Magny-Cours effect : Supermicro packs 96 CPU cores in a 2U case

[informal]

Continuing with its aggressive expansion plans, Supermicro launched its servers and for the Magny-Cours i.e. Opteron 6100 Series processors. Among a series of different form factors, we noticed one that looked a bit special.

Supermicro's eight Socket server for AMD's Opteron 6100 seriesIn the past two years, Supermicro started to put blade motherboards into 1U rack chassis and internally connecting them with a high speed bus. Given the hot plug nature of the motherboards in question, downtime was non-existent if one of motherboards would go awry - a nightmare for the server admins. With the Opteron 6100, Supermicro launched a "2U Twin2 system" - no less than four hot pluggable motherboards with two sockets each. Long story short, this setup will support 96 processing cores and no less than 1TB of DDR3-1333 of system memory. Charles Liang, President and CEO of Supermicro stated that "The introduction of quad-channel DDR3 on these new 8- and 12-Core servers doubles the memory capacity and can accelerate memory performance up to 66%, especially for large data sets."

According to AMD's specs, bandwidth per processor is 42.66 GB/s which would mean that this 2U can offer a staggering 341.25 GB/s of aggregated bandwidth. Unfortunately, that is what DDR3 at 1.33 GHz can provide in four DIMM per Socket mode. The truth, though, is somewhat different. By clocking the memory controller and the L3 cache at only 1.8 GHz [well-known painful spot of K10/10.5 architecture], AMD crippled its own bandwidth to "just" 28.8 GB/s per socket, and this server will offer you a grand total of 230.4 GB/s. All in all, these are quite impressive figures, even with this self-castrated processor.

Supermicro's power supplies are some of world's best self-manufactured 80 Plus Gold certified 1400W power supplies. Just like everything else, the whole setup is being manufactured in Supermicro's facilities in San Jose, CA. In a way, it is interesting to see how a progressive company such as Supermicro concluded that off-shoring the manufacturing to the Far East doesn't pay off in the long run - at the same time when its competitors moved even the design teams to the Far East or just hire Supermicro or a similar company to create a product for them.

Naturally, this high-density system cannot offer GPU support, however there are 12 3.5" Hot-Swap HDD Bays [three per node]. In order to connect the four motherboards, you can opt for a 40Gbps QDR InfiniBand connection.

http://www.brightsideofnews.com/news...a-2u-case.aspx



Notice the power supplies in these servers .MC is proving to be one extremely power efficient chip.

[[XC] Oj101]

Hmmm, at 2.1GHz that's 201.6GHz of crunching power in one machine - more than ten times mine

[Russian]

Notice the power supplies in these servers .MC is proving to be one extremely power efficient chip.

Nice copy paste there.

[god_43]

link?

[onex]

about 10k for a system like this...
should be very productive .

[mstp2009]

We own two of three of these setups, just older mobos (two are Istanbul based, one is Nehalem based).

They are NICE for HPC Cloud Computing, and the PSUs are about as efficient as ANY I have ever used.

[informal]

Nice copy paste there.

Sorry,forgot the link
http://www.brightsideofnews.com/news...a-2u-case.aspx

[BatteryOperated]

10 years from now, this system will go the way of quad pentium pro.

[3NZ0]

10 years from now, this system will go the way of quad pentium pro.

Wow. Hardware gets outdated and outperformed. Who'd have thought it?

[K404]

Thats a sweet bit of engineering

Wonder if SuperMicro could be tempted to expand further into the desktop market with PSUs. Brand power + reliability would serve them well

[josh1980]

Hmmm, at 2.1GHz that's 201.6GHz of crunching power in one machine - more than ten times mine

You do realize that although if you add the total Ghz of all of the processors to get your mathematical number, it is NOT equivalent to a 201.6Ghz processor, right? Not by a VERY, VERY long shot. Think about this...If some 3D process took a 2Ghz computer 5 minutes to perform a task, your "201.6Ghz" computer would not do the same work 100x faster(3 seconds by your calculations). There's dozens of reasons why your math doesn't work for real actual performance gains. Not to mention that all cores are not created equal computationally, and you will have certain limitations of the memory performance.

But don't get me wrong, I'd LOVE to have a 201.6Ghz computer, but I fear that will not happen in my lifetime. There's a reason Ghz hasn't moved up in the last 5 years, just more cores. As frequency goes up there's limitations that mankind has not been able to solve, get around, or cheat our way through.

[Manicdan]

^per clock performance has also increase substantially too, 3ghz today is not the same as 3ghz K5. even before DDR2 came out we had to start useing theoretical number, aka 2500+, even though it was only 1.8ghz (i think, or maybe it was a little faster)

todays 3.4GHz Phenom II core is probably as fast as 7-8GHz would be from 1999, the only way to really tell is to use a very old benchmark, o like SuperPi, and downlock the crap out of a chip to match speeds from back then.

[Particle]

It's funny trying to do that, too. Modern chips are some times unstable at such low frequencies. hehe The late 90s was a fun time, wasn't it? Back when every year or so you could nearly double the clock speed of your previous processor. It was a lot more obvious to see the difference back then too, since performance in general was marginal compared to today.

[josh1980]

Particle-

I was just discussing the observation you mentioned yesterday. I was wondering if the reason why processor performance was so much more "apparent" back then compared to today is:

1. Programs were much more efficient in the 1990s because there was very limited RAM, HD space, etc. Today's processors can remove those inefficiencies via the CPU's branch predictor and other such optimizations in the CPUs of today.
2. Back then hard drives and other storage media could dish out data about as fast as the CPUs could feasibly process it. By today's standards, and with the advent of SSD it has become very apparent that hard drives are a MAJOR hindrance of performance for the computer.
3. Combination of the above.

With the advent of SSD and the new architectures being released by Intel and AMD, I have been wondering if we are hitting a point where CPUs are so fast that the average user really doesn't need all of the performance they can buy at the store. I'm seeing more and more computers that are not being replaced because that dual core processor is plenty fast enough. Typically, when more processing power is available via Intel or AMD, software engineers find a way to make consumers "need" that power. But, there's a limit to how much we can "need" at any given time.

What computer user (not an xtreme user) would have Outlook open checking email, sync their iPod via iTunes, chat online, surf the net, and play a game of solitare simultanously? Even doing all of these things simultaneously wouldn't max out a dual core computer. Are we looking at a possible "pullback" of the computer industry because we've exceeded our own desires?

Of course, this is all just a theory I have questioned. I know, this is really deep thinking, but that's how I roll.

[Motiv]

Particle-

I was just discussing the observation you mentioned yesterday. I was wondering if the reason why processor performance was so much more "apparent" back then compared to today is:

1. Programs were much more efficient in the 1990s because there was very limited RAM, HD space, etc. Today's processors can remove those inefficiencies via the CPU's branch predictor and other such optimizations in the CPUs of today.
2. Back then hard drives and other storage media could dish out data about as fast as the CPUs could feasibly process it. By today's standards, and with the advent of SSD it has become very apparent that hard drives are a MAJOR hindrance of performance for the computer.
3. Combination of the above.

With the advent of SSD and the new architectures being released by Intel and AMD, I have been wondering if we are hitting a point where CPUs are so fast that the average user really doesn't need all of the performance they can buy at the store. I'm seeing more and more computers that are not being replaced because that dual core processor is plenty fast enough. Typically, when more processing power is available via Intel or AMD, software engineers find a way to make consumers "need" that power. But, there's a limit to how much we can "need" at any given time.

What computer user (not an xtreme user) would have Outlook open checking email, sync their iPod via iTunes, chat online, surf the net, and play a game of solitare simultanously? Even doing all of these things simultaneously wouldn't max out a dual core computer. Are we looking at a possible "pullback" of the computer industry because we've exceeded our own desires?

Of course, this is all just a theory I have questioned. I know, this is really deep thinking, but that's how I roll.

I'll add the tl;dr for you

tl;dr: We don't need more than 640k of ram.

[Manicdan]

I have been wondering if we are hitting a point where CPUs are so fast that the average user really doesn't need all of the performance they can buy at the store.

i think thats about it. no one really needs more than a 2.0GHz duel core cpu to feel like everything is fine, but then the average user only checks email and plays flash video games on AOL still. people who are tech savy probably have a cell phone or laptop for the basic needs, and use consoles for any gaming. then theres the small handful of us left who know that your pc can decode video, or do serious gaming, and how to get the most out of it with a good solid cpu, which even up to today, still are not strong enough (basically gaming not being very good at scaling still)

[freeloader]

I can imagine in the future when process technology allows, we will have everything on one chip. CPU, GPU and main memory all on one die.

[Movieman]

I wonder if I could get 6 more MC's?
Although I think SM would take my name off the rolodex if I aked for that 2U setup!

[Manicdan]

I can imagine in the future when process technology allows, we will have everything on one chip. CPU, GPU and main memory all on one die.

the future of which market though?

what sets products and standards is really not determined by technology. if we have CPU, GPU, and RAM all in one die, for 50W and fits in a nettop, why should we stop building a machine that can consume over 1000W and play crysis 5 maxed out? the answer is we wont, since PCs size and price really have not changed for like 20 years. they all are about 20 to 50lbs, and all go for about 500 to 4000$. (depending if its mATX or full ATX, but all qualify as a desktop)

[vitchilo]

tl;dr: We don't need more than 640k of ram.[/B]

Well if they worked their butts off once in a while, no we wouldn't.

[owned666]

i never ever believe something on the 1st of April

[Manabu]

I can imagine in the future when process technology allows, we will have everything on one chip. CPU, GPU and main memory all on one die.

This already exists in mobile space (see OMAP3, for one example). And in desktop space, this integration trend is also going on for a long time. The last two examples are memory controllers and IGPs on CPU die.

[owned666]

This already exists in mobile space (see OMAP3, for one example). And in desktop space, this integration trend is also going on for a long time. The last two examples are memory controllers and IGPs on CPU die.

yeah and in terms of power
the people who made the psp's GPU said the gpu power of the ps3 will be in the psp after 2 or 4 years which i hardly believe but i never believed something as small as a stick of gum(tegra) can run a 1080p video

[freeloader]

the future of which market though?

what sets products and standards is really not determined by technology. if we have CPU, GPU, and RAM all in one die, for 50W and fits in a nettop, why should we stop building a machine that can consume over 1000W and play crysis 5 maxed out? the answer is we wont, since PCs size and price really have not changed for like 20 years. they all are about 20 to 50lbs, and all go for about 500 to 4000$. (depending if its mATX or full ATX, but all qualify as a desktop)

There's no such thing as a lasting standard in the computer industry. I can't even begin to imagine what a desktop PC is going to look like twenty years from now. About the only thing that I can't see changing is our desire for a decent monitor size. I can't see myself and many others for that fact, trying to make due with some small ten or elven inch screen on a desktop. I'm looking forward to sub 10nm process technology.

[freeloader]

This already exists in mobile space (see OMAP3, for one example). And in desktop space, this integration trend is also going on for a long time. The last two examples are memory controllers and IGPs on CPU die.

That's one nifty device.