More Knight's Landing Details Surface!

[patrickjp93]

http://cdn3.wccftech.com/wp-content/uploads/2015/04/SZ15_SFTS003_100_ENGf.pdf

 

wccf obtained the slide deck for an upcoming presentation at IDF 2015 about Knight's Landing. Inside a whole bunch of jargon about revolutionizing parallel computing can be found, but most importantly are the specs for Intel's new 715mm^2 beast chip which can only be dwarfed in size by IBM's newest AI CPU. Mind you this is built on the 14nmFF SHP process too. Transistor counts are not yet known.

 

Knight's Landing will come in variants of core counts ranging from 60-72 in both socketed package and Xeon Phi form. All socket packages will come with 6-channel DDR4 2400MHz support capable of addressing up to 384GB each. Also, the large L3 cache of the original model has been replaced by 8 or 16GB of JEDEC-compliant 3D-stacked HBM produced by Micron on the Intel 14nmFF process (a shot across the bow to SK Hynix). There is still a 36MB shared L2 cache.

 

The HBM is supposed to offer over 1TB/s of internal bandwidth due to 8-channel configuration it's in.

 

Also be aware this is PCIe 4.0 (redouble your PCIe bandwidth) compliant as will be released with Skylake-E in 2016.

 

The theoretical performance figure many are currently quoting is 6TFlops SP and 3TFlops DP, but new to this conversation is that this is for the lowest-end version (60 cores). Clock speeds have not been given.

 

We can directly estimate the speed for the lowest model like so

Edit: I forgot there are 2 AVX 512 units per silvermont core.

6*10^12 FLOPS = 60 cores * (1024/32 = 32 SIMD units/core) * (2 FMA + 1 M = 3 op/(SIMD*cycle)) * cycles/second

clockspeed = 2.0833...*10^9 cycles/second = 1.0416666...  GHz

 

72 * 32 * 3 * 1.04666*10^9 = ~7.2TFlops SP / 3.6TFlops DP

It should be noted the previous (currently selling) generation increases in clock speed as core counts increase, so in truth these numbers are a bit conservative for the 72-core model.

 

And this is promised in a 300W TDP...

 

For reference, the nearest competition is the Nvidia K40 and K80 Tesla based on the Kepler architecture. the theoretical performance and memory bandwidth figures can be found here: http://www.nvidia.com/object/tesla-servers.html

 

As you can see, Intel has taken the DP performance crown by a fair margin. AMD's best FirePro is rated for 2.58 TFlops DP, but obviously neither it nor Nvidia get near this according to Linpack benchmarks. However, the Xeon Phi coprocessors already get much closer to their theoretical performance bounds. So, AMD and Nvidia better bring the performance in 2016 or this will be a one-sided slaughter.

 

Enjoy the supercomputing hype y'all.

[cesrai]

I'm enjoying it for some reason that I will never know.

 

So:

 

16 or 8 GB of HBM instead of L3 cache.

2304 GB of RAM - theoretical limit - 

14nm.

Two form factors.

36MB of L2 cache.

and more.

 

Bloody amazing and expensive.

[Deletive]

Hmmm, This is super cool. Wonder how it'll compare to Tesla's if you can even compare those two.

[AlexGoesHigh]

16GB HBM? the hard on is real

[patrickjp93]

Hmmm, This is super cool. Wonder how it'll compare to Tesla's if you can even compare those two.

Xeon Phi are already getting better real world performance than Kepler Teslas in a number of instances. Now that Intel additionally has the theoretical performance advantage as well, well, I don't think 2016 is going to be kind to Nvidia with or without Pascal.

[Shakaza]

I want one! I have no use for it, nor can I afford it if I want to go to college, but I want one!

[Deletive]

Xeon Phi are already getting better real world performance than Kepler Teslas in a number of instances. Now that Intel additionally has the theoretical performance advantage as well, well, I don't think 2016 is going to be kind to Nvidia with or without Pascal.

I'd say the next big competitor to what intel has to offer is volta in 2017. ~300w with HBM and Nvidia link.

[Klerk]

I thought this was gonna be Game of Thrones related...Silly me...

[MrRavens]

I thought this was gonna be Game of Thrones related...Silly me...

 

Out...OUT NOW! *shuts door*

 

Anyways, cool stuff Intel, can't wait.

[Opcode]

I'd say the next big competitor to what intel has to offer is volta in 2017. ~300w with HBM and Nvidia link.

Fiji will probably be the next big contender with 8.6 TFLOPS of single precision and 4.3 TFLOPS of double precision at 300w TDP.

[dizmo]

I thought this was gonna be Game of Thrones related...Silly me...

Not gonna lie that's immediately what I thought too!

So disappointed. Haha.

[patrickjp93]

I'd say the next big competitor to what intel has to offer is volta in 2017. ~300w with HBM and Nvidia link.

PCIe 4.0 will offer 16GT/s or up to 156Gb/s bandwidth vs. NVLINK 2.0 offering 200. If the bottleneck still ends up being PCIe then Nvidia could still do well, and then maybe Intel can work to use NVLink with its own coprocessors since PCI-SIG is so damn slow...

[patrickjp93]

Fiji will probably be the next big contender with 8.6 TFLOPS of single precision and 4.3 TFLOPS of double precision at 300w TDP.

Yes, but AMD gets nowhere near its theoretical numbers, and there's still the problem of OpenCL's ugliness/clunkyness vs. OpenMP and the lack of HSA adoption as well.

 

Also, what are the values for the AMD flops equation? edit: nvm

 

5.632*10^12 flop/second = 2816 SPs * (2flop/SP/cycle) * 1*10^9 cycles/second

[Opcode]

Yes, but AMD gets nowhere near its theoretical numbers, and there's still the problem of OpenCL's ugliness/clunkyness vs. OpenMP and the lack of HSA adoption as well.

No chip on the market will reach its theoretical compute performance including the Phi. OpenCL isn't as bad as most make it out to be. Most I would figure have no experience in using it as it doesn't take much to establishing a device driver and begin executing kernels. If you dislike it you can always migrate over to OpenACC.

 

Also, what are the values for the AMD flops equation?

 

5.632*10^12 = 2816 cores * (x simd units) * (instructions/clock) * clock speed

4096*2*1050 = 8.6 TFLOPS SP + (1/2 ratio DP) = 4.3 TFLOPS DP.

[patrickjp93]

No chip on the market will reach its theoretical compute performance including the Phi. OpenCL isn't as bad as most make it out to be. Most I would figure have no experience in using it as it doesn't take much to establishing a device driver and begin executing kernels. If you dislike it you can always migrate over to OpenACC.

4096*2*1050 = 8.6 TFLOPS SP + (1/2 ratio DP) = 4.3 TFLOPS DP.

No chip reaches its theoretical values, but some get much closer than others. On the IBM platform Nvidia actually gets closer than the earlier figures of the Xeon Phi, and Linpack has been slow to update, but Intel's Xeon Phi is performing between 75 and 95% theoretical performance now (yay branch predictors), which has started putting Intel in better and better positions to unseat Nvidia entirely from its performance crown moorings. AMD's FirePro are performing in the 35-55% range it seems.

OpenCL is okay, but it's just not as nice to set up kernels as CUDA and it's definitely nowhere near as clean as OpenMP/OpenACC. I think that's one big thing AMD shot itself in the foot on. It should have jumped on those standards instead of striving for HSA first.

And given the Titan X offers 9.9 TFlops SP, it might want to do better than that. Doesn't AMD yet have FMA + M?

[RagnarokDel]

I was expecting some game of thrones content, honestly slightly disappointed.

[FakezZ]

I'd rather have a trillion raspberry pis in a cluster /s

Wow even the lowest end one is a damn beast!

[Aircooled Brain]

Donger Raised.

[BookerDeWitt]

The Fiji XT based FirePros and the R9 390X will wipe the floor with this. Full HSA support, over 8 Teraflops of single precision compute and over 4 Teraflops of double precision compute, 16GB of HBM and undoubtedly half the price with 250 TDP.

And let's not forget all of this is going to be on TSMC's 28nm process and a 550mm² chip. The Phi is over 700mm² large on 14nm and still only manages 300W TDP.
With the same process node Fiji XT would literally be half the size and half the TDP with a 25% performance advantage.

[patrickjp93]

The Fiji XT based FirePros and the R9 390X will wipe the floor with this. Full HSA support, over 8 Teraflops of single precision compute and over 4 Teraflops of double precision compute, 16GB of HBM and undoubtedly half the price with 250 TDP.

And let's not forget all of this is going to be on TSMC's 28nm process and a 550mm² chip. The Phi is over 700mm² large on 14nm and still only manages 300W TDP.

With the same process node Fiji XT would literally be half the size and half the TDP with a 25% performance advantage.

Facepalm* I haven't seen a fanboy post this bad in a while...

No, the Fiji FirePro will not simply wipe the floor with this. If that was remotely true the current Hawaii FirePro would be wiping the floor with both Nvidia and Intel in the HPC space right now based on the theoretical numbers. However, AMD barely has any supercomputer sales, and the current Hawaii FirePro already support OpenCL 2.0 and most of HSA.

HSA is not seeing much adoption or even exploration. That's because most of the HPC code base is already in OpenMP and OpenACC which do 95% of what AMD proposed in HSA, and they're easy programming models to learn.

Intel has one huge advantage over GPUs when it comes to parallel compute: fully asynchronous computation. Every thread can be individually handled and programmed, on a much bigger scale than AMD's 8 ACEs. Furthermore, the Xeon Phi have branch predictors, allowing you to put both the computations and decision tree on the Xeon Phi and not have to ping-pong kernel I/O, leaving th Phi saturated for far greater percentages of time. This is why Intel's real world performance is so much closer to its theoretical and then beats its competition despite having lower theoretical performance numbers.

As per die sizes, you forget AMD is already using its HDL move and Intel isn't. At 14nmFF with no HDL Fiji would easily be as big as 700mm^2. Intel is also pulling double the clocks of AMD's GPUs. And both chips will have a 300W TDP.

And actually Intel's Xeon Phi are very price-competitive against Tesla and FirePro products.

Only in truly embarrassingly parallel workloads will Fiji eke out a small performance win over Knight's Landing. The problem with using this as your measure is HPC has many different parallel workloads, and you can't eliminate all branching for all workloads for dynamic sizes no matter how hard you try. As Linpack has published, Intel wins in 18 of 22 major tests on the current Knight's Corner generation vs the K40 Tesla the the W9150 FirePro, some by small margins and others by large ones.

[Misanthrope]

Wouldn't this chip technically work as a GPU with so many smaller cores?

[patrickjp93]

Wouldn't this chip technically work as a GPU with so many smaller cores?

It's been confirmed DX12 and Vulkan can work with these, but Xeon Phi lack the nice TMUs and ROPs that GPUs depend on, so they won't do so well, especially for the price. It is a highly parallel computation unit though similar to GPUs in that respect.