AMD’s Boltzmann Initiative could make HPC development easier

AMD wants to make software development easier for high-performance computing (HPC) clusters, like those based on its FirePro S9170. At SC15 this week, AMD is showing off a set of tools called the Boltzmann Initiative that it says will help developers move CUDA-based applications to C++, and make HPC development more accessible to a wider audience. 

The most important part of the Boltzmann Initiative seems to be a Heterogeneous Compute Compiler (HCC) for C++. The company says the HCC will make HPC applications easier to write because of that language's popularity. The HCC can purportedly dig through an application's code to determine the tasks that are best-suited to the high serial performance of the CPU and those which are best-suited to the much more parallel GPU. 

To go along with the compiler, AMD says its Heterogenous-compute Interface for Portability (HIP) tool will help developers move their code from CUDA to C++ in the first place. The company says that "in many cases," 90% or more of a given program's CUDA code can be automatically converted by the HIP tool, leaving developers with a relatively small amount of manual effort to port what remains. To show off the combined potential of the HCC and HIP tool, AMD is demonstrating the CUDA-based Rodinia benchmark suite running on the company's own GCN-based hardware at SC15.

The final tool that AMD revealed today is an HPC-focused headless driver and runtime for Linux. The company says that this release will focus on low-latency dispatch, increased PCIe transfer speeds, and create a single unified memory pool out of the memory on each card. The driver will also support peer-to-peer GPU communication and Remote DMA from InfiniBand. All of the tools AMD announced are expected to become available early next year.

Despite the new tools and move to C++, AMD says it remains committed to OpenCL, too. The company has announced expanded GPU-compute libraries for that programming language. AMD says these new libraries are aimed at scientific computing tasks. 

Ben Funk

Sega nerd and guitar lover

Comments closed
    • kuttan
    • 4 years ago

    With this new approach application developers can target both AMD and Nvidia hardware much easier than designing CUDA and OpenCL separately.

    • Zizy
    • 4 years ago

    Just 90%? This sounds very bad. Especially as even those 90% are unlikely to be properly optimized to get max performance. Phi has 100% source compatibility to your existing x86 code and it is still a huge pain in the ass, to the point you might as well use CUDA.

      • NoOne ButMe
      • 4 years ago

      I think the extra 10% versus trying to make x86 core massively parallel is probably easier.

        • Zizy
        • 4 years ago

        Sure, but when your competitor has 100% CUDA compatibility and offers similar performance for not much more (unsure how much either company is willing to drop prices for a typical cluster though), why bother in the first place?
        This x86 part was just that compatibility doesn’t mean much in the end, code needs to run fast or forget about it.

        So, here we have 2 paths:
        One path = new NV cards, your existing CUDA code and people that know how to write it.
        The other = new AMD cards, this porting tool and effort to make it work, hopefully similar code so the people can easily retrain.

        Which do you believe most will take? Unless there is a lot of salary budget but limited hardware budget (yes, sometimes budget has such stupid splits), the first one.

          • NoOne ButMe
          • 4 years ago

          But, AMDs cards are typically faster for the cost, and for heavy FP64 cases better in pref per watt. Assuming this program hits expectations.

          I think the best course for companies would be to buy one AMD card to see if their port is easy and if it runs faster. Generally speaking a small loss if Nvidia is better, and likely an overall gain in performance or lower cost of AMD is better.

          Also CUDA is a lock on while AMD’s C++ should offer an “escape” should they wish to choose vendors. For some that could decide it. But I prefer companies test to figure out what is best for them and go with that. Nvidia will win the largest share probably, AMD will win some, Intel will win some. Just Nvidia would likely lose overall market share.

    • ronch
    • 4 years ago

    I wonder how relevant this will be, ultimately. Most of AMD’s products and technologies pretty much just fall over the wayside these days. That goes for their APUs, their embedded chips, Firepro (I wonder how many folks actually get Fireproof over Nvidia’s Quadro), Mantle, TrueAudio, misc. software tools, etc. Really, apart from Zen, some FX models, and Radeon, people are just mostly ignoring them. Even Carrizo doesn’t seem to be getting much love these days despite efficiency gains.

    • DoomGuy64
    • 4 years ago

    Hmm, ports CUDA to C++. Wonder if someone could use this to port PhysX. That would be interesting…

      • TopHatKiller
      • 4 years ago

      No it ain’t porting. the idea, unless i’m wrong, is easier translation between cuda and opencl. for hpc computations; valuable for that market but this has, as far as i understand, no value for consumer graphics. In any case the horrid ‘physX’ thing is pretty much dead. [least i hope so]
      However we might laugh at the idea of people buying millions of dollars of hardware just to play a nv-sponsored Batman game or some such – they probably don’t. Sadly.

        • mctylr
        • 4 years ago

        It sounds like a source-level conversion from CUDA to OpenCL, i.e. source based “porting.”

        Agreed it would most likely not be applicable to PhyX, and is strictly a high-end market (HPC) oriented effort.

    • guardianl
    • 4 years ago

    Nvidia’s Tesla business revenue has been down Y/Y two quarters in a row. The revenue comes in fits and starts, but in general, Tesla+CUDA (2006) has been around for almost 10 (!) years now and while it’s been a nice side thing for Nvidia there is no huge untapped market for GPU acceleration of general programs.

    Nvidia gets about $400 million/year in revenue from Tesla etc. There’s no huge secret market in compute, and what’s there is going to get eaten almost completely by Intel’s Knight’s Landing because software (compatibility, ease of development etc.) is king.

    Another mis-adventure by AMD management.

      • mctylr
      • 4 years ago

      In the HPC market, software performance is the end goal. The theoretical performance (of hardware) is of marginal interest, used mostly by the hardware vendor as a marketing / promotional aid, compared to the actually achieved performance. The gap can be 10-20% in may cases, depending on how radical any platform changes are.

      For every generation of hardware, HPC centres spend about a year migrating their existing operations to the new platform, spending a large amount of time benchmarking and tuning the performance, as well as ensuring reliability. Tools that make this easier are always welcome, but there is still a planned migrating phase where software optimization is manually performed.

      While Intel’s MIC efforts, e.g. Knight’s Landing is a solid contender, GPGPU does key compute (double precision floating point matrix calculations) well (massively parallel) enough to not be a foregone conclusion.

      The military and intelligence customers (NSA, GCHQ, etc.) spending is not well documented (don’t appear on TOP500) portion of the HPC segment by design. I don’t think anyone knows how large that portion actually is.

      I do agree with the idea that it isn’t likely enough to be AMD’s saviour, unless they become like Cray and SGI whom attempted and failed. to reinvent themselves as a smaller, supercomputer focused, company.

        • guardianl
        • 4 years ago

        I find your comments insightful, but I will nitpick a little bit anyways.

        >> In the HPC market, software performance is the end goal.

        In every software market the end goal is to solve problems. HPC is more of a marketing term than a realized concept. You can do effective compute-work by renting time on EC2 (including GPGPU of course), but a lot of universities and national labs will avoid it because it’s not a good way to maximize their grant/budget proposals.

        i.e. I can propose to spend $100 million to build a new cluster on campus, or hope I can get $20 million every year for the next 5 years for renting time on EC2. In the first case, I only need approval once and it’s a sure thing for years of research. It also has a tangible quality which tends to help with the approval.

        >> The military and intelligence customers (NSA, GCHQ, etc.) spending is not well documented (don’t appear on TOP500) portion of the HPC segment by design. I don’t think anyone knows how large that portion actually is.

        I’ve seen the inside of that rats nest, and without saying anything specific, military procurement is more about politics and business relationships even in SIGINT, HPC etc. The “Pentagon Wars” is hilariously relevant for software/compute hardware at DoD. Whoever succeeds here will probably be selling with the blessing of Boeing/Raytheon/BAE etc. If Nvidia succeeds over Intel or vice-versa, the difference won’t be due to benchmark results.

      • NoOne ButMe
      • 4 years ago

      Revenue falls because market is saturated on 28nm? Large growth once 16nm hits?

    • chuckula
    • 4 years ago

    [quote<]The most important part of the Boltzmann Initiative seems to be a Heterogeneous Compute Compiler (HCC) for C++. The company says the HCC will make HPC applications easier to write because of that language's popularity. The HCC can purportedly dig through an application's code to determine the tasks that are best-suited to the high serial performance of the CPU and those which are best-suited to the much more parallel GPU.[/quote<] Well anything that makes programming for Knight's Landing easier is a good thing, but why is AMD doing this?

      • maxxcool
      • 4 years ago

      Because AMD cannot survive on ZEN sales or GPU sales alone. Without a strong presence in the HPC arena they will fold.

        • mesyn191
        • 4 years ago

        Fold might be a little strong. HPC is definitely a high profit market segment that they should be targeting so this does make sense for their bottom line if they want to sell Firestream GPU’s.

          • maxxcool
          • 4 years ago

          At this juncture nobody WANTS to buy a AMD cpu. ATI is their meat and potatoes. ZEN when it pans out will be ‘good enough’ more than likely but that does not make instant marketing, and does not change peoples minds even with glorious Team Red slides on every website on earth. So they literally NEED to survive to zen 2.0 to show that management can do it, and that engineering … the pool of folks they have fired from multiple times will still be there.

          Sure the fans will but up ZENS 2 or three at a time and proclaim their ‘told you so’s” .. but fans do not bump the needle so to speak. ZEN needs a volume purchaser and xbox and ps5 with its RAZOR thin margins is not their friend.

          Unless ZEN is spectacular.. or unless dell starts selling droves of zen enabled laptops AMD will slow their bleeding but not solve it.

          thus imo, internet opinion at that 😉 ZEN + ATI is not enough. they NEED a massive NV style deal for network learning or AWS style purchase or some fun Database widget that shows amd having a clear 20% performance advantage.. maybe with CRAY ? I dunno. But they NEED to every win possible contract that pays them upwards of 11% – 25% profit on every unit sold.

          OEMs are afraid of AMD. their long term bleeding leaves fear … And fear is the mind killer. Even with stiff discounts oems will have a hard time committing large piles of money until they can see the product their investing in will have a future.

          honestly. it makes me sad, because if ZEN IS really good it still won’t turn things around initially. maybe even for a year plus which puts us back in in the dreaded 2nd quarter earning report.

        • TopHatKiller
        • 4 years ago

        wow. you know how Zen is gonna do? Wish I did.

          • maxxcool
          • 4 years ago

          It’s an easy prediction. Switching to SMT, going full fpu per core with AMD’s traditional fpu and strong integer will put them thuslly::

          imo..

          ZEN will trade blows in the INT realm. Some apps will be stronger than Ivy bridge, some will be close enough that it won’t matter. unless there is a issue with the scheduler.

          ZEN will do ‘well’ in FPU but will not beat Ivybridge most of the time. They will do extremely well and hit 90% in most benches compared to IVY.

          ZEN will do EXTREMELY well in gaming compared to excavator. ZEN may pull a few games out of the hat to beat Intel on but the majority will be within striking distance of IVY bridge (+ or – 5%) unless there is a issue with the scheduler.

          imo, ZEN will be good enough to spend money on… as long as its 25% cheaper than the comparative part.

          again… intenretz opinion and all that.

          Goodday sir! /tips hat/

            • maxxcool
            • 4 years ago

            The REAL gamble\question. Will we see ZEN 2.0 before the buyout?

            • TopHatKiller
            • 4 years ago

            mhm.erf. tapping fingers. if you’ve got that spot on I’ll Eat My Hat. we’ll see.
            ps. thanks for your sweat politeness and, i cheat, my hat is imaginary.

        • NoOne ButMe
        • 4 years ago

        Um… How much of Nvidia’s pro GPU revenue and hence profits comes from non-HPC? The majority.
        How much of Intel’s professional COU revenue comes from HPC? Not much.

        HPC won’t make a difference to AMD is the other segments cannot pull it back. And Zen’s design makes is bad for HPC to start. Because HPC doesn’t matter for Zen.

          • maxxcool
          • 4 years ago

          When your drowning every breath helps.

      • anotherengineer
      • 4 years ago

      “To go along with the compiler, AMD says its Heterogenous-compute Interface for Portability (HIP) tool will help developers move their code from CUDA to C++ in the first place. The company says that “in many cases,” 90% or more of a given program’s CUDA code can be automatically converted by the HIP tool, leaving developers with a relatively small amount of manual effort to port what remains. To show off the combined potential of the HCC and HIP tool, AMD is demonstrating the CUDA-based Rodinia benchmark suite running on the company’s own GCN-based hardware at SC15.”

      Probably to try to push sales of their compute cards with higher margins to try to make more money.

      Just a guess but I could be wrong 😉

      • zorg
      • 4 years ago

      This won’t make the programing for Knights Landing easier. Sure there is a possibility to write a compiler that compiles the HIP source for Xeon Phi, but this product family has hardware design problems. The way that SIMD is implemented on x86 is really inefficient, because the programers don’t want to deal with multithreading, prefetching, and small SIMD registers. The concept how Intel support this is also really bad. They constantly add new instructions with new SIMD register sizes, which will make your code outdated. The only thing that could help the Knights family is a new data-parallel ISA, without any x86 compatibility.

    • blastdoor
    • 4 years ago

    One thing that would be super helpful from AMD would be to design, manufacture, and sell high performing, competitively priced CPU and GPU products. Maybe they could look into trying that.

      • xeridea
      • 4 years ago

      GPUs are competitively priced on various performance levels. Compute wise, they are a better deal because they have a lot better double precision performance than Nvidia cards (other than Titan, which is overpriced). This is where their extra power usage comes from.

      For CPU, Zen is about a yearish away, which should get them near Intel. Bulldozer has its flaws, though the architecture has improved greatly, and it still does fine for what it was meant for, multiple threads. I think the biggest issue is they were never able to go to 20/22nm due to manufacturing issues (outside of their control), so Steamroller and Excavator products are limited.

        • blastdoor
        • 4 years ago

        I’m not so sure that manufacturing process is their biggest issue. Do they have any 32 nm CPUs that are competitive with Sandy Bridge (which was 32 nm)?

        Good point regarding double precision FP for GPUs.

        edit:

        here’s a quick look at two 32nm CPUs from Intel and AMD:

        [url<]http://www.anandtech.com/bench/product/1289?vs=994[/url<] Even ignoring the power difference and looking only at performance, my read is that manufacturing process isn't AMD's only (or even primary) problem.

          • xeridea
          • 4 years ago

          Process shrink wouldn’t be a magic fix, but it would have helped. The main thing would be allowing Steamroller/Excavator on FX chips instead of only APUs and laptops (in the case of Excavator). So their whole line up would benefit from IPC and multithreading improvements with uArch revisions. This would lessen the performance delta and help them sell more chips. It would be a good benefit to APUs since GPU cores benefit more from shinks than CPUs.

      • NTMBK
      • 4 years ago

      AMD’s graphics hardware is fine in raw compute power. It’s the software that has lagged.

        • guardianl
        • 4 years ago

        The raw ALU performance is there, but there’s a lot more to graphics performance than shader ALUs. Good drivers would go a long way, but software isn’t what’s causing (most of) the huge perf/watt disadvantage compared to Maxwell.

      • Ninjitsu
      • 4 years ago

      Looking at the CUDA story, the impression I’m getting is that for HPC, software is key.

        • Leader952
        • 4 years ago

        Remember the quote “Nvidia is a software company” that got ridiculed here.

        Well Nvidia was right in that software drives hardware sales.

        [url<]http://www.nvidia.com/object/gpu-applications.html?All[/url<] Here is a just announced example: Nvidia Tesla GPUs To Accelerate VASP Material Simulation Application [url<]http://www.tomshardware.com/news/tesla-k80-supercomputing-vasp-acceleration,30576.html[/url<]

      • Convert
      • 4 years ago

      Hahaha! I don’t know why you are getting down voted but I gave you three up votes.

        • NoOne ButMe
        • 4 years ago

        Because he includes GPUs which are typically competitive and almost always priced competitively.

    • PrincipalSkinner
    • 4 years ago

    Yes but where is 380x?

      • NTMBK
      • 4 years ago

      It’s a supercomputing conference…

        • ronch
        • 4 years ago

        Boom.

      • Ninjitsu
      • 4 years ago

      IN MY POCKET

Pin It on Pinterest

Share This