Clock for Clock, Vega VS FuryX and discussion

[TheoneandonlyMrK]

I'm not saying it's representative of what it will be, if you read that correctly, I'm saying that v100 was paper launched in may 2017 and available in q3 same year, ppl suggesting that we're have to wait 1.5 year for the gaming variant is a a stretch.

Your right in regards to time but calling it voltas a stretch , why would Nvidia all of a sudden put the 64bit back in consumer land or the tensor units or an 18000$ chip for that matter.
Its going to be Another pascal/maxwell/kepler refresh called volta.

 

[Fluffmeister]

The fact that GV100 is already in the hands on some researchers bodes well as it's largest most complex chip in the lineup, stands to reason stripping back elements that are of no relevance to gamers shouldn't be a big deal.

As it stands, the still have the performance crown with a gently sliced GP102 anyway so no hurry, besides this is all massively off topic.

 

[Vya Domus]

I'm not saying it's representative of what it will be, if you read that correctly, I'm saying that v100 was paper launched in may 2017 and available in q3 same year, suggesting that we have to wait 1.5 year for the gaming variant is a stretch. Also, unless he means that amd will just do a straight rebrand of the cards that literally just launched, expecting team red to launch another series before volta is not something I can really logically comprehend given how much time it usually takes them to release a high end card.

Why is it a stretch ? AMD screwed up the typical release schedule , Nvidia is in a tight spot even if RX Vega did not dethrone them.

They simply don't have a clear cut plan anymore as of right now , they spent so much time and money piling up features on the Tesla cards while stripping down the Geforce line-up and now AMD is offering monstrous FP16 performance for peanuts.

I mean what are they going to do ? Use Volta in it's current iteration with Tensor Cores on gaming cards and cannibalize the astronomically priced Tesla's while offering 0 advantage to gaming ?

Enable full FP16 performance on them and push it to developers ? Well Vega already has that , if they get developers to focus more on FP16 and get an uplift in performance using it , so will Vega.

Volta has no extra features to give an edge in gaming compared to Pascal. Going from 16nm to 14nm is a small jump , not enough to account for huge increases in clock speed or shader count like they could with Pascal , at least not without major compromises on power efficiency and costs.

They have no reason to release Volta as in a competitor for RX Vega, they would just be wasting money. It would be a smart idea to wait one year or even more and come up with a different/modified design from Volta to fight off Navi. Remember that Maxwell stayed around for about 2 years as well.

 

[Fluffmeister]

Market share for Vega will remain too small for devs to start pushing FP16 too much, we know AMD have got into bed with Bethesda but I don't see much changing anytime soon.

 

[Frick]

Market share for Vega will remain too small for devs to start pushing FP16 too much, we know AMD have got into bed with Bethesda but I don't see much changing anytime soon.

Ewww you mean AMD will only make dissapointing, shallow, stupid, overhyped products from now on? Wait, they already do, but the difference is people for some reason love Bethsoft specifically because they can write stories as well as they can write code, which is on the level of a horny 13-year old. "Immershun".

 

[TheoneandonlyMrK]

Market share for Vega will remain too small for devs to start pushing FP16 too much, we know AMD have got into bed with Bethesda but I don't see much changing anytime soon.

You keep saying fp16 , that's a general thing thats been about years but was principally pointless to use in that if all your code passes through 32bit compute engines and the engine in question cant go quicker using 16bit ops then their is no point using 16bit ops and people don't ,Rpm is a step ahead in its application in that Now a card will do four through a 64bit array in one pass ,calling it simply fp16 is either missleading or miss understanding its nature.
As is believing its not going to have much market share , given pc and console use cases it should see reasonable adoption by devs.
It only needs to be addopted by engine devs in reality ie Ea using it in one game on one use case enabled all Eas offices to also use it ,they may not but they already can.

 

[Fluffmeister]

You keep saying fp16 , that's a general thing thats been about years but was principally pointless to use in that if all your code passes through 32bit compute engines and the engine in question cant go quicker using 16bit ops then their is no point using 16bit ops and people don't ,Rpm is a step ahead in its application in that Now a card will do four through a 64bit array in one pass ,calling it simply fp16 is either missleading or miss understanding its nature.
As is believing its not going to have much market share , given pc and console use cases it should see reasonable adoption by devs.

Of course, but I just don't see loads of devs focusing that much support on it initially, when the vast majority of cards won't benefit.

Again, this is too much off topic.

 

[RejZoR]

Depending on how much work that means for developers. If it's a thing of flipping a switch on stuff devs build, then it could become a thing rather quickly. In the end, everyone can benefit from it. First AMD, if it proves to be useful, NVIDIA might jump on it as well. Or it might even become a part of DirectX eventually.

 

[the54thvoid]

More speculation going on here than a room full of investors at a speculative investment meeting.

If you all just take a step back and try making TPU less like a school playground and more like a tech debate, that'd be nice. As for Volta, stop talking crap about a product few people seem to actually have any knowledge or interest in.

https://devblogs.nvidia.com/parallelforall/inside-volta/

This explains a lot of the changes Volta makes over Pascal. There is a large drive to increase parallelisation through changes to the warp scheduler as well as other elements that taken together make Volta a very flexible GPU.

 

[xkm1948]

How come we are talking Volta now? We were discussing the architecture improvement of Vega over Fiji

 

[cucker tarlson]

Why is it a stretch ? AMD screwed up the typical release schedule , Nvidia is in a tight spot even if RX Vega did not dethrone them.

They simply don't have a clear cut plan anymore as of right now , they spent so much time and money piling up features on the Tesla cards while stripping down the Geforce line-up and now AMD is offering monstrous FP16 performance for peanuts.

I mean what are they going to do ? Use Volta in it's current iteration with Tensor Cores on gaming cards and cannibalize the astronomically priced Tesla's while offering 0 advantage to gaming ?

Enable full FP16 performance on them and push it to developers ? Well Vega already has that , if they get developers to focus more on FP16 and get an uplift in performance using it , so will Vega.

Volta has no extra features to give an edge in gaming compared to Pascal. Going from 16nm to 14nm is a small jump , not enough to account for huge increases in clock speed or shader count like they could with Pascal , at least not without major compromises on power efficiency and costs.

They have no reason to release Volta as in a competitor for RX Vega, they would just be wasting money. It would be a smart idea to wait one year or even more and come up with a different/modified design from Volta to fight off Navi. Remember that Maxwell stayed around for about 2 years as well.

Explain two things to me. To what extent is that monstrous fp16 performance going to affect gaming performance and how does that compare to nvidia's ability just to release a 600mm2 chip like they did with GM200 and upgrade it with ridiculosly fast GDDR6. I think nvidia has the ability to make a card that would smash V64. I mean even looking at the gap between v64 and 1080Ti, do you take it for granted that with rapid packed math amd is certain to top that by such a big margin that volta will be in such a tight spot?
And second thing, are tensor cores necessary for nvidia to release a card with double fp16 performance like vega ? AFAIK P100 offers double fp16 as well, I don't see no tensor cores on it.

 

[cdawall]

Well, I can tell you that there won't be any Volta till 2018 for sure. Partially because Volta is so expensive to make and partially because AMD isn't really pushing the envelope with Vega.

I was gaming pretty comfortably with GTX 980, going to GTX 1080Ti means I'll be good for minimum 2 years like I was with GTX 980. Meaning my next replacement will be just in sync with Volta and Navi if all goes as predicted.

So how do you know how much all of the Vega Volta architecture costs? All brands have somewhat accelerated launch schedules on cards, I would not be surprised if we saw something new (or refreshed) for this holiday season.

 

[Vya Domus]

This explains a lot of the changes Volta makes over Pascal. There is a large drive to increase parallelisation through changes to the warp scheduler as well as other elements that taken together make Volta a very flexible GPU.

Flexible indeed , for compute that is. With each new iteration they are reintroducing and improving the hardware scheduling/control capabilities that they got rid off with Kepler. Volta is a compute oriented architecture , even more so than Vega.

A truly unexciting time to be a gamer , almost every advancement now is meant to benefit GPGPU/heterogeneous computing.

Explain two things to me. To what extent is that monstrous fp16 performance going to affect gaming performance and how does that compare to nvidia's ability just to release a 600mm2 chip like they did with GM200 and upgrade it with ridiculosly fast GDDR6. I think nvidia has the ability to make a card that would smash V64. I mean even looking at the gap between v64 and 1080Ti, do you take it for granted that with rapid packed math amd is certain to top that by such a big margin that volta will be in such a tight spot?
And second thing, are tensor cores necessary for nvidia to release a card with double fp16 performance like vega ? AFAIK P100 offers double fp16 as well, I don't see no tensor cores on it.

Not exactly sure what are you talking about , my point was Volta is not fit/meant to compete with Vega from a gaming point of view and odds are Nvidia might prefer to wait on this one , all those things you said are exact reasons on why this might just be the case.

The monstrous FP16 performance is the reason why Nvidia will not give the average Joe fully capable Volta cards , at which point you might just as well call it a Pascal refresh.

Yes they have the ability to make a 600 mm^2 Pascal chip , but why ? To beat whom ? Their own cards ? You know how expensive these dies are do you ? Nvidia is probably paying TSMC money hand over fist for those V100 cores. They are not Intel you know , they'll try and avoid using such low yields chips for consumer parts in order to make room for what they ship to data-centers.

GDDR6 ridiculously fast ? Where did you get that from ? First batch of GDDR6 chips are supposed to be 14Gbps ,that's faster than current GDDR5X chips by about what , 20% ? Ridiculously fast ? By what standards ?

You do know that for example on Vega with HBM2 a doubled bus-width would give you 1 terabyte of bandwidth. Hell P100 already has about 700gb/s bandwidth. There goes your ridiculously fast GDDR6. Cheaper than HBM perhaps , ridiculously fast nah.

 

[BiggieShady]

Sometimes performance per clock (or IPC) is lower in a more modern architecture (or process), because it makes something else possible, so in the end performance is higher. You're totally ignoring this aspect.

I'm not ignoring it in this case since clock for clock vega is still faster than fiji ... but I can see how my post may have come across as a general statement

I don't see how, yes you can run crysis on both at same clocks and compare how they both run in effect legacy software .
But most of the architecture changes were for non functioning in that test ,features ,and i dont buy gpus for games i can already run fine.
I buy them for games due to be released.
But it Is an easy way to troll Amds software team who admittedly have not yet enabled CBR or rpm etc etc.

I'm certain AMD driver team is still finishing vega drivers, it will be interesting to run these vega/fiji clock for clock tests in a year or so

Again, this is too much off topic.

Discussing one of the improvements in the thread about architectural improvements is not off topic.

Devs can get creative when opportunity to extract additional performance exists ... speeding up complex shaders by using fp16 instead fp32 where it isn't visually noticeable (half precision = banding on gradients). Good example is using half precision for reflections in PBR shaders. Shader compiler should use packed math automatically on vega (also a wait for that to be finished)

 

[the54thvoid]

Flexible indeed , for compute that is. With each new iteration they are reintroducing and improving the hardware scheduling/control capabilities that they got rid off with Kepler. Volta is a compute oriented architecture , even more so than Vega.

A truly unexciting time to be a gamer , almost every advancement now is meant to benefit GPGPU/heterogeneous computing.

Time to buy a console....

 

[Vya Domus]

Time to buy a console....

The situation is still not that bad in order for that to happen thankfully

But yeah the days when engineers would only have to worry about figuring out ways to make games run faster are unfortunately over.

 

[Fluffmeister]

How come we are talking Volta now? We were discussing the architecture improvement of Vega over Fiji

Indeed, let's start with the sins!

https://videocardz.com/blog/all-amd-radeon-rx-vega-64-sins

Some really dodgy console ports may or may not save the day.

 

[xkm1948]

Indeed, let's start with the sins!

https://videocardz.com/blog/all-amd-radeon-rx-vega-64-sins

Some really dodgy console ports may or may not save the day.

At least according to buildzoid testing, primitive discarding is 100% functional. HBCC should not affect performance as long as vram is not bottlenecked. All those millions of extra transistors and extra silicon real estate are purely for boosting clock speed. I can see RTG taking a page from Maxwell and Pascal. It may be the bad node manufacturer of GloFo that is causing such hideous power consumption. If Vega's power consumption is within 5% of 1080/1070, then the situation would look a lot better.

 

[Vya Domus]

GloFo undoubtedly bears some guilt for this situation.

 

[INSTG8R]

GloFo undoubtedly bears some guilt for this situation.

Except if you've seen the news about "different" chips being discovered and AMD confirmed they were using multiple fabs.

http://www.guru3d.com/news-story/am...iffer-physically-and-quite-significantly.html
Now that might just mean assembly and it's a one man show on the actual chip fab, but...

 

[Vya Domus]

Except if you've seen the news about "different" chips being discovered and AMD confirmed they were using multiple fabs.

http://www.guru3d.com/news-story/am...iffer-physically-and-quite-significantly.html
Now that might just mean assembly and it's a one man show on the actual chip fab, but...

Hmm , even if they would use another manufaturer ( which would most likely be TSMC ) they still need to build it around the same node with the same spec. While TSMC for example might be able to provide a better node , they would still have to cater to the lowest denominator , which would be GloFo.

One thing is for sure , GloFo is not among the best and they have been a hamstring for AMD for quite some time.

 

[xkm1948]

GloFo undoubtedly bears some guilt for this situation.

I agree. Some design flaw mixed with the incompetence of GoFlo. Perfect combination for a power hungry chip.

With RTG's current situation(being spun off as a separate group within AMD) they will probably have to milk both Polaris and Vega to death before getting to a new arc of GPU design. We may see power efficiency fine tuned version of Polaris and Vega next year. If not power effciency fine tuned, at least higher clocked(1.8GHz~2GHz?) version of Vega+

It is highly likely we will be seeing GCN for another 2~3yrs before we have anything exciting from AMD that can challenge the top end of graphic market

 

[phanbuey]

I agree. Some design flaw mixed with the incompetence of GoFlo. Perfect combination for a power hungry chip.

With RTG's current situation(being spun off as a separate group within AMD) they will probably have to milk both Polaris and Vega to death before getting to a new arc of GPU design. We may see power efficiency fine tuned version of Polaris and Vega next year. If not power effciency fine tuned, at least higher clocked(1.8GHz~2GHz?) version of Vega+

It is highly likely we will be seeing GCN for another 2~3yrs before we have anything exciting from AMD that can challenge the top end of graphic market

I agree, but i think GCN is about to bust into the high end space as they are just going to start gluing these suckers together with infinity fabric...

very likely they will, as you say, start powertuning, and then go massive mcm at the same time.

 

[xkm1948]

Now you mentioned IF, or shall we say FineGlue? There were something in the Vega reviews asking the connection speed of IF between Vega core and HBM2. This may be another limiting factor of Vega performance.

 

[phanbuey]

Now you mentioned IF, or shall we say FineGlue? There were something in the Vega reviews asking the connection speed of IF between Vega core and HBM2. This may be another limiting factor of Vega performance.

Any interconnect will choke performance, but if the approach is that you can pump out 4 downclocked/harvested vega (say 4x56) cores sitting on IF - even with a penalty - that should outperform any monolithic design currently available.

I agree with you though - im sure in the beginning it will be a bit of a frankenchip.

https://www.overclock3d.net/news/gp...ega_utilises_their_new_infinity_fabric_tech/1

"well this new vega sure runs great at 4K, but due to the IF, if you try to run it above 200fps, your house explodes"...