AMD Radeon Vega RX 64 8GB Review

[InvalidError]

And perhaps even the double-rate fp16 capability of the P100 might not reach any lower than the V102.

To me, FP16 seems like a logical fit for HDR if you want to be able to calculate arbitrarily bright or dark pixels without clipping while also maintaining enough precision to avoid obvious banding from rounding.

 

[TMTOWTSAC]

I invested in Freesync monitor and now need something better than my current RX 580 to power it. I will likely just buy the Vega 56 once the custom designs come to market, or buy stock version and waterblock, assuming I can find one at $399 USD. People are using the increased power requirements as a major showstopper for buying Vega. Lets say that if you buy Vega and find out that there is a change in full load power requirements from your old card to Vega of 85W, the cost in additional power is minimal, based on my power costs here in Nevada. If I play 20 hours of gaming a week, I end up paying an additional 20 cents per week (12c per kwh in Nevada) based on calculation wattage x hours used ÷ 1000 x price per kWh= cost of electricity
If I refer to gaming load results at Anandtech, during BF1 gaming, there is a difference of 51W between the RX 580 and the Vega 56 and 142W between the RX 580 and Vega 64, so if I use the example above for calculating additional power costs then the swap to a Vega 56 will cost an additional 12 cents per week and swap to Vega 64 will cost an additional 34 cents per week.

Perhaps my math is wrong? While the talk of increased TDP and power requirements are initially alarming, the additional costs look to be minimal. These increased power requirements won't stop me from buying Vega.

Having said this the additional cost will hurt miners who are using their computers 24/7 to mine.....

If you compare the air cooled 64 against the 1080 FE in the same Anand BF1 test, you get a difference of 149 watts. So for your 20 hours a week usage, that amounts to about $18.60 more per year. If you can find a Vega 64 for $18.60 less than a 1080, it's basically a wash if you use it for a year. If you hang on to it for 2 years, you'd need to find it for $37.20 less. That could take a while until the pricing and availability settles down.

Another potential issue is the total power requirements. A lot of people have invested in more efficient, smaller PSU's. If you have to upgrade your PSU that's more added cost. The Anand test system could probably have gotten away with a 500w PSU. But that's with no CPU OC, or even any hard drives.

Finally, there's the heat. Every review I've read so far tested with a watercooled CPU. Anyone on big air would be dealing with a lot more heat in their case, which would cause the card to hit its thermal limits that much sooner. And an extra 149 watts for the surrounding room is non-trivial too, especially during summer in Nevada. That's about 1/10 what an actual spaceheater puts out on high setting.

All else being equal efficiency is a big selling point. Obviously all else isn't equal though, such as Freesync. It would take an awfully long time for 149 watts to add up to the added cost of G-Sync...11 years?

 

[ammaross]

@Chris Angelini
The $399 and $499 price points are launch-day only "early adopter" prices for cards without bundled games. Since those dried up, the actual MSRPs are likely to jump $100 since there's only bundled Vegas going forward. A conclusion update/change may be in order after an RFC to AMD.

 

[10tacle]

All else being equal efficiency is a big selling point. Obviously all else isn't equal though, such as Freesync. It would take an awfully long time for 149 watts to add up to the added cost of G-Sync...11 years?

You should add $30 more at least for the better PSU.

It's interesting watching people try and defend Vega's horrendous power demands compared to the GTX 1070/1080. The very same people who brag about Ryzen being more efficient in wattage per core than Kaby Lake (which is a valid point to be sure). Suddenly efficiency is not important when it comes to GPUs?

Now if the Vega 64 was 25% faster than the GTX 1080 while consuming 25% more power, then I could validate it. But zero performance improvement over the 1080 using nearly twice the power? Uhm, no.

 

[InvalidError]

It's interesting watching people try and defend Vega's horrendous power demands compared to the GTX 1070/1080.

Nobody was defending Vega's worse performance per watt. The point being made earlier was that going with GSync would cost you ~$200 extra on the monitor up-front compared to AdaptiveSync from going with Vega and that it'd take you 10+ years to recover that cost difference if Vega costs you ~$20/year extra in power and cooling.

Vega may be less power-efficient but it is more cost-effective over its probable useful life once you account for the difference in monitor prices.

 

[rush21hit]

It was never about the bills. We know already that its still cheaper than 1080 while comparable in performance. Even for the long run. It would took years for it to matter. This has been discussed a long time ago. A 300watt GPUs are not that absurd to behold. As long as it deliver comparable performance on its competitor of its gen.
It was about the power it needs to run. Which, in effect, will demand a higher wattage of a reliable PSU. Ask anyone. Most people bought the 500 and 600watts PSU, tops. Look on the forum, only few use 750watts and up. The reason, ironically, is how GPUs from the last 2 gens from both sides became even more power efficient. The people who bought the GTX600 and 900 series, heck, even the R7/R9 200 series KNOWS they don't need a hefty PSU (the 290/x were borderline, but stil, Vega is even worse of a power hog). These are the people that probably would consider to choose between either Vega or 1070/1080. Also, anything 80+ at 750watt level is pretty darn expensive compared its own to lower wattage brethren. Was and still is today.

To adopt whichever Vega in their rig would demand a new PSU to complement it. For some, it may not be a problem. For everyone else, they'd get the 1080 instead to avoid the hassle.

 

[bit_user]

And perhaps even the double-rate fp16 capability of the P100 might not reach any lower than the V102.

To me, FP16 seems like a logical fit for HDR if you want to be able to calculate arbitrarily bright or dark pixels without clipping while also maintaining enough precision to avoid obvious banding from rounding.

Agreed. There should be opportunities to use fp16 in graphics. In fact, fp16 isn't new, even in GPUs. I guess there just wasn't enough demand to justify the extra complexity in the hardware to make fp16 fast.

Even packed-int8 arithmetic seems like it could find some uses in things like texture filtering and effects.

The question is whether this stuff will get used. AMD's shader compiler can probably find limited opportunities to use these, but the real benefit will at least depend on engine developers adding specific optimizations.

 

[Karadjgne]

And right after that ground breaking breakthrough in gpu performance, it'd get bought out by nvidia and squashed. Oh wait, my bad, that already happened with 3dfx, ageia etc

 

[Lucky_SLS]

On a side note, Nvidia has its game works that is being used in almost all the latest game engines. What's there for AMD? I don't think Vulkan has any feature tech, only a performance driven api.

 

[InvalidError]

Agreed. There should be opportunities to use fp16 in graphics.

I suspect it is going to be far more than just an opportunity: to do HDR, you need to do your image processing with at least 12 bits of INT precision if you don't want your LSBs on 10bits output to immediately turn into rounding error noise. A 12x12bits INT multiplier is about just as big as an FP16 multiplier (FP16 has only 11 significant bits to multiply, exponent bits get added), so you get no significant die area or power saving there. While FP16 may have one less LSB than INT12, you have to keep in mind that FP16 has millions of times more dynamic range coded in the exponent that will become the MSBs when the FP16 color space gets compressed to RGB10/12 output format.

Once HDR goes mainstream, I bet the whole pixel pipeline will eventually go FP16.

 

[bit_user]

On a side note, Nvidia has its game works that is being used in almost all the latest game engines. What's there for AMD?

AMD has a number of libraries to support games and other interactive graphics. These are now all open sourced.

http://gpuopen.com/games-cgi/

 

[bit_user]

Once HDR goes mainstream, I bet the whole pixel pipeline will eventually go FP16.

FYI, I think most shader code is currently fp32. The only reason to go fp16 would be as an optimization, and then only when it's fast on enough hardware (especially the low-end).

The good news, on that front, is that Intel GPUs have had 2x fp16 support since Broadwell.

 

[10tacle]

Vega may be less power-efficient but it is more cost-effective over its probable useful life once you account for the difference in monitor prices.

Well certainly that's true as I previously agreed with someone else on, but it's a variable. Ryzen Threadripper beats Skylake X in core watt efficiency per core, but you can throw just about any variable in the mix to justify one over the other. I'm just dumbfounded on the near double power draw of the RX 64 compared to the GTX 1080.

 

[Johnny_38]

"The GTX 1080 has been out for 15 months now"

It really doesn't matter how long the 1080 has been out. It is like saying Ryzen is a flop because Intel has been faster for ten years, since Ryzen also does not overshadow Intel everywhere.

Yes, we have hoped for more, but at least they have caught up with nVidia a lot. There is still a lot about Vega that we still will see from Vega, this is not the full story. Do you know what Radeon Pro SSG (Vega) or GPU compute is? Not everything is about games, is it?

Personally I also think FP16 will make a big difference in upcoming games.

AMD had to focus where the money lies, in that, they have chosen the best route!

 

[InvalidError]

Once HDR goes mainstream, I bet the whole pixel pipeline will eventually go FP16.

FYI, I think most shader code is currently fp32. The only reason to go fp16 would be as an optimization, and then only when it's fast on enough hardware (especially the low-end).

Sounds like you think that FP16 is slower. An FP16 multiplier is about 1/4th the size of an FP32 multiplier and in many cases, FP16 multiplication is achieved by partitioning the FP32 multiplier instead of using completely separate circuitry.

The only reason you are seeing FP16 being slower than FP32 on consumer GPUs is arbitrary vendor restrictions, as evidenced by how the same architectures in "professional" variants mysteriously become twice as fast at FP16 than FP32. Why only twice as fast instead of my 4X claimed above? Because without extending the input register size by one method or another, halving the operand size only doubles the number of possible inputs and outputs from those same registers.

If GPUs went FP16-centric, FP16 performance would be quadruple that of FP32 performance.

 

[kyotokid]

...OK, so for 8 GB of VRAM we are looking at a price about 60$ more than that for an Nvidia GTX 1080 which is less power hungry. So other than HBM2 memory, what is the advantage for 3D rendering (particularly since the only Open CL engine available right now is buggy LuxRender)?

 

[Lucky_SLS]

Latest news? The 400 for rx56 and 500 for rx64 are introductory offers. We can expect a 100 buck increase later...

 

[riz_76]

Are AMD or NVidia even aware that most of us end up buying the Radeon cards because of free-sync monitors being way cheaper that their G-sync counterparts?

If it wasn't the case, AMD sales would be even lower...

 

[bit_user]

There is still a lot about Vega that we still will see from Vega, this is not the full story. Do you know what Radeon Pro SSG (Vega) or GPU compute is? Not everything is about games, is it?

Personally I also think FP16 will make a big difference in upcoming games.

AMD had to focus where the money lies, in that, they have chosen the best route!

AMD did something pretty innovative with their SSG product.

As for compute, they have a price advantage (GFLOPS/$) over Nvidia. But that's street prices and once you account for the power efficiency differences (and don't forget to factor in additional air conditioning costs), it's not obvious to me they still have a winning proposition for real-world TCO (Total Cost of Ownership).

That said, I do appreciate what they're doing with ROCm (Radeon Open Compute) and HIP (their CUDA compatibility mechanism). Those efforts are important enablers, but the economics also have to be right.

 

[bit_user]

Sounds like you think that FP16 is slower.

Because it is. You're so wrapped up in thinking about the hardware implementation that you're forgetting the vast majority of GPUs out there have token support for it, at best. All of the consumer Pascal GPUs have fp16 performance that's 1/64th the performance of fp32. They only included one 2x fp16 unit per SM.

An FP16 multiplier is about 1/4th the size of an FP32 multiplier and in many cases, FP16 multiplication is achieved by partitioning the FP32 multiplier instead of using completely separate circuitry.

That's how Nvidia did it, in the GP100, but I don't know about Intel or AMD. It still costs die area and power, as well. All for a feature that is unused in most/all games, today. That's probably why it didn't take hold until machine learning offered a sufficiently large benefit to justify adding it.

The only reason you are seeing FP16 being slower than FP32 on consumer GPUs is arbitrary vendor restrictions, as evidenced by how the same architectures in "professional" variants mysteriously become twice as fast at FP16 than FP32. Why only twice as fast instead of my 4X claimed above? Because without extending the input register size by one method or another, halving the operand size only doubles the number of possible inputs and outputs from those same registers.

Great, so not only do they have the overhead of partitioning for fp16, but you also want them to double their registers and all the datapaths connecting them? Do you realize the GP102 has 2 million 32-bit registers? You could enable the use of register pairs (ignoring, for a moment, all the complexity that would add), but you'd still need to double the data paths.

If GPUs went FP16-centric, FP16 performance would be quadruple that of FP32 performance.

It won't happen. The reason being that fp16 can only be substituted for fp32 in the minority of cases. Otherwise, they'd have done it, already. Like I said, fp16 (aka half-precision) has been kicking around in shading languages, for a while. x86 CPUs even have a pair of instructions for converting to/from it, which I'm pretty sure was a nod to GPUs, since those are the only fp16 instructions they currently have.

The route Nvidia seems to be taking to unlock the full potential of fp16 is with their special-purpose tensor engines. Great for machine learning, but not very interesting for other fp16 opportunities, like your HDR example. To my earlier point, the rise of machine learning might not actually benefit graphics processing horsepower, in the end.

 

[bit_user]

other than HBM2 memory, what is the advantage for 3D rendering (particularly since the only Open CL engine available right now is buggy LuxRender)?

AMD has ProRender, which they say is supported in 3DS Max, Blender, Maya, and SolidWorks.

https://pro.radeon.com/en-us/software/prorender/

They also can use the GPU's 8 GB as a cache, potentially enabling you to render much larger scenes. You'll want to find some benchmarks showing how well it actually works for rendering, before you buy.

 

[Sakkura]

It was never about the bills. We know already that its still cheaper than 1080 while comparable in performance. Even for the long run. It would took years for it to matter. This has been discussed a long time ago. A 300watt GPUs are not that absurd to behold. As long as it deliver comparable performance on its competitor of its gen.
It was about the power it needs to run. Which, in effect, will demand a higher wattage of a reliable PSU. Ask anyone. Most people bought the 500 and 600watts PSU, tops. Look on the forum, only few use 750watts and up. The reason, ironically, is how GPUs from the last 2 gens from both sides became even more power efficient. The people who bought the GTX600 and 900 series, heck, even the R7/R9 200 series KNOWS they don't need a hefty PSU (the 290/x were borderline, but stil, Vega is even worse of a power hog). These are the people that probably would consider to choose between either Vega or 1070/1080. Also, anything 80+ at 750watt level is pretty darn expensive compared its own to lower wattage brethren. Was and still is today.

To adopt whichever Vega in their rig would demand a new PSU to complement it. For some, it may not be a problem. For everyone else, they'd get the 1080 instead to avoid the hassle.

You can get an EVGA BQ850 for $65. That's really not bad considering a decent unit in the 550-600W neighborhood will run you at least $40-50 most times.

 

[InvalidError]

Because it is. You're so wrapped up in thinking about the hardware implementation that you're forgetting the vast majority of GPUs out there have token support for it, at best. All of the consumer Pascal GPUs have fp16 performance that's 1/64th the performance of fp32.

And you are wrapped up in thinking this won't change in future GPUs now that HDR and AI which are perfectly fine with reduced precision are a thing. Can't use those with older GPU designs that either cripple FP16 performance or have no/limited native support for it.

As for "token support for FP16 in the consumer space", Vega does bring packed FP16 to the consumer space. I'd expect many games that are still actively supported to get updates to add FP16 shaders for GPUs that fully support FP16. I wouldn't be surprised if that was part of the secret behind the few titles where Vega 64 beats the GTX1080.

 

[bit_user]

Because it is. You're so wrapped up in thinking about the hardware implementation that you're forgetting the vast majority of GPUs out there have token support for it, at best. All of the consumer Pascal GPUs have fp16 performance that's 1/64th the performance of fp32.

And you are wrapped up in thinking this won't change in future GPUs now that HDR and AI which are perfectly fine with reduced precision are a thing.

No, that's not my position.

What happened is that you suggested HDR would be a trigger for shaders to start using fp16 (seeming to imply that they currently used integer arithmetic) and for hardware to support it. Then, I pointed out they already use fp32, and fp16 would only be used as an optimization, but it would first need to be fast enough on low-end hardware. What I meant by that last clause was that low-end GPUs would need more than token fp16 support, but you seemed to think I meant that fp16 wasn't fundamentally more efficient. Then, I explained what I meant by the current lack of support for it and the chicken-and-egg problem of getting it widely supported, and that it's still not a given that it'll become de facto standard.

So, I'm neither saying that it won't, nor that it shouldn't happen. However, I am saying it's not a small matter for them to implement general-purpose fp16 arithmetic at 4x the rate of fp32. I didn't know if you appreciated the sheer scale of what you were criticizing everyone for not doing.

Vega does bring packed FP16 to the consumer space.

Source, please? This would be great news. Then, we'd just need AMD to trickle it down to their next-gen entry-level GPUs and for games/engines to start using it.

BTW, game engines have supported HDR rendering for quite a while. The difference that HDR displays make is that it would potentially avoid the renderer needing to perform tone-mapping. Although I don't know much about new HDR displays, I'd be surprised if renderers didn't still need to perform some degree of dynamic range compression.

 

[InvalidError]

I didn't know if you appreciated the sheer scale of what you were criticizing everyone for not doing.

There is no 'sheer scale' for doing it if you do it from the ground up instead of attempting to shoehorn it into an existing architecture.

Vega does bring packed FP16 to the consumer space.

Source, please? This would be great news. Then, we'd just need AMD to trickle it down to their next-gen entry-level GPUs and for games/engines to start using it.

Just do a search for Vega FP16. A bunch of current and near-future games have already announced support for FP16 so I'd say support from the game engine side of things appears to be on a good start.

BTW, developers who have benchmarked FP16 on the PS4 Pro are seeing ~30% performance boost from simply switching to half-precision.