The Inquirer: Moving backwards going forwards
|
|
|
AMD hits back at Intel's Nehalem
IT MAY HAVE taken AMD a few days, but Little Big Chip is finally recovering from its winding at the hands of Intel's new Nehalem server chip, Xeon 5500, and has thrown a few counter-punches of its own.
AMD's server chief Pat Patla led the counter-offensive with an outspoken interview to tech blog TechPulse360.
Asked whether Nehalem's slowest server chips could still beat any Opteron chip hands down, Patla said it seemed incredulous. After all, he pointed out, the slowest Nehalem chip is dual-core, whilst Opteron is quad core and Intel take great care to only buzz about benchmark results from their top-end parts.
Of course, Intel doesn't actually claim that its dual-core Nehalem chip beats Shanghai's fastest, so this is a bit spinful of AMD. What Chipzilla does claim, and is true, is that its slowest quad-core Xeon 5500 outperforms the fastest Shanghai.
Patla goes on to say that both QuickPath Interconnect (QPI) and hyperthreading are overrated, claiming QPI is nothing but a rip-off of open standard HyperTransport made proprietary with speeds which drop dramatically the lower you go on the server chip scale.
Applications which rely on high I/O and high memory throughput but don't need masses of compute power, for instance, would need to cough up for the fastest Nehalem processor to get the sort of high speeds touted by Intel, according to Patla, adding that AMD offers the same HyperTransport speed on all Opteron chips.
Meanwhile, while hyperthreading may appear enviable from afar, "real men use cores" say AMD, whilst pounding its hairy chest.
True, hyperthreading is a bit like a multitasking woman, but so what? SO WHAT?? Echoes AMD in its big, booming, manly voice. 'So, that pretty-looking multitasking woman is only giving you a 10 to 15 per cent performance bump for real applications workload,' says AMD (although we did paraphrase this slightly).
Not only is Nehalem a bit girly, it's also not a cheap date, as Patla points out. Whilst a Dell server with a Shanghai processor at 2.7GHz costs somewhere in the region of $3,000, a Nehalem 2.93GHz-packing server costs about $6,100, a not insignificant difference.
Of course, there are good reasons why Nehalem chips are so ludicrously expensive. For one thing, says Patla, the DDR3 memory they sport are still far too expensive and power-sucking. AMD reckons it will wait until 2010 when the latency has been lowered, and the price drops before going the DDR3 route.
Second of all, Nehalem counts three channels of memory whilst Opteron has only two, apparently making them 50 per cent more expensive in DIMMs and 50 per cent more power guzzling from a memory perspective.
AMD also poo-poos Intel's boasts that it can consolidate nine single-core servers on just one Nehalem server, noting, "They're not the only platform that runs virtualisation."
Patla notes that AMD's support for virtualisation platforms like VMware, Microsoft HyperV and Xen allow one dual-socket server to support on average 5 to 10 virtual machines.
As for Intel's claims of an eight-month ROI, AMD calls it "disingenuous" because it's not only about the hardware, it's also about software, lifecycle management, licensing, power and security.
"If your hardware is about 10 per cent of the cost of the total solution, how are they coming up with an ROI of eight months?", queries Patla, adding, "I'm sure they are doing the math thinking 'if you're buying the server today and you unplug 10 single core servers, the amount of power that you'd save would pay off this server'."
Intel, of course, claims Nehalem is actually a cash machine, with spinner Nick Knuppfer telling the INQ that its "extraordinary performance can be turned into actual dollars for the IT manager."
Knupffer reckons consolidating old single-socket servers achieves a 9:1 ratio while keeping the same level of performance, enthusing, "This can turn into a payback of just eight months! After that, Nehalem becomes a cash machine". WHOA!!!!
AMD, curbing Knupffer's enthusiasm a bit, does admit Nehalem has greatly reduced its idle power, but goes on to say that's just not very useful in a data centre, where parameters are typically turned up to 11.
But Knupffer, never happy unless he gets the last word, told us that, when it came to Nehalem, it really wasn't about anything other than the core.
"The uncore is a bit like a car gearbox, but the core itself is the engine. And Nehalem features an astonishingly powerful core, and the results speak for themselves," he said before we hung up on him and went to make some dinner. µ
Share this:
Share
PS: sorry, me again: I bet Bolt playing chess...anyone cares? But I didn't dare challenging him running...LOL...get it???
Folks, let's face it: Intel's Nehalem chip ROCKS! And I'm glad to see you pick Linpack to compare Nehalem vs Istanbul: it's the 1 (out of 100) benchmarks where AMD outperforms Intel...please, let's try and be serious when commenting...
We have completed HPL benchmarks on both. The results are detailed on our blog. http://www.advancedclustering.com/company-blog/high-performance-linpack-on-xeon-5500-v-opteron-2400.html
It looks like there has been a role reversal between AMD and Intel on performance. It is now just as important than ever to test your app before buying. There is still not a definitive "winner" between the two.
It really boils down to some simple arithmetics: on Nahelm you have 4 cores with 4 execution units each which, thanks to the HT and IMC, can actually be fed and do something useful in a multi-threaded environment. The Istanbul thing will have 6 cores with 3 execution units each, but their efficiency will be a bit under Nahelm's (because it's less efficient per unit when using 3 units for one thread then 4 units for two threads on HT). So in the end you have 16 units on Nahelm vs 18 slightly less efficient units on Istanbul, so it should be a textbook tie. Well, actually not, because the efficiency of the k10 units is higher then core2's, so to rise, i expect Istanbul to clearly beat Nahelm on a per-clock basis, but question is, will they be able to push it at least at 3GHz (and even so Istanbul will still have a 10% handicap vs Nahelm). All in all, the next half-year intel will reign supreme, but with Istanbul out it's gonna be a tie again.
Most of what has been posted and the claims in the article themselves are bunk. The issue of whose processor is faster was settled by DARPA and NITRD in the HIGH Performance computing competition. The SGI/Intel Nehalam proposal was eliminated first round in evaluations performed by Jack Dongarra, Horst Simon, Kathy Yellick and others who each have forgotten more than the collective knowledge of the computer tabloid press. The winners after the second round were IBM's Blue Waters with the P7 and the AMD/Cray Cascade with the Shanghai. The Networking and Information Technology Research development Office http://www.nitrd.gov/ has more resources, does more rigorous testing, and has more knowledge than the manufacturers themselves. Ask yourself the following two questions why 12months after Roadrunner went to 1 petaflop has Intel failed to produce a petaflop machine? Pleiades is 8 months away. Second if Nehalam is so great do the agencies where cost is no issue buy AMD instead of Intel? By the time Intel hits 1 petaflop AMD and IBM will be approaching 6 petaflops. The answer is very simple Intel doesn't scale past one socket due to limitations of RDMA. I would suggest that every one make the trip to Rome nest month to the IPDPS conference and listen to the presentation of this paper http://post.queensu.ca/~afsahi/PPRL/papers/HPCS08.pdf Intel's architecture is less than 50% efficient in its current configuration when doing nontrivial apps using more than 8k page files.
Comments like 'rip-off of open standard HyperTransport made proprietary' have nothing to do with one product being 'better' than the competition. That's just wining, and no one cares. And the irony of AMD saying Intel needs more cores...
8 month * 30 days * 24 hours = 5760 hours. Let's assume that a single core server plus it's cooling system consume 400 watt ( 0.4 Kilowatt ).
So 9 server operating for 8 month will need 5760 hours * 9 servers * 0.4 kW = 20736 kWh ( kW per hour ). If you pay 29 cents per kWh ( 0.29 $/kWh ) than the cost is 20736 kWh * 0.29 $/kWh = 6013.44 $. ROI depends on how much you pay per kWh and the power consumption of the server and it's cooling system. Spin giant AMD did there "calculation" for a server based in some were cold with the lowest kWh prices in the US.
AMD is low in performance because their microprocessor socket pins count (1207 LGA)still lower than intel (1366 LGA).
I still rememeber when they use to say "Real men own fabs" what happened with that? You might want to go to amdzone they are running low on Kool Aid.
""real men use cores""
Hahahahahaha can't stop laughing at that moron. Hahahahahaha
Real men deliver performance, it doesn't matter _how_ they do it. AMD lacks a feature and he says real men go without that feature. Next he'll say "real men use slower processors".
Reminds me of when AMD accused Intel of using dual-dies for their Core 2 Quads instead of making native quad cores. Intel's Paul Otellini's reply was ""The initial ones are multi-chip, but so what?' You guys are misreading the market if you think people care what's in the package."
Performance is what matters, and AMD has failed to deliver that. They can go on citing Intel's benchmark numbers cut in half or one-fourth (Hyper threading delivers 10% performance improvement?? STHU!). Nobody's going to buy their crap, except fanboys, of course.
The truth is slowest Nehalem quad-core still beats fastest AMD quad-core. AMD should go and die.
Read more about our thoughts on this claim here: http://links.amd.com/Hype.
Hey, can you let us know when AMD "hits back" at Intel with an actual chip, rather than just words? And nothing less than that. K, thanks so much.
There actually is one dual-core Nehalem, the E5502.
http://www.intel.com/products/processor/xeon5000/specifications.htm?iid=products_xeon5000+tab_specs
SoftWare Selection Makes Server its Particular flavour. Heres some more: Open Graphics Language & Open (standard) COmputer Language. Logic? has anyone tried Logic? if 2 Intels can equal 8 AMD, then seems cool stuff is more likely to work on integrated machinry. Yet with simple-soft Ink, you need turn off fouling software to return to aero, as learned to glitche, & speed makes it smooth. its just check box & its NT5 to NT6. So Much depends on Quality Made In Today. Certainly $75,00 to open bootcamp is steepe, as Intel corrected its' Erring 25M, probably with data gathers from todays Hot Blade, if you bought 100 Intel SSD, Why'd I'd Want any o' this. Intel will trade old for new if you call 2day. So theres quality of Service issue to. If its speed demon SSD, Maybe Speedie Demon MainBoard, too be BIG Booster. Domios, Dialbos' chowMein. Expect Fast Action thruout ''9 playing season.USB3 drashek ithinkicanihopeican DrashTel M.D.