AMD Collaborates with US DOE to Deliver the Frontier Supercomputer

[R-T-B]

It doesn't matter if it's open or closed.

And if the driver is a closed binary unavailable for your platform?

Least resistance. It is arguably easier to port a driver than pay for a new closed one to integrate with a moving target (OSS kernel).

I do advocate opensource, but only when it's actaully helpful. Their build design choices suggest it must be, or they'd probably have went with a CUDA based system for the versatility of the end users running code.

 

[notb]

And if the driver is a closed binary unavailable for your platform?

Least resistance. It is arguably easier to port a driver than pay for a new closed one to integrate with a moving target (OSS kernel).

I'm not sure why porting a driver would be easier than paying for a new closed one. Porting takes time and costs. And you'll likely outsource it either way.

Also, we're talking about an HPC system. Cray delivers the whole package: configured and ready to run.

 

[notb]

Because you'll need to pay for a port update everytime the open source kernel you are integrating with updates. Open source integrates easier with open source and is inherently cheaper to maintain.

You buy a server with service. Cray (or any other OEM) provides support - including drivers.
It's way more cost effective as well.

 

[medi01]

Perhaps that's why Threadripper 2 has disappeared.

 

[Aquinus]

Well, the reality is that Nvidia cluster can run CUDA and this can't.

The reality is that nVidia owns the copyrights for CUDA like all of their other IP and everyone else doesn't. If it's not nVidia, it's not going to do CUDA. If your software uses CUDA, that's called vender lock-in.

 

[Frick]

I live in a time when there are three exaflop supercomputers in the works. Crazy.

 

[lexluthermiester]

supercomputers-exceeding a quintillion, or 1018, calculations per second

?!? A quintilion is much higher than a mere 1018.

 

[AleksandarK]

?!? A quintilion is much higher than a mere 1018.

Thanks for pointing that out. Should have been 10^18th

 

[Patriot]

AMD made ROCm and HIP for a reason, HIP can convert cuda code... and often it is faster on nvidia cards after the conversion than before...
ROCm has been lagging behind 2-3 versions in support but they have been catching up quite well to CUDA's feature set.

If you want to do workloads that rely heavily on tensor ops, then you go with V100s, if you need simple double, single or half precision than AMD is a solid option.
If you need just need inferencing... shiiiit options are wide open.

Edit: Shit, they have caught up on version support... https://rocm.github.io/dl.html

 

[R-T-B]

You buy a server with service. Cray (or any other OEM) provides support - including drivers.
It's way more cost effective as well.

And Cray has an interest in reducing the price of the service they provide?

As I said, I can't be certain. But I think their parts choices point down that road...

You're naive if you think you have an argument.

And your naive if you think calling me naive constitutes an argument.

AMD sold the idea to DoE in 2011. Nvidia & Intel followed.

And the wright brothers invented the airplane. Who cares who follows who in relation to marketshare? Furthermore, call me when their system is even in the TOP100 charts.

 

[Deleted member 158293]

Starting with open solutions is the quickest way to program.

AMD made ROCm and HIP for a reason, HIP can convert cuda code... and often it is faster on nvidia cards after the conversion than before...
ROCm has been lagging behind 2-3 versions in support but they have been catching up quite well to CUDA's feature set.

If you want to do workloads that rely heavily on tensor ops, then you go with V100s, if you need simple double, single or half precision than AMD is a solid option.
If you need just need inferencing... shiiiit options are wide open.

Edit: Shit, they have caught up on version support... https://rocm.github.io/dl.html

Interesting to see the need for, and always good to have an exit ramp from vendor lock in.

 

[prtskg]

It's written in the article that ROCm software stack will be used. It's open source and with version 2.4, it's quite competitive with CUDA feature wise. Though I wouldn't be surprised if the next supercomputer win belongs to IBM+Nvidia combo. Let's wait and watch.

 

[Patriot]

Furthermore, call me when their system is even in the TOP100 charts.

Titan is still #9 and AMD based... There is a Zen based chinese small node # supercomputer that is #38
This will be #1 and 7x faster than the current #1 and 50% faster than Intel's system that may or may not get finished first.
I think this is the first time I have seen AMD cpu+gpu be in the top 10, everything before has been AMD CPU/Nvidia GPU.

 

[HD64G]

Interesting you accuse me of not reading, because I already addressed this.

Frontier is based on Zen, which is a complete ground up redesign since Bulldozer based fusion. They are about as related as an Apple and a Potato, or Netburst and Sandy Bridge.


Pretending you are smarter than everyone is not making you look great here.

or maybe I am confused... what is your argument here, exactly?

If you are just doing a generic "AMD IS DAH BEST," you at least aren't exclusively wrong. Consider the following:

Zen is better suited for this than Intel for certain. It's their GPU-choice I find intriguing, and only because they could've likely been more power efficient (important in super computer clusters) with nvidias line.

For servers the GPUs aren't rated by their gaming performance but in TFlops=raw compute power. And in that, AMD has efficient GPUs as Vega arch is mainly a compute targeted one for data centers. And in 7nm AMD GPUs are much more efficient. Radeon 7 is ~30% faster than Vega 64 using 10% less power.

 

[Caring1]

And the wright brothers invented the airplane.

And here I was thinking they were the first to actually do a documented flight, and they invented planes too....

 

[Prima.Vera]

31MW of power. Talking about "global warming"?

We are in 2019. 4 wind turbines like those can provide more than adequate power. Also they can install additional solar panels on the roof and can have 100% ECO energy.

 

[mtcn77]

We are in 2019. 4 wind turbines like those can provide more than adequate power. Also they can install additional solar panels on the roof and can have 100% ECO energy.

31MWh is still 7.4 megacalories per second, or 7.4°C/s per ton of water...

 

[SoNic67]

We are in 2019. 4 wind turbines like those can provide more than adequate power. Also they can install additional solar panels on the roof and can have 100% ECO energy

And DOE can make calculations only when wind blows. Good investment.
Seriously, every single one of those wind turbines has to be paired up with a quick-reacting gas turbine, ready to pick up the load. Spinning reserve. Those are expensive to buy, to maintain and they are not fuel efficient compared to a slow reacting coal or nuclear plants (those cannot be used as spinning reserves because cannot be turned on-off so fast).

So by using the "free" wind, you just increased the price of electricity...

Did you miss the part where I worked here just a year ago?

What happen, did they finally fired you?











Just joking, sorry, could not help, was so easy...

 

[bogmali]

This thread is not for chest beating purposes, you can continue those offline or via PMs. Thread cleansed and reply bans issued.

 

[Prima.Vera]

Seriously, every single one of those wind turbines has to be paired up with a quick-reacting gas turbine, ready to pick up the load. Spinning reserve.

No need for those. The Datacenter will be connected to the main grid anyway ready to pick up the missing load in case no wind/sun from the Mother Nature.

But I am also curious about the future upgradebilitty for those server farms. Can the CPUs/GPUs be easily upgraded in the future on the fly? And I mean without changing motherboards and such?

 

[SoNic67]

The Datacenter will be connected to the main grid anyway ready to pick up the missing load in case no wind/sun from the Mother Nature.

You don't get it. That fabled "main grid" ready to jump to help... it's composed of many individual steam or gas turbines like I said above.
In every microsecond of the day, 24/7, 365 days/year, the amount of electricity produced has to be equal to the electricity consumed. If a power generator drops quickly, somewhere in the system, preferably close by, another one has to pick up as quickly that exact deficit of power.
Steam turbines (coal fired) have a spare capacity of a few minutes of steam, and by that time, the gas turbines have to start-up already to pick up the slack.
You can't mess up with a nuclear power plant up and down that way. Or... you can, but might end up with Chernobyl.

 

[Caring1]

You don't get it. That fabled "main grid" ready to jump to help... it's composed of many individual steam or gas turbines like I said above.
In every microsecond of the day, 24/7, 365 days/year, the amount of electricity produced has to be equal to the electricity consumed.

WUT!
Wrong and so off topic, I'm leaving it there.

 

[HwGeek]

After Watching this about Milan and the possibility that it will include 15 Chiplets and maybe SMT4, I think [my Imagination] I Have an Idea where AMD is going with it's future(Maybe Custom design?) HPC EPYC design on 7nm+:
1)Each CPU chiplet will be 6C/24T to save space/power while giving similar or better then 8c/16t performance.
2)Adding 4 custom Instinct GPU chiplets.
3)Adding 2 custom AI accelerator [Asics] chiplets.
4)1 I/O chiplet with HBM memory stack.

So the final EPYC Milan(?) can be HPC beast with:

• 48C/192T Zen CPU cores.
• 4 custom Instinct GPUs.
• 2 AI accelerator Asics.
• 1 I/O Chiplet with HBM 3D staking .

EDIT: I see that there was already great article on such HPC APU design:
https://www.overclock.net/forum/225-...lops-200w.html

https://www.computermachines.org/joe/publications/pdfs/hpca2017_exascale_apu.pdf

You can see the EPYC PCB design in "Figure 2. Exascale Heterogeneous Processor (EHP) ".

So after reading some of it I changed my illustration:
No CPU Chiplet ontop of I/O:[Took Vega Pro 20CU image and placed the HBM on top and shrank it to 7nm+ level]
IMO the GPU's could take ~150W + 75W~100W rest of the CPU+I/O= around 250W TDP.

Or CPU Chiplets ontop of I/O- it can still be 14nm or 7nm- but it gonna stay large chiplet anyway to place cpu chiplets on top,

And 8 Milans could be installed in Cray’s Shasta 1U with Direct Liquid Cooling:

https://www.anandtech.com/show/13616...liquid-cooling