• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

IBM z10 Processor

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

Shiggity

Member
Joined
Dec 16, 2007
Location
Chicago, IL
Fastest processor in the world.

The highest-end z10 processors use five quad-core die packages and two service cores; that's 20 cores at 4.4 GHz, 60 MB of L2 cache and 48 MB of shared L3 cache on a single processor!!!!!!!!!!!!!!!!!!!!!

They are used in z10 based servers, price starts at 1,000,000$. Get your checkbooks out.

IBM is going to run the interwebz on one server for realz.

http://www.dailytech.com/article.aspx?newsid=10882

http://news.yahoo.com/s/ap/20080226/ap_on_hi_te/ibm_new_mainframe
 
Hmmm, I wonder how much folding ppd one of those would do...

Not what you might assume. IBM's mainframe CPUs are not optimized for floating point performance, but for integer performance. Prior to the Z series their mainframe CPUs didn't even have floating point units. All floating point math was performed via software routines, and it was painfully slow: less than 1/1000th the FP performance of an X86 chip at the same clock speed.

The Z10 core has a single un-pipelined 64 bit floating point unit capable of 1 Flop per cycle. At 4.4 GHz this equates to 4.4 GFlops theoretical peak. Not too shabby, but it's all due to clock speed. One core of a current 3 GHz Intel CPU will peak at 6 GFlops peak due to SSE registers being 128 bits wide and heavily pipelined.

If you really want to pump out some folding results, you need this machine:
http://sicortex.com/products/high_capability_system_sc5832

It will do just over 8,000 GFlops peak using 5832 64 bit MIPS cores. Each CPU is a system-on-chip containing 6 CPU cores, a dual channel DDR2 memory controller, PCIe controller, and Gig-E controller. Each chip is a node in the system, and each node runs a copy of Linux. So, in essence, you'd have 972 six-way servers each running a folding client running 6 threads. Not really what the machine is designed for, but it could easily be done as it runs Linux. Although, I'm not sure if there are MIPS binaries available. If the folding client is open source, you could compile it yourself to run on this system.

Oh, what's that? You don't have $6 million for one of these? Then buy the smallest sibling in the family, the deskside SC072 for $23,695.00. It has 9 of the chips, for a total of 100 GFlops, and it even plugs into a standard wall outlet and runs very quietly. Seems pricey? It's still $10K cheaper than what an IBM System X Xeon server of 100 GFlops would cost you.
 
Hardwarefreak,
well using NV's cuda with a Telsa card is cheaper route to go.
Tesla C1060 Computing ($1.7k,price from Tigerdirect)
Single Precision floating point performance (peak) 933
Double Precision floating point performance (peak) 78
http://www.nvidia.com/object/product_tesla_c1060_us.html
or get this
http://www.nvidia.com/object/product_tesla_s1070_us.html
or even this one
http://www.nvidia.com/object/personal_supercomputing.html
for $10k would be a better cost/performance ratio.


CGR,
im sure that z10 would be better at seti or other projects that are integer based. would be :drool:
 
Hmm, smack in 2 4870x2s and bam, the perfect WoW rig :beer:
 
Not x86 architecture = fail :(


you cant put windows on it, or run your favorite applications :weep:
 
Crysis should be hard... You could probably install windows inside a linux-VM, but then again you'd have to emulate a gfx-card (since mainframes don't have any and 'lack' the slots to add one in).
However, once you got it installed, it conveniently fits into the z10's RAM (which is mostly a TB btw).

However these machines were never, are not and will never be designed or meant to play games.
They're servers.
For servers. Yes, servers for servers. Once IBM started a script to start new linux-VMs on a z990, the system stopped popping up new VMs when the count reached 41,000. Of course that's not a setup you would encounter, but these machines barely run with only one OS.

And they don't shut down. They don't reboot. They don't fail. Ever. (Actually, the last 40 years not one mainframe failed)
Because that would be costy ;-)


You might be glad to know IBM has plans for the Cell-BE to be used with the IO-Monster (it is, z10 has 280 GByte/s of IO) mainframe, that will be interesting =)
 
Back