't fared well with that software for full 1080P H.264 decoding either.
Build it, Drive it, Improve it! Hybridz.org
Accorcing to Wikipedia (http://en.wikipedia.org/wiki/Hamming_code#Hamming_codes_with_additional_parity_.28SECDED.29), a scheme that can correct 1 bit error and detect 2 is typically used. So it can correct single errors. The most common reason is some form of radiation (see http://en.wikipedia.org/wiki/Soft_error#Causes_of_soft_errors).
Against at least one of those (cosmic rays), even quality RAM is not immune. This said, only the vendors of quality RAM seem to be in the business of making the ECC version anyway. Which makes it more expensive, because you're buying more of the good stuff (the parity bits for the SECDEC algorithm require an extra chip on the module). But considering the overall price of RAM, that should be not a problem.
C - the footgun of programming languages
Sometimes it's cheaper to buy the new stuff than not to. I bought a new Athlon X2 5400 a little over a year ago to replace my Athlon 1.2. I plugged the whole system into a Kill-A-Watt to see the power draw. I calculated that it would take 10 months for the energy savings to completely pay for the upgrade. So, my faster computer is now not only free, but actually saving me money. Because of this, I have also upgraded my wife's computer, my kid's computer, and my server. My wife's computer and my server should be breaking even over the next month or so, and my kid's computer will have payed itself off in June.
I also downgraded my speakers. My old Klipsch 5.1 surrounds sound speakers sounded great, but they drew something like 45 watts. I replaced them with a generic set of 2.1 speakers that don't sound as good, but are more than adequate for the purpose, and I am now only drawing 2 watts.
You subscribe to the bathtub curve of reliability. For many components, this isn't an accurate model.
For many components, after you get past infant mortality, the devices remain consistently reliable. I've seen 386s and 486s that are still running, day in day out, today. PDP11s simply don't die, and there are some that are just sitting in a corner quietly doing mission critical tasks in industry.
All you have to do is identify common failure modes and do maintenance to mitigate them. For example, the dominant failure mode of PCs tends to be fan failure due to dust build-up, followed by hard disk failure. The first can be avoided by cleaning the fans regularly on a schedule chosen based on the rate of dust build-up. The second isn't really a dominant failure mode -- I don't know anyone who has ever had to buy a replacement hard drive. Regardless, to protect against failure, simply keep back-ups of important data.
It's been a long time.
When hacking my Iopeners, I learned you can pull the bios chip from a running computer, put in the bad one, then just re-flash it.
Just have to be careful putting it in not to short the pins. Worked great.
-Mike
-Unresolved symbol? Byte me!
The higher Numerical Aperature lithography tools are definitely helping for making narrow lines (hence faster transisters) for both Intel and AMD alike. However, the biggest advantage Intel has in the chip-making business is the use of hafnium for forming the gate of the transister. As Gordon Moore put it, "It's the biggest change in transister technology in the past 40 years."
The rest of the industry is feverishly trying to match/duplicate the hafnium process improvements which Intel discovered. Unless there's some equivalent breakthrough at AMD (which is highly unlikely), Intel will retain the crown for performance.
Disclaimer: I work for Intel, the above is my opinion and I am not a spokesman for Intel. Heck, I'm just a lowly peon. I'm not even authorized to tell you the time of day!
I'm an individual! Just like everyone else!