Radeon Vega Frontier Edition packs 13 TFLOPS and 16GB of HBM2

It seems that lately, 10 out of every five news pieces are about machine learning. During its Financial Analyst Day presentation, AMD talked a lot about going after high-margin, fast-growing markets. Machine learning is one of those markets, and AMD has unveiled its first product made specifically for compute performance. Meet the Radeon Vega Frontier Edition, meant for "data scientists, immersion engineers, and product designers."

The Radeon Frontier Vega packs a whole 16 GB of HBM2, capable of running through data sets four times larger than the Radeon Fury X could hold. Performance for FP16 compute is pegged at around 25 TFLOPS, while the all-important FP32 compute speed is around 13 TFLOPS. (Nvidia's GeForce GTX 1080 Ti manages around 11.5 TFLOPS, but that's a very rough comparison if you're trying to get a sense of how these cards stack up for graphics power.) A couple of 8-pin PCIe power connectors feed the card some much-needed juice.

AMD has been on a product release roll since the beginning of the year, and the pace isn't letting on anytime soon. The Radeon Vega Frontier Edition will arrive in "late June." We can already spot its silhouette riding in from the sunset.

Comments closed
    • ronch
    • 3 years ago

    Hope they’ve addressed their efficiency issues with Vega. AMD really needs to get that part right.

    • psuedonymous
    • 3 years ago

    Same number of stream Processors as Fury.
    Raw double-precision rate: Fury = 13TFLOP, Vega = 8.6TFLOP. Delta = 1.51x
    Clock speed: Fure = 1.59GHz, Vega = 1.05GHz. Delta = 1.51x

    Apart from single-precision performance which has more than doubled (due to packed operations as we saw with the PS4 Pro update), and a slight reduction in total memory bandwidth (from 512GB/s on Fury to ~480GB/s on Vega) it’s looking like, in consumer/gaming workloads at least, that Vega is a Fury clock-bump, similar to the Polaris refresh.

    Will be interesting to see how ‘big Vega’ morphs into a consumer card. Dropping from the two 8GB HBM2 stacks to a pair of 4GB HBM2 stacks is an obvious one, but a SKU using GDDR5X might be an interesting option too. Maybe we’ll see a Titan-esque card launch using binned big Vega dies as a ‘halo’ product for consumers and a direct challenge to Nvidia’s high-end performance crown.

      • Pholostan
      • 3 years ago

      Yeah, it sure looks like a Fiji refresh. Very disappointing.

      • Bumper
      • 3 years ago

      How drunk are you right now?

      • Chrispy_
      • 3 years ago

      We still have absolutely no idea how much of an improvement the new architecture will make to gaming/consumer.

      Raw TFLOPS numbers have consistently been utterly useless at predicting game performance. NCU architecture could be [i<]slower[/i<] than GCN for gaming, it ought to at least match it or ideally be much faster. Undoubtedly some of that performance will come from drivers too, yet more reason why the raw throughput in TFLOPS isn't much help.

        • psuedonymous
        • 3 years ago

        [quote]Raw TFLOPS numbers have consistently been utterly useless at predicting game performance. /quote]In isolation I agree entirely. But within the same architecture the correlation is very strong. We have seen with PS4 -> PS4 pro the architecture change is minimal enough to maintain console back-compat, so I’m not expecting massive differences between Vega and other GCN cards.

          • Chrispy_
          • 3 years ago

          That’s the whole point though, this [i<]isn't[/i<] the same architecture. PS4 Pro is not Vega (NCU), it's Polaris (GCN); Specifically RX480 at 911MHz, so you shouldn't try and predict Vega based on PS4Pro performance.

      • chuckula
      • 3 years ago

      Unless there’s a pretty radically different piece of silicon that AMD has been uncharacteristically quiet about, there won’t be a GDDR5/5X Vega product. The memory technologies are radically different and require big changes in silicon.

      It’s possible that there is another piece of Vega silicon out there, but given how much AMD likes to show off Vega, it would be kind of strange for them to be completely silent about another Vega design.

        • ImSpartacus
        • 3 years ago

        I agree. Vega 10 probably only has hbm2 controllers.

        Vega 10 probably couldn’t use GDDR (of any variety) even if AMD wanted to.

        And with all of the fancy “high bandwidth cache” stuff that AMD put in Vega, even Vega 11 might need to use HBM2 (and many folks think Vega 11 might be a clean “half” of Vega 10 due to the rumor stating it’ll be at around 150W).

      • ImSpartacus
      • 3 years ago

      I agree that a “Titan” makes sense, though this “Frontier Edition” might be AMD’s Titan. Price isn’t really relevant. The titan program is about dick measuring. Victory for a halo product DOES trickle down to the lower brackets, even if it shouldn’t.

      Consumer Vega 10 will almost certainly sport 4 or 8 GB of HBM2. I think that might’ve been confirmed a couple months ago. There’s a video floating around where Raja goes into how Vega is more efficient with its use of vram capacity and that’s why they claim that an 8GB card can game like a 16GB (or more aptly, 11-12GB) competing card and a 4GB card can game like an 8GB competing card.

      But overall, it’ll still be two stacks of HBM2. That’s all but guaranteed.

        • psuedonymous
        • 3 years ago

        [quote<]I agree that a "Titan" makes sense, though this "Frontier Edition" might be AMD's Titan.[/quote<]Not if they price it for the HPC market. If you think the Titans are overly expensive, don't even go look at the price of Tesla or top-end Quadro cards!

    • backwoods357
    • 3 years ago

    AMD is a branding mess and they need to get their shit together and get the RX Vega out the door in the near future or they will lose traction. If they delay until Q4 or late Q3 people will just wait it out for the next GTX.

    • Ninjitsu
    • 3 years ago

    Could have called it Radeon Lyra, but no, “eypc frontier gamerz elite pro edition”.

      • ImSpartacus
      • 3 years ago

      Yeah, it’s pretty cringey.

      Sometimes I don’t know what AMD marketing is thinking.

      • nico1982
      • 3 years ago

      [quote<]Could have called it Radeon Lyra[/quote<] Nice!

    • kuttan
    • 3 years ago

    Impressive. AMD should name the new Vega as Radeon Koduri Edition lol.

    • the
    • 3 years ago

    Well the theoretical performance matches up, I guess it’ll determine how well the new architecture runs and how mature the drivers are to see if this Vega card can surpass the GTX 1080 Ti/Titan XP. The potential is there.

      • ET3D
      • 3 years ago

      At least for SpecViewperf it seems to work well, according to AMD: [url<]http://pro.radeon.com/en-us/vega-frontier-edition/[/url<]

    • caconym
    • 3 years ago

    gotta admit, I like the little LED “voxel”

      • UberGerbil
      • 3 years ago

      It looks good in the 3D render. I wonder what it’ll look like when it’s actually on real hardware?

        • backwoods357
        • 3 years ago

        [url<]http://images.techhive.com/images/article/2017/05/radeon-vega-100722929-large.jpg[/url<]

    • Beahmont
    • 3 years ago

    Good News Everybody!

    It’s got to be cheaper than GV100 because it’s slower at FP32 and FP16 and doesn’t even mention FP64 performance.

      • Voldenuit
      • 3 years ago

      I think a Tesla Model S might be cheaper than GV100 at this stage…

        • Beahmont
        • 3 years ago

        Psh… are you a real techie or what? Sure you can get one Tesla Model S for $70k, but why would you do that when you can get a workstation with 4 GV100’s for the same price! Priorities man.

          • Krogoth
          • 3 years ago

          What SKU of the GV100 are we taking about? The GV100 isn’t commercially available yet bro. It was just announced last week. Give it another quarter for commercial availability. I doubt the SKUs aren’t going to cheap unless there’s stiff competition from AMD (Not likely in the near-future)

    • tsk
    • 3 years ago

    Poor Volta. /s

    • lycium
    • 3 years ago

    > immersion engineers

    wat

      • RephRayne
      • 3 years ago

      Pool cleaners.

      Or possibly Lifeguards.

        • lycium
        • 3 years ago

        Aren’t lifeguards immersion QAs?

          • UberGerbil
          • 3 years ago

          No, immersion QAs wouldn’t save you, they’d just file a report describing exactly what you looked like while you drowned and the steps to exactly reproduce your drowning.

          Severity 1, Priority… 3, maybe.

            • Redocbew
            • 3 years ago

            Feedback: Attempted to reproduce issue by tossing a random sample of people in the pool. They all scrambled out dripping on the pavement and screaming profanities. Could not reproduce bug. Will not fix.

            Sounds like a fun job to me.

            • willmore
            • 3 years ago

            This is a disturbingly accurate comparison.

            • derFunkenstein
            • 3 years ago

            QA guys have it bad enough just trying to get stuff entered into Jira. You can’t possibly expect them to fix the problem, too.

            • Anonymous Coward
            • 3 years ago

            Well, if you want to prevent drowning, it would be useful to have detailed records of how it happens.

      • RAGEPRO
      • 3 years ago

      A very fancy way of saying “VR developers”, I suspect.

      • rika13
      • 3 years ago

      [url<]https://www.youtube.com/channel/UC4MGZcDG5hnzpi3hDpORkXw[/url<] This guy is an immersion engineer.

Pin It on Pinterest

Share This