http://www.businesswire.com/news/ho...-World’s-Fastest-Supercomputer-Reveals-Future
(...)
Knights Landing” – A Choice of Coprocessor or CPU
Intel revealed details of its second generation Intel Xeon Phi products aimed to further increase their supercomputing capabilities. Codenamed “Knights Landing,” the next generation of Intel MIC Architecture-based products will be available as a coprocessor or a host processor (CPU) and manufactured using Intel’s 14nm process technology featuring second generation 3-D tri-gate transistors.
As a PCIe card-based coprocessor, “Knights Landing” will handle offload workloads from the system’s Intel Xeon processors and provide an upgrade path for users of current generation of coprocessors, much like it does today. However, as a host processor directly installed in the motherboard socket, it will function as a CPU and enable the next leap in compute density and performance per watt, handling all the duties of the primary processor and the specialized coprocessor at the same time. When used as a CPU, “Knights Landing” will also remove programming complexities of data transfer over PCIe, common in accelerators today.
To further boost the performance for HPC workloads, Intel will significantly increase the memory bandwidth for all “Knights Landing” products by introducing integrated on-package memory. This will allow customers to take full advantage of available compute capacity without encountering memory bandwidth bottlenecks experienced today.
That machine is powered by by unreleased processors. :hmm:
I guess that's one way to do your validation testing.
Nvidia cards and AMD CPUs made on less advanced process technology still seem to have greater efficiency.
Correct, they're currently around 13% more efficient than Intel's first product targeting the market. They're also slightly less efficient in in terms of Rmax vs Rpeak. If that trend continues on actual workloads then there's no question that Kepler is the better product for this market... but unfortunately all we get to compare these systems by is a very generic 'synthetic' type of metric.
It'll be quite interesting to see what kind of improvements Knight's Landing brings. Even if it just continues to trade blows with NVIDIA's next generation... yeah, that isn't good for NVIDIA. I'd imagine it's a fair bit more difficult for them to sell K20 at over $3k each when a comparable Xeon Phi is available for under $2k. (The Tianhe-2 is listed as using Xeon Phi 31S1P, which I'd guess to be a custom variation of the 'inexpensive' 3100 series.)
Efficiency of KC does not seem that brilliant though:
http://www.anandtech.com/show/7075/june-2013-top500-list-published-xeon-phi-takes-top-spot
Titan using Nvidia cards and AMD Bulldozer CPUs made on less advanced process technology still seems to have greater efficiency and was an upgrade of an earlier system(not built from scratch too),and the same goes for the IBM powered systems using chips made using even older process technology.
This is not correct. These systems have different configurations and KS is not the only component consumes power. In fact the most power efficient system for now according to green500 is Xeon Phi based (with 2,499.44 Mflops/watt). As for R(max)/R(max) efficiency you can actually find much more efficient systems than Tianhe-2 based on Xeon Phi by navigating through top500.org
Correct, they're currently around 13% more efficient than Intel's first product targeting the market. They're also slightly less efficient in in terms of Rmax vs Rpeak. If that trend continues on actual workloads then there's no question that Kepler is the better product for this market... but unfortunately all we get to compare these systems by is a very generic 'synthetic' type of metric.
It'll be quite interesting to see what kind of improvements Knight's Landing brings. Even if it just continues to trade blows with NVIDIA's next generation... yeah, that isn't good for NVIDIA. I'd imagine it's a fair bit more difficult for them to sell K20 at over $3k each when a comparable Xeon Phi is available for under $2k. (The Tianhe-2 is listed as using Xeon Phi 31S1P, which I'd guess to be a custom variation of the 'inexpensive' 3100 series.)
The Green500 list has no K20 or K20X based systems with Intel CPUs(since there are none around ATM),and it still does not change the fact that for a large scale system Tianhe-2 is still less efficient than a older upgraded system. Tianhe-2 is a brand new installation.
Well, you do know that the top500 list is simply a run of optimized linpack, right?While certainly synthetic I think it's safe to say a lot more care is put into making sure Spec represents performance, in tasks common to installations on the super computer list, than with a desktop synthetic benchmark for desktop tasks.
Luckily Green500 is simply a derivation of the top500 and hence we can easily calculate TFLOP/s*MW. Titan is 2142 with K20x and Opteron 6274. HPCC (#53 on top500) is is 2243 with K20m and Xeon E5-2665. Todi (#108 on top500), another Cray XK7 install like Titan, is actually also 2243.4 with K20 and Optern 6272. This is compared to Xeon Phi based systems at 1901 for Tianhe-2, 1146 for Stampede, 1849 for Conte, 1932 for Discover, 1264 for Endeavor, 1685 for MVS-10P, 1613 for Maia, 904 for an unnamed one at #158, 1000 for #249, 2455 for Beacon at #397...
Basically a long way of saying that the K20 based systems are far more consistent for whatever reason while the Xeon Phi ones vary a surprising amount. There's no evidence of Titan being held back by its 'old' infrastructure or its choice of CPU - the Xeon based K20 system is exactly the same efficiency as the smaller XK7 system.
Well, you do know that the top500 list is simply a run of optimized linpack, right?
The Green500 list has no K20 or K20X based systems with Intel CPUs(since there are none around ATM),and it still does not change the fact that for a large scale system Tianhe-2 is still less efficient than a older upgraded system. Tianhe-2 is a brand new installation.
Moreover,the top system is only a relatively small scale installation(number 397 in the Top500 list),with an AMD S10000 based system coming second(52 in the Top500) and the list at the top is mostly dominated by IBM systems with CPUs made on a 45NM node,with the next Xeon Phi based systems being dozens of places down. Maybe the updated list will change that,so it will be interesting to see it.
Also,doom and gloom predictions about Nvidia and AMD being finished according to hardware enthusiasts on tech forums,is starting to get to silly levels now.
Intel has given them more competition no doubt,but its not like their competitors will all of a sudden give up. Anyway,2014 and 2015 are going to be interesting and I will leave it at that.
No, thought they'd use SpecFP. Given it's just Linpack isn't it kind of silly to even be discussing the Top 500 list in terms of component performance?
It would put all systems with accelerators (except may by xeon phi based) out of the competition. Porting SpecFP to OpenCL/Cuda is not trivial task (if possible at all).
Let see some statistics:
http://www.top500.org/statistics/list/
November 2012:
Nvidia - 50 systems
Ati - 3 systems
Intel - 7 systems
June 2013:
Nvidia - 39 systems
Ati - 3 systems
Intel - 11 systems
The trend can clearly be seen.
OK but, given the constraints, using the Top 500 list for any sort of argument about general capability seems pretty pointless.
Looks like Nvidia will have a tough time with its Tesla business.
Didint take long for Intel to take the HPC crown. And Phi will be on 14nm before Tesla will be on 20nm I guess.
Efficiency of KC does not seem that brilliant though:
http://www.anandtech.com/show/7075/june-2013-top500-list-published-xeon-phi-takes-top-spot
Titan using Nvidia cards and AMD Bulldozer CPUs made on less advanced process technology still seems to have greater efficiency and was an upgrade of an earlier system(not built from scratch too),and the same goes for the IBM powered systems using chips made using even older process technology.
KC is way easier to program for.