AMD Radeon Open Compute Platform hits version 1.3

A year ago at the intuitively-named SC15 supercomputing conference, AMD announced its intention to improve its software stack for high-performance computing. The name of that program was the Boltzmann Initiative, and the project's first software fruits came out in April, collectively known as the Radeon Open Compute Platform, or "ROCm." Now, at SC16, AMD is announcing the release of version 1.3 of the platform.

While the red team has been clawing back some gaming market share from the green giants, Nvidia still owns the high-performance compute market. Put simply, ROCm is AMD's answer to Nvidia's CUDA, and aims to assist developers in coding compute-oriented software for Radeon GPUs, along with converting existing CUDA software to run on GCN hardware. At SC16, AMD showed a demo of the CUDA-based Caffe deep-learning framework running on its Radeon GPUs. The company says it was able to translate 99.6% of the code using its HIP tool, and as a result it took less than four days to complete the port. That kind of porting turn-around time could make adopting Radeons a real option for companies that currently rely on CUDA.

The new ROCm version has a lot more than the updated HIP tool, though. Previous versions of ROCm only supported the Hawaii and Fiji GPUs, meaning that ROCm deployments in the field were power-thirsty by nature. ROCm 1.3 now supports the Polaris GPU family, bringing the minimum spec down to the Radeon RX 460. The new version also packs an update to the LLVM-based Heterogeneous Compute Compiler. ROCm software can now be used in virtual environments, too, thanks to new support for Linux KVM GPU pass-through.

Besides announcing the new release, AMD is talking about the future of ROCm at SC16. Software built on the framework currently requires an Intel Haswell or newer CPU, but AMD says future releases will support ARM's AArch64, IBM's Power8, and of course, its own Zen processors. The company also says it plans to rebuild its OpenCL support on top of the ROCm platform to give OpenCL applications "direct-to-metal" access, which might improve their efficiency.

Comments closed
    • ronch
    • 3 years ago

    That first marketing slide.

    AMD marketers.

    Sigh.

    • Waco
    • 3 years ago

    Anyone from TR at SC this year? I’d love to chat HPC with fellow gerbils or gerbil-masters.

    Also, AMD seems pretty tight lipped thus far – I’ll see if there’s any more details to eek out on this and Zen.

      • chuckula
      • 3 years ago

      The technical sessions at the conference apparently really get going later this week.
      If AMD is going to talk up Zen (and they should) then we will know more relatively soon.

      TR hasn’t reported on it yet but the new Top 500 list came out ahead of this conference too: [url<]https://www.top500.org/lists/2016/11/[/url<] Interesting notes are that there are now two Knights Landing systems in the top 10 at positions 5 and 6. Nvidia also has a P100 system at number 8. I'm actually more interested in the report on the newer HPCG benchmark, which is intentionally designed to stress I/O and other components beyond LINPACK's embarrassingly parallel number-crunching process to test how well systems actually scale for complex tasks. [url<]http://www.hpcg-benchmark.org/[/url<]

        • Waco
        • 3 years ago

        Trinity should come in at #2 should we decide to run Linpack on it. I honestly hope we don’t. HPCG utilitizes far more of any given machine, much more like a real app.

        I’ll dig if there’s anything good from AMD, but I didn’t see much on the schedule regarding their x86 parts. It’s possible I’m terrible at reading schedules, I’ll check again and/or stop by their booth.

          • chuckula
          • 3 years ago

          #JellyThatYouGetToGo

            • Waco
            • 3 years ago

            It’s part of my job, between presentations, meetings, and networking. I miss my systems though…

            • chuckula
            • 3 years ago

            Just setup your crontab for them to email you every two hours. It’ll be like you never left!

            As TR’s unofficial/official man on the ground, tell us if you see anything cool.

            • Waco
            • 3 years ago

            I think I might get in trouble if I bridged my networks onto anything Internet. Lol

            I’ll keep my eye out for cool stuff.

            • BluePanda
            • 3 years ago

            Oh but the wife is here too…so now we’ll have compete for who can find the coolest stuff.

            • Waco
            • 3 years ago

            Some cool stuff on the floor, but the real interesting stuff is all NDA.

            Seagate has HAMR drives on the floor with a live demo though!

      • Srsly_Bro
      • 3 years ago

      A Naples rumor originated here.

      [url<]http://dresdenboy.blogspot.co.il/2016/11/leaked-zennaples-benchmarks-appeared-in.html[/url<]

        • Waco
        • 3 years ago

        Nothing from AMD this week. The CPU team isn’t even here, just the Radeon crew…

    • DragonDaddyBear
    • 3 years ago

    I’m not a programmer. I struggle to understand the significance of this. It SOUNDS cool, though. It’s not a TR specialty, but I would love to see an article that compares and contrasts the Green vs Red vs “other” type platforms in this space. Maybe TR could work with a gerbil who knows a thing or two in this area?

      • brucethemoose
      • 3 years ago

      Optimization compared to OpenCL?

      I’ve worked with an interesting example: here’s the same image filter ported to 2 different platforms:

      OpenCL: [url<]https://github.com/HomeOfVapourSynthEvolution/VapourSynth-Waifu2x-w2xc[/url<] Caffe (CUDA): [url<]https://github.com/lltcggie/waifu2x-caffe[/url<] On a GTX 980m, the caffe version was over 2x as fast. Another interesting note: Fiji is a [i<]monster[/i<] in w2xc, churning out upscaled frames in 0.5s vs almost 2s for the 980m, if I'm remembering it correctly. My lowly 7950 was also outpacing it. Not sure how meaningful that is, it could just be Nvidia's lack of focus on optimizing their OpenCL drivers. So, if Radeons could run that ported CUDA code with this project, they could get a nice performance boost AND save some programmer 90% of the trouble of porting whole CUDA programs like that to OpenCL.

        • Goty
        • 3 years ago

        [quote<] it could just be Nvidia's lack of focus on optimizing their OpenCL drivers. [/quote<] I'm almost certain that's the reason for the apparent discrepancy.

        • xeridea
        • 3 years ago

        Could be architecture differences.

        Could be that Nvidia doesn’t care about OpenCL, so drivers probably suboptimal.

        Could be that Nvidia doesn’t care about OpenCL, so they don’t support OpenCL 2, which makes many things faster/easier/possible. Any GCN card supports OCL 2.2.

        Could be Nvidia likes vendor lockin, so they intentionally don’t optimize drivers for OpenCL so people are forced into CUDA.

          • brucethemoose
          • 3 years ago

          It’s probably “all of the above”.

          Doesnt change the point though. Open Compute seems to be “closer to the metal”, hence it represents a performance boost over OpenCL either way.

      • morphine
      • 3 years ago

      Well, simplifying a lot:

      Right now the “GPU-accelerated” HPC market belongs to Nvidia and its Tesla thingamabobs. CUDA is Nvidia’s “programming language” for GPU-accelerated applications. You can download the SDK and run it on your GeForce. A huge part of NV’s success came from the fact that even early on, the company offered an SDK of excellent quality. The critical mass built up along the years, coupled with Nvidia’s success with designing “GPUs” that fit the purpose very well.

      ROCm is CUDA’s counterpart, and wants to offer the same ease-of-use for developers and the capability of converting CUDA code so that it runs on AMD’s hardware. In short, AMD’s competing with Tesla/CUDA directly. This isn’t the red team’s first effort, but at first sight, it appears to be their best take on it so far.

      Disclaimer: I’m a developer, but not a CUDA developer. 🙂

        • beck2448
        • 3 years ago

        I doubt many professionals are going to want to lay another software on top to Port Cuda when they can just buy an Nvidia card which is always going to run it better anyway

          • ptsant
          • 3 years ago

          If I had to buy cards for a research project and spend like $20000 on hardware, I would not “just buy” nVidia. I would actually test both and consider the cost of porting. I’d like to think that many of the buyers are making rational decisions at this scale, not just submitting to habits. Could be wrong, though. Humans are not always rational.

            • beck2448
            • 3 years ago

            It’s simple logic. When ultra precision is critical, few pros, scientists, engineers, etc are going to fool around with another extraneous layer of software which can only produce more errors.

      • MLSCrow
      • 3 years ago

      The significance of this is in the fact that with this, companies now have a cheaper alternative to purchasing the Intel and nVidia combo in order to get what they want done. They will now have AMD Zen as an option on the CPU side of things and AMD as an option on the GPU side of things.

      Now, that alone isn’t enough to get anyone to jump the ship they are already in, however, AMD will be releasing server based APU’s, like Raven Ridge, which will be a much lower cost solution than buying a server CPU and discrete server GPU separately.

      Since nVidia doesn’t have an x86 license, people generally had to purchase Intel CPU’s and nVIdia GPU’s. Now, they can purchase a single APU, saving them money over purchasing CPU+GPU, while being able to port over any previous code that was used for previous hardware, with ease.

      It’s a rather big deal in that market space.

        • chuckula
        • 3 years ago

        [quote<]They will now have AMD Zen as an option on the CPU side of things and AMD as an option on the GPU side of things. [/quote<] Is there any evidence that Zen is cheaper than a comparable Xeon? Is there any evidence that HPC-grade AMD GPUs are actually cheaper than their Nvidia counterparts?

          • Lord.Blue
          • 3 years ago

          This is more for constructing a lower cost solution(ie one without a discrete card) that would be greatly expandable with the addition of AMD Fire board solutions or even nVidia Tesla board solutions, but the stand alone lower cost solution would still be able to run the code due to the on-board GPU in the APU.

            • Amiga500+
            • 3 years ago

            There are a few things that spring to mind.

            1. The migration of AMD APUs toward a unified single addressable memory space with the HSA program.
            2. Evolution of HBM in the discrete GPU space, possibly moving to high-end APUs in the future.
            3. AMD’s awareness of I/O on their SSG Radeons.

            **If** (and I’ve been burnt with Barce and Bulldozer, so its a big if) Zen is competitive, then AMD could offer a compelling HPC ecosystem that is maintained under one roof that has good communication between x86 and GPU, large memory bandwidth and excellent I/O, along with the resulting robustness and efficiency that **should** result from that relative to their competition.

            Yes, Intel may have a quicker x86 CPU, Nvidia may have a quicker compute GPU, Nvidia may have a more accessible language in CUDA and Intel may have a faster storage medium in Octane – but AMD could be able to put the whole lot together such that their sum exceeds the performance of the individual parts of the competition.

          • beck2448
          • 3 years ago

          Not to mention Nvidia has entrenched solutions in the space, has customers galore including Microsoft, Amazon, the US Government, Princeton, Stanford and a host of other top Tier clients who will ALWAYS choose the first class option. AMD is fighting an uphill battle on both the hardware and software side, and dual GPU ??? Get serious.
          Branding and marketing has not been great for AMD for a long time.

        • ptsant
        • 3 years ago

        I don’t think any APU has (a) professional driver support and (b) the power to compare with Teslas.

        There are rumored server APUs, but for the moment the compute cards from AMD are the S-series, like the S9300 (2×4096 GCN cores). The fastest APU is stuck at 384 cores.

          • Waco
          • 3 years ago

          I spoke with AMD reps yesterday, no server APUs are on the roadmap.

          So, they’re either crazy, or they’re just keeping things close (which is also crazy IMO).

    • chuckula
    • 3 years ago

    I know I shouldn’t, but I’m going to anyway:

    In 2016 we’re gonna ROCm!
    In 2017 we’re gonna SOCm!
    LET’S MAKE HPC GREAT AGAIN!

      • Ari Atari
      • 3 years ago

      ROCm SOCm robots? =p

Pin It on Pinterest

Share This