New Performance Benchmarks of AMD's Vega Frontier Edition Surface

[RejZoR]

Rejzor.
I learned one thing from this. How to add someone in the ignore list.

Ignorance is bliss. Stay dumb forever.

 

[jabbadap]

Stop right there, I'll have to correct you:
Vega FE does not have full fp64, 819 Gflop/s fp64. AMD is planning a Vega20 to arrive next year with higher fp64 performance to compete with Quadro.

...

Not exactly wording I would have used. VEGA FE has full fp64, which vega10 chip can offer it's not neutered like i.e. hawaii series was. But other than that I agree Vega20/navi should be the next fp64 chip for scientist that really needs higher precision. Now that chip is Firepro W9100.

 

[okidna]

GeForce cards (gaming ones) also have "Optimize for Compute" setting in NV CP which has a disclaimer that can negatively impact games, especially those that use sparse textures.

Not all GeForce cards, only Maxwell based.

 

[Th3pwn3r]

The claim that prosumer and professional cards suck in gaming is a myth

Ah, so cards priced far above others that offer similar performance to far cheaper cards are GOOD FOR GAMING? BRB spending 5k on a card instead of buying a 1080ti to play games with it, that 'makes sense' doesn't it?

 

[rtwjunkie]

Rejzor.
I learned one thing from this. How to add someone in the ignore list.

Come on, you are better than that. This is the real world. People will disagree and that's a wonderful thing! It means we are all free thinkers, and actually using our brains.

We should be happy for a world in which everyone doesn't think the same and expresses a different viewpoint. To ignore him because you disagree is to ignore your own humanity.

 

[efikkan]

Not exactly wording I would have used. VEGA FE has full fp64,

No, it does not have full fp64 performance.
If you bothered to check the link I provided, you'll see Vega FE have 819 GFlop/s fp64
Compared to Quadro GP100 with 5200 GFlop/s fp64

 

[xkm1948]

I really would like to see VEGA FE clocked at 1050 with HBM overclocked to the same bandwidth as FuryX and pitch them toe to toe.
Also it seems nobody has mentioned HBM2 timing issue. One thing I learned a lot when playing with my FuryX is the HBM is affect by both timing and speed, with timing affect performance a lot more than speed.

 

[DeathtoGnomes]

Is it misguided when they tell you what it is



People might have been expecting something else and it seams the issue is with their expectations.

The first time I agree with you on this thread. All these expectations from FE card is from selfish motives, the proverbial "I want ...", and it applies to those people bashing the card for not living up to their ignorant expectations.

 

[jabbadap]

No, it does not have full fp64 performance.
If you bothered to check the link I provided, you'll see Vega FE have 819 GFlop/s fp64
Compared to Quadro GP100 with 5200 GFlop/s fp64

Yeah, so? Vega10 chip has 1/16 ratio(819GFlops*16=13.1TFlops of fp32) of fp64/fp32 as is all gcns since tonga, you can't just change ratio without changing actual hw inside the gpu. So I'm not quite follow you how it does not have full fp64 performance. GP100 has full fp64 performance by ratio of 1/2 fp64/fp32, gp102 has full fp64 performance by 1/32 of fp64/fp32, R9-290x has neutered fp64 performance by 1/8 fp64/fp32 ratio while firepro w9100 has full fp64 performance by 1/2 fp64/fp32 ratio(like kepler titans with full vs 780&780ti neutered).

 

[john_]

Come on, you are better than that. This is the real world. People will disagree and that's a wonderful thing! It means we are all free thinkers, and actually using our brains.

Not always
I don't have a problem with disagreement. I have a problem with people, who keep reacting like they don't understand what you are saying, which is annoying especially when they are saying almost the same thing with you, but you know, with the right technical terms, making them experts and you "forever dumb". In this case, it's not actually an indication of brain activity, but more of game of brain endurance. I didn't had time (or pop corn around) to play that game, so I deleted my last two posts and blocked him. Based on his last post, by doing so, my last chance to learn something in this world, is gone. Forever.

We should be happy for a world in which everyone doesn't think the same and expresses a different viewpoint. To ignore him because you disagree is to ignore your own humanity.

Too much philosophy in this one. We are in a world where everyone has a different viewpoint. And that viewpoint is the correct one. Any other viewpoint is "nonsense", "invalid", "blindness", "ignorance", "dumb". Not my words.



Oh, I almost forgot.

On topic.
AMD confirms Radeon RX Vega is launching at SIGGRAPH | VideoCardz.com

 

[Th3pwn3r]

"We waited 2 years, we can wait another 29 days"

We can wait forever, it's not a good thing really.

 

[cdawall]

"We waited 2 years, we can wait another 29 days"

We can wait forever, it's not a good thing really.

Waits for paper launch of cards

 

[notb]

Yeah, so? Vega10 chip has 1/16 ratio(819GFlops*16=13.1TFlops of fp32) of fp64/fp32 as is all gcns since tonga, you can't just change ratio without changing actual hw inside the gpu. So I'm not quite follow you how it does not have full fp64 performance. GP100 has full fp64 performance by ratio of 1/2 fp64/fp32, gp102 has full fp64 performance by 1/32 of fp64/fp32, R9-290x has neutered fp64 performance by 1/8 fp64/fp32 ratio while firepro w9100 has full fp64 performance by 1/2 fp64/fp32 ratio(like kepler titans with full vs 780&780ti neutered).

1/2 is "full", because if a GPU had FP64 cores only, it'll have a 1/2 ratio as well (as a FP64 core can perform two FP32 operations).
As a result a 1/2 ratio is "full" and everything below is considered not optimal for FP64-intensive scenarios.

 

[RejZoR]

I think AMD didn't expect people to be throwing games at a card clearly meant for professional use. Sure, it can run games, but they never intended this, which is why they released it without half of new Vega performance boosting features for games, assuming they'll gradually add stuff when it's ready for gaming version RX Vega. I mean, the fact tiling doesn't work at all is a telling thing (and this is what actually brings the significant boost and efficiency to Maxwell/Pascal). Which makes you wonder what else isn't working yet for games. Primitive Shaders (since they are suppose to be a switchable thing when possible via profiles)? HBC? The whole DX11/DX12/Vulkan stack?

For compute you don't really need any of this apart from HBC maybe for massive data sets, you just throw data to it and it'll grind it with what it has already. Games aren't that simple really. Which is why I'm still very optimistic about it.

 

[xkm1948]

I think AMD didn't expect people to be throwing games at a card clearly meant for professional use. Sure, it can run games, but they never intended this, which is why they released it without half of new Vega performance boosting features for games, assuming they'll gradually add stuff when it's ready for gaming version RX Vega. I mean, the fact tiling doesn't work at all is a telling thing (and this is what actually brings the significant boost and efficiency to Maxwell/Pascal). Which makes you wonder what else isn't working yet for games. Primitive Shaders (since they are suppose to be a switchable thing when possible via profiles)? HBC? The whole DX11/DX12/Vulkan stack?

For compute you don't really need any of this apart from HBC maybe for massive data sets, you just throw data to it and it'll grind it with what it has already. Games aren't that simple really. Which is why I'm still very optimistic about it.

Sure bro. Luv your optimism. Keep it up and we will discuss this again once rx vega drops.

 

[RejZoR]

We will. Then we can at least spit at it justifiably and objectively.

 

[R-T-B]

Well... I've written on this forum before that new generations of gaming GPUs could be limited in mining and virtually no one believed me.

I expect the same from the accelerated Pascal Refresh.

That's silly considering mining is a general compute task. You limit mining, you limit compute, which many games use. Won't happen.

Also, mining isn't really a task that takes a ton of optimization. It's generally a generic opencl kernel.

You literally said they intentionally lower IPC to increase clocks. Which is complete BS. Yes, that is a correction and yes, you do get offended too easily.

Rej, lowering IPC increases clock ceilings via a generally less complex chip that eats less raw wattage. I really don't see how you aren't connecting the dots here. Less Transistors = less heat and odds that one won't clock up, in exchange for less IPC. Please try not to throw tantrums at users using accepted principles on issues you clearly lack understanding of yourself.

I tried to word this nicely, but this is all I could manage.

 

[efikkan]

Yeah, so? Vega10 chip has 1/16 ratio(819GFlops*16=13.1TFlops of fp32) of fp64/fp32 as is all gcns since tonga, you can't just change ratio without changing actual hw inside the gpu. So I'm not quite follow you how it does not have full fp64 performance. GP100 has full fp64 performance by ratio of 1/2 fp64/fp32, gp102 has full fp64 performance by 1/32 of fp64/fp32, R9-290x has neutered fp64 performance by 1/8 fp64/fp32 ratio while firepro w9100 has full fp64 performance by 1/2 fp64/fp32 ratio(like kepler titans with full vs 780&780ti neutered).

You know very well a fully featured GPU has double fp32 performance vs. fp64, since fp64 units can be designed to calculate two fp32 operations.

I think AMD didn't expect people to be throwing games at a card clearly meant for professional use.

Seriously? Are you still using that excuse? You have been around long enough to know there is no reason a prosumer card should perform any worse in gaming.

Sure, it can run games, but they never intended this, which is why they released it without half of new Vega performance boosting features for games, assuming they'll gradually add stuff when it's ready for gaming version RX Vega. I mean, the fact tiling doesn't work at all is a telling thing (and this is what actually brings the significant boost and efficiency to Maxwell/Pascal).

Which specific features are you talking about?

Tiled rasterization is a hardware feature. If it's is defective in Vega FE then it will be defective in RX Vega too.

 

[RejZoR]

And you think hardware features just magically operate without driver layer? Oh dear. Especially since there was info floating around about Vega's capability to switch between tiled and "classic" rendering mode. Meaning the software layer plays a huge role and is clearly not functioning as it should at the moment. Just because something is implemented in hardware, that doesn't mean it's not software/driver dependent.

Also all this "prosumer" BS, has anyone bothered actually looking at Vega FE webpage? Like 90% of the time they talk about compute and professional use and in the end they briefly mention about testing games as developer. But sure, go on with the "prosumer" nonsense. Whatever the hell that even means. Vega is Quadro class card. Quadros aren't "prosumer", they are for professional use only. Titan cards are pure gaming cards no matter what NVIDIA or anyone else says. They are for those gamers with endless wallets. That's the end of it.

Vega FE is what FirePro used to be in the past. They are just shifting the product presentation to different looks and naming schemes. Instinct will be what NVIDIA has with Tesla.

 

[efikkan]

And you think hardware features just magically operate without driver layer? Oh dear. Especially since there was info floating around about Vega's capability to switch between tiled and "classic" rendering mode. Meaning the software layer plays a huge role and is clearly not functioning as it should at the moment. Just because something is implemented in hardware, that doesn't mean it's not software/driver dependent.

Tiled rendering is enabled in hardware. The main purpose of it is to avoid memory bottlenecks and stalls, which is detected and the feature enabled in the hardware. Unless AMD has purposely disabled a central hardware feature, you have no case here.

Also all this "prosumer" BS, has anyone bothered actually looking at Vega FE webpage? Like 90% of the time they talk about compute and professional use and in the end they briefly mention about testing games as developer. But sure, go on with the "prosumer" nonsense. Whatever the hell that even means. Vega is Quadro class card. Quadros aren't "prosumer", they are for professional use only. Titan cards are pure gaming cards no matter what NVIDIA or anyone else says. They are for those gamers with endless wallets. That's the end of it.

Vega FE is what FirePro used to be in the past. They are just shifting the product presentation to different looks and naming schemes. Instinct will be what NVIDIA has with Tesla.

Deflection is not going to help your case. Both Titan and Vega FE is intended for the same market; developers, researchers and content creators. None of them can compete with a Quadro in fp64, which is why AMD will be bringing Vega20 next year to do exactly that. Claiming Vega (FE I assume) is a Quadro class card is a lie, so is claiming that Titan is a gaming/consumer product. Please stop spewing this nonsense.

 

[RejZoR]

AMD didn't DISABLE a "central" feature, they didn't ENABLE it since their drivers are clearly in a very immature state at the moment. But whatever, I have no case here apparently when being skeptical about how fully functional it is in current state.

Deflection XD Titan is a freaking gaming card. With your logic, GTX 1050 is a professional card as well then.

I'm done with Vega discussions till RX Vega is out, this shit is reaching hilarious levels of ignorance and stupidity. And it ain't from my side.

 

[jabbadap]

1/2 is "full", because if a GPU had FP64 cores only, it'll have a 1/2 ratio as well (as a FP64 core can perform two FP32 operations).
As a result a 1/2 ratio is "full" and everything below is considered not optimal for FP64-intensive scenarios.

You know very well a fully featured GPU has double fp32 performance vs. fp64, since fp64 units can be designed to calculate two fp32 operations.

So kepler was not fully featured by your wording. Tesla k20/k20x/k40/kepler titans/kepler quadro k6000 has 1/3 of fp64/fp32 ratio. What if amd releases hpc chip with 1/1 fp64fp32 ratio would you call it over filled or what?

EDIT: Ahh just readed your wording more precise. We just had a different PoV. Mine full was what hardware have, your full what hardware possible might have. So carry on.

 

[efikkan]

AMD didn't DISABLE a "central" feature, they didn't ENABLE it since their drivers are clearly in a very immature state at the moment. But whatever, I have no case here apparently when being skeptical about how fully functional it is in current state.

The mode switching from immediate to tiled rendering is done in hardware in real time, it actually switches back and forth due to changes in the workload. This feature is in no way implemented in the driver, so if it's not "functional", then the defect is on the hardware side.

Deflection XD Titan is a freaking gaming card. With your logic, GTX 1050 is a professional card as well then.

Both Vega FE and Titan are targeting developers, researchers and content creators. Nvidia has even stopped selling them in stores to emphasize this is not a consumer product. And to use your own argumentation, if Titan is a consumer product, then why isn't Vega FE just the same? There is nothing making Vega more professional than Titan. And your whole argument that Vega FE is "intended" for professionals, well Titan is too. You simply can't have it both ways; applying one logic to one of them, and defying that logic for the other, whichever fits your agenda at the time. You need to learn how to make a logical argument for your case instead of attacking your opponents.

 

[RejZoR]

So, tiling and immediate mode is done in hardware and hardware is magically suppose to know which one to use. LOL? Of course this stuff works on hardware level, do you seriously believe that I think it's done in software? The decision making which of these hardware implementations are used is however done in SOFTWARE. And clearly, AMD has massive issues with drivers at the moment. So, which part of it you don't understand now?

NVIDIA is not selling Titans in stores to have a higher margin and keep it all for themselves. C'mon, it's business 101. They don't have time to sell every GeForce by themselves, but they sure can handle sales of just 1 type of card...

 

[DeathtoGnomes]

stop using "if" here. Is it disabled on the hardware level? or is it disabled on the driver level?? The drivers tell it to turn on, its not automatically on when you boot your system.

The mode switching from immediate to tiled rendering is done in hardware in real time, it actually switches back and forth due to changes in the workload. This feature is in no way implemented in the driver, so if it's not "functional", then the defect is on the hardware side.


Both Vega FE and Titan are targeting developers, researchers and content creators. Nvidia has even stopped selling them in stores to emphasize this is not a consumer product. And to use your own argumentation, if Titan is a consumer product, then why isn't Vega FE just the same? There is nothing making Vega more professional than Titan. And your whole argument that Vega FE is "intended" for professionals, well Titan is too. You simply can't have it both ways; applying one logic to one of them, and defying that logic for the other, whichever fits your agenda at the time. You need to learn how to make a logical argument for your case instead of attacking your opponents.

the rest of this is absolutely mind boggling deduction here!