AMD Unveils the Radeon Vega Frontier Edition: Vega Starts in the Pros [Anandtech]

[Bacon1]

http://www.anandtech.com/show/11403/amd-unveils-the-radeon-vega-frontier-edition

This is the Professional based card, not the gaming one.

 

[Rifter]

but i thought 16GB HBC was impossible! lol

 

[raghu78]

but i thought 16GB HBC was impossible! lol

AMD already said at the start of the year that Vega will come with 8 and 16GB capacity. The 16GB variant is for pro graphics and uses 2 HBM2 8 Hi stacks. The consumer graphics version would use 2 HBM2 4 Hi stacks.

 

[ddogg]

Really excited to see this card (consumer) and thread-ripper debut. Hoping it gives Nvidia's 1080 and p100 cards a run for their money.

Sent from my Nexus 6P using Tapatalk

 

[Mopetar]

Really excited to see this card (consumer) and thread-ripper debut. Hoping it gives Nvidia's 1080 and p100 cards a run for their money.

The one benchmark that they showed off has Vega doing better than P100, but I'm guessing that's their best case benchmark so I wouldn't use it to draw broad performance comparisons.

 

[ddogg]

Yeah that slide needs to be taken with a lb of salt I never look much into company performance slides.

Sent from my Nexus 6P using Tapatalk

 

[Guru]

According to their slide the card is significantly faster than the Titan XP. Of course I don't take this at face value, but even if we assume AMD cheated by 5% on those slides in their best case scenario, that still leaves up to 10-15% performance advantage, which is really impressive.

We'll see how this translates to games though.

 

[alcoholbob]

According to their slide the card is significantly faster than the Titan XP. Of course I don't take this at face value, but even if we assume AMD cheated by 5% on those slides in their best case scenario, that still leaves up to 10-15% performance advantage, which is really impressive.

We'll see how this translates to games though.

Shouldnt we just go by their own tflop claims? 13 tflops is 2.24x rx480, which is
21% faster than gtx 1080.

That makes it 90% as fast as a stock 1080 ti.

 

[mohit9206]

Shouldnt we just go by their own tflop claims? 13 tflops is 2.24x rx480, which is
21% faster than gtx 1080.

That makes it 90% as fast as a stock 1080 ti.

Which pretty much confirms that consumer Vega will not be competing with 1080Ti?

 

[Paratus]

Shouldnt we just go by their own tflop claims? 13 tflops is 2.24x rx480, which is
21% faster than gtx 1080.

That makes it 90% as fast as a stock 1080 ti.

Which pretty much confirms that consumer Vega will not be competing with 1080Ti?

I don't think we can assume that Vega FLOPS compare to Pascal FLOPS the same as Fiji or Polaris FLOPS.

I mean Vega is the same layout as Fiji but has roughly 50% more transistors and 50% more clock speed. It's likely they have closed some, most, or all of the performance / FLOP gap with Pascal.

 

[Paratus]

Is it wrong to desperately want to see a laptop with 4X 8stacks of HBM2 (32GB @ ~1Tbps) and the largest Ryzen / Vega based APU that will fit?

 

[beginner99]

I'm pretty certain it will beat the 1080Ti at gaming. The more important question when will the consumer version actually be available and what will it cost? H1 2017 certainly has slipped for consumer Vega and this Frontier Edition thing is probably just to say we launched in H1, will cost and arm and leg and will have a tiny quantity.

If it takes till August for them to ship Vega in meaningful quantities, that simply said would be very, very late if not too late. My first goal is to upgrade CPU/mobo. So not much in a hurry bit if I needed a new gpu soonish, I would not wait any longer and buy NV now.

 

[antihelten]

I don't think we can assume that Vega FLOPS compare to Pascal FLOPS the same as Fiji or Polaris FLOPS.

I mean Vega is the same layout as Fiji but has roughly 50% more transistors and 50% more clock speed. It's likely they have closed some, most, or all of the performance / FLOP gap with Pascal.

Have we actually gotten any official numbers on the transistor count (or die size) of Vega?

This is the Professional based card, not the gaming one.

The most interesting thing here to me, is where the FP64 rate ends up. If Vega frontier is capable of a 1:2 FP64:FP32 ratio*, then Vega would actually stack up quite nicely against GV100 (as long as you don't need the tensor flops of GV100). Vega would then have FP16/FP32/FP64 performance of 26/13/6.5, compared to 30/15/7.5 for GV100, so about 15% slower, but that's 15% slower against the huge 815mm2 ~$15000-18000 monster that is GV100, with a 500-550 mm2 GPU.

*I know that the leaked slides for Vega 10 indicated a 1/16 ratio, but we also know that the FP64 ratio of Vega is configurable, and that certain previous versions of GCN was capable of a 1:2 ratio (Hawaii for instance) .

 

[Krteq]

Nope

 

[xpea]

The most interesting thing here to me, is where the FP64 rate ends up. If Vega frontier is capable of a 1:2 FP64:FP32 ratio*, then Vega would actually stack up quite nicely against GV100 (as long as you don't need the tensor flops of GV100). Vega would then have FP16/FP32/FP64 performance of 26/13/6.5, compared to 30/15/7.5 for GV100, so about 15% slower, but that's 15% slower against the huge 815mm2 ~$15000-18000 monster that is GV100, with a 500-550 mm2 GPU.

Can we stop this ~$15000-18000 non sens ? DGX Station has an official price of $69k with 4 GV100 but non only. it includes a full water cooling system (CPU and GPU), dual Xeon E5-2698 20 core CPUs, 256G DDR4 RAM, 4 x 2TB SSDs in RAID, 1500W Platinum PSU, and premium chassis. Sorry but all these parts don't come for free ! in Fact and from reliable source, V100 is less than 20% more expensive than P100. So depending on the model, you will get V100 PCIe cards from $6000...

and finally Vega FE has low FP64 performance (1/16 FP32). So not, it wont compete with V100 in HPC. Vega 20 will have 1/2 FP32 in 2018

Off-topic is not allowed.
Markfw
Anandtech Moderator

 

[Krteq]

Water cooling? Where?

//Nevermind, you was talking about DGX Station, not DGX-1.

And again, stop derailing every AMD thread with nvidia stuff, thx in advance

 

[antihelten]

in Fact and from reliable source, V100 is less than 20% more expensive than P100. So depending on the model, you will get V100 PCIe cards from $6000...

You're confusing the Quadro GP100 (which was roughly $5000) with the Tesla P100. The Tesla version of GP100 was selling for upwards of $10,000 (in the SXM2 formfactor), so 20% more would put V100 at $12,000. This is admittedly well below the $15000-18000 range, but also well above what Vega Frontier will likely be selling for (their last high end workstation card the W9100 launched for $4000, and the recent W7100 is a meager $800).

Of course GV100 based Quadro cards will undoubtedly also arrive somewhere down the road, and they should be much cheaper, but they aren't here yet, and probably won't be so for quite some time (it took about a year for GP100 to arrive to the Quadro lineup).

and finally Vega FE has low FP64 performance (1/16 FP32). So not, it wont compete with V100 in HPC. Vega 20 will have 1/2 FP32 in 2018

We have not received any official information on the FP64 performance of Vega Frontier, so you simply cannot know this with any kind of certainty. It's perfectly possible that it will end up there, but we just don't know for certain yet.

The only information we have in this regard are the leaked slides from Videocardz, but we now know that these are fake (or at the very least outdated), given that they mention Vega 20 on 7nm (AMD just confirmed that Vega will only be 14nm and 14nm+, with 7nm being used for Navi).

 

[Flapdrol1337]

We have not received any official information on the FP64 performance of Vega Frontier, so you simply cannot know this with any kind of certainty. It's perfectly possible that it will end up there, but we just don't know for certain yet.

If it were half of fp32 they probably would've mentioned it.

 

[antihelten]

If it were half of fp32 they probably would've mentioned it.

Yes probably, but my point is simply, that even if a 1/16 rate is the most likely, we still don't have any official confirmation one way or another.

 

[Elixer]

If you guys have questions for Raja, right now is the time to ask them! Will be going on until 6PM EST today.

I want to start things off today but saying thank you to everyone for all of your excitement, energy and enthusiasm for all things AMD, and in particular, for Vega.

Earlier this week we were thrilled to launch Radeon Vega Frontier Edition. We think it will have a big impact on machine intelligence and content creators. I also know some of you are disappointed that we didn’t launch RX Vega as well. I wanted to hold this AMA and have an open discussion with you about our Vega launches. And while we're not launching RX Vega today -- so I won’t be talking about pricing or launch date -- there are lots of rumors and innuendo I want to put to bed, and there are plenty of questions I can answer.

I know you guys can’t wait to see Radeon RX Vega. I know that a lot of you guys obsess over when you’ll be able to game on Vega. Your passion for all things Radeon is what drives every single person in the Radeon Technologies Group to push hard on all fronts – hardware and software engineering, display technologies, and successful developer relationships. Everyone at AMD sees how much you guys talk about Vega and how eager you are to get your hands on it, and it fuels us.

Please know that we’re working incredibly hard on Radeon RX Vega. You are the lifeblood of Radeon, and you all deserve a graphics card that you’ll be incredibly proud to own.

https://www.reddit.com/r/Amd/comments/6bklro/we_are_radeon_technologies_group_at_amd_and_were/

You can see all of Raja's replies by looking at https://www.reddit.com/user/gfxchiptweeter

 

[Paratus]

Well this answers my 8pin vs 6pin question:

I grabbed an engineering board from the lab on the way to the Sunnyvale auditorium, and that boards works well with a 6 and an 8 pin. We decided to put two 8 pin connectors in the production boards to give our Frontier users extra headroom

 

[Paratus]

This was interesting too:

Infinity Fabric allows us to join different engines together on a die much easier than before. As well it enables some really low latency and high-bandwidth interconnects. This is important to tie together our different IPs (and partner IPs) together efficiently and quickly.
It forms the basis of all of our future ASIC designs.

We haven't mentioned any multi GPU designs on a single ASIC like Epyc, but the capability is possible with Infinity Fabric.

They haven't mentioned it not they haven't done it!

Let the rumors commence!

 

[Bacon1]

Someone did a AMA Recap:

Thought I would recap the information that has been confirmed during the RTG Vega Frontier AMA today.

• Main goal of Vega was to create an architecture that can handle large data sets and game at 4K at 60fps.
• The demo during Financial Analyst day was an air-cooled Frontier Edition, not an RX Vega card. There will be water cooled versions of Vega that will run slightly faster.
• Frontier uses 2 x 8GB stacks of HBM2.
  
• Both HBM1 and HBM2 provide plenty of bandwidth.
  
• Raja will look into OC'ing HBM2 and a 16GB RX card.
• RX Vega will be shown at Computex. It will not be available the same week but nothing else has been ruled out.
• Frontier runs comfortably using 1x6-pin and 1x8-pin but RTG put 2x8-pin on the production card for more headroom.
• Raja is keeping his beard until Vega launches.
• Infinity fabric allows for the joining of multiple engines on a single die, and offers high bandwidth and low latency. There has been no mention of using Infinity fabric with multiple GPUs.
• Frontier was designed for an array of workload usages. RX Vega is for gaming and will be faster than Frontier.
• Vega will support Tensorflow, Cafe2, Cafe, Torch7 and MxNet via MIOpen.
• Pro versions support hardware virtualization. He did not state out right if this included Frontier/Vega or or not.
• The High Bandwidth Cache Controller (HBCC) helps increase minimum framerates and can improve performance even more if it's specifically coded for.
• Developing drivers for GPUs is really hard.
• Raja expects to grow ROCm to improve machine learning and compute. Another ROCm comment.
• Raja replies to a comment regarding particle physics simulation, saying this will be improved via the new cache and infinity fabric.
• New geometry pipeline in Vega improves throughput per clock cycle and will require no extra work on dev's part to utilize.
• Radeon Vega Frontier will be the fastest single GPU solution for compute.
• Radeon Instinct will provide dramatically better performance per dollar compared to the competition
• RX will have different drivers than Frontier that are optimized for gaming as well as additional goodies.
• Vega is the first GPU architecture to use Infinity Fabric and is in no way a re-hash of Polaris
• Radeon Chill will continue to be improved and will be updated soon.

Link to the full AMA.

https://www.reddit.com/r/Amd/comments/6bzu05/raja_kaduri_ama_recap/

 

[krumme]

I dont care for avg framerates only mins and 1% so hbcc sounds very nice here. Great.

I guess i can get a discount as 99.99% of consumers still buy from avg numbers