Vega/Navi Rumors (Updated)

[ZGR]

My question has nothing to do with HBM2 or GDDR.
That's utterly irrelevant to what I'm asking.
Vega 11 is supposed to take that GTX 1070 spot and compete in that market. How does it do that in 2018 when the GTX 1070 is being phased out and the GTX 2070 is coming in?

Again, I'm not asking about what memory Vega 11 uses. That's not relevant.
I'm asking about the actual performance of the chip.

If Big Vega can't beat Big Pascal, how in god's name will Vega 11 be competitive against Volta midrange in 2018?

Fury wasn't the most competitive chip when it came against the 980 ti, but the RX 480 was a great competitor to the GTX 1060.

 

[Bacon1]

My question has nothing to do with HBM2 or GDDR.
That's utterly irrelevant to what I'm asking.
Vega 11 is supposed to take that GTX 1070 spot and compete in that market. How does it do that in 2018 when the GTX 1070 is being phased out and the GTX 2070 is coming in?

Again, I'm not asking about what memory Vega 11 uses. That's not relevant.
I'm asking about the actual performance of the chip.

If Big Vega can't beat Big Pascal, how in god's name will Vega 11 be competitive against Volta midrange in 2018?

I thought you were going to buy a 1080 / 1080 Ti, did that work out for you or are you still waiting on Vega that you don't think will compete?





Stop this back and forth with tential.


esquared
Anandtech Forum Director

 

[tential]

Fury wasn't the most competitive chip when it came against the 980 ti, but the RX 480 was a great competitor to the GTX 1060.

Neither of these examples are relevant to what we're talking about.

What does a midrange Vega chip in 2018 accomplish?

That's a chip slower than big Vega, but years after the competition. That doesn't raise an eyebrow to anyone and wonder how amd will sell that chip?

No one has any conjecture beyond the current thoughts that it competes against the gtx 1070 and 1080 and is just years late to the party?

Edit: I'll give my 2 cents, I didn't want to give my brilliant answer away but fine.
I don't think Vega 11 is meant to compete directly with the gtx 1070/1080.
I think it will replace Polaris, and come in with 2-4 chips with performance levels that don't remotely correlate with Nvidia and are just priced well. There won't be a way to direct compare like the rx 480 vs 1060. They'll slot in weirdly, and Polaris won't be needed beyond maybe a couple of cards at the lowest ends.(think fury(not x) , which didn't really directly compete with any card and was just priced well it's closest competitor was a cheaper gtx 980 and the 390x also. They sandwiched the gtx 980 which 2 better price/perf cards)

If we go off naming scheme, isn't Vega 11 made after Vega 10? Just how Polaris 10 released before Polaris 11.

Hence why I'm more curious about Vega 11.

Vega 10 makes sense, and is boring at this point we know the performance level of that product. Vega 11...im curious about.

Doesn't Polaris 11 have more features than 10? Isn't Vega 11 an updated 10?

 

[Karnak]

And what does a midrange Volta chip in 2018 accomplish?

 

[tential]

And what does a midrange Volta chip in 2018 accomplish?

Midrange Volta will be faster than what we have on the market currently. Unless you're suggesting it will be slower than GTX 1080Ti. Which fine, I'll eat that bait. Lets say a GTX 1180 is slower than the current GTX 1080Ti.
Fine, but Vega 11 isn't faster than Vega 10.

So Vega 11 and Midrange Volta aren't even on the same playing field.

Or do you believe otherwise? If so please explain, because so far, everyone has replied with a cheeky one liner or by just asking a question because they don't have any actual thoughts on Vega 11.

One of their bigger points is that you really, really don't need a crystal ball to predict Volta's arrival times/performance etc with a decent degree of accuracy. Heck you don't really need it for 'son of Volta' or even 'Grandson of Volta' come to that. NV are very predictable indeed these days. Mostly in a good way.

Trying to predict what AMD will do/have achieved with Vega is quite another matter

This man gets it. We know what to expect with Volta... And we know generally what to expect from Vega 10. So the real question mark is Vega 11.

Edit:
Forget it,
I'll make another thread to discuss Vega 11's potential performance. This does not seem to be the spot for it.
Edit2:
Looks like you're only allowed to discuss Vega in this thread. So I guess I'll talk about Vega 11 here.

 

[Head1985]

And what does a midrange Volta chip in 2018 accomplish?

Like always much lower TDP with little better performance.Probably finaly new architecture.First since maxwell because pascal is just overclocked maxwell.
GTX2080 10%faster than TITANXP and with 180w TDP.

 

[Karnak]

Like always much lower TDP with little better performance.
GTX2080 10%faster than TITANXP and with 180w TDP.

So we're calling a GTX 2080 "midrange"? Okay then..

The 1070 is not midrange, the 1080 is not midrange. A 2070 won't be midrange and a 2080 won't be midrange. Midrange is 1060/580 performance or with Volta 2060 performance.

 

[Head1985]

They are midrange technically.But NV have monopoly and sell them for more than twice they sell them few years back.X04 was always midrange and x06 was low end.

 

[zinfamous]

They are midrange technically.But NV have monopoly and sell them for more than twice they sell them few years back.X04 was always midrange and x06 was low end.

at this point, one must learn to start asking themselves if it is more important to define midrange as a cost proposition or a performance proposition, because those are two very distinct and non-overlapping categories with nVidia these days, especially when you look at those xx04 vs xx06 and xx07 lines. You already understand that they charge drastically more compared to those historic ancestor cards, simply because they live in a vacuum (so, without competition, how do you adequately make a real value/performance argument in that category now?) Well, you can't. So if they are now in a situation to define the price of whatever performance they want to deliver, due to lack of competition, you now accept that midrange and high end is also going to have to be re-defined, either way you go. One out of the two variables is now completely without measurable comparisons (thus creating a baseline)

 

[misuspita]

Yes, but hopefully AMD will be doing the same thing it did to Intel and recalibrate prices vs performance.

Both Intel and Nvidia enjoyed a monopoly at the high end, but that means AMD has the same opportunity as in the cpu market, to come up with a expensive chip which still is lower priced than the competition. So the asp 's will grow on the gpu front, even if they sort of undercut Nvidia

 

[amenx]

They are midrange technically.But NV have monopoly and sell them for more than twice they sell them few years back.X04 was always midrange and x06 was low end.

Would be funny if Nvidia priced the 1080/2080 at mid-range prices or half of what they would have sold for a few years ago. Although I dont think AMD would be laughing. Vega would then have to be pretty cheap too. It would basically amount to a death sentence for AMD, or call it quits on the GPU front at least.

 

[KompuKare]

Like always much lower TDP with little better performance.Probably finaly new architecture.First since maxwell because pascal is just overclocked maxwell.
GTX2080 10%faster than TITANXP and with 180w TDP.

All this Volta speculation is a bit off-topic in the Vega thread, but since the argument is that Vega will be doomed since GV104 is imminent, let's see what we do know from GV100 and its transistor count and die size:
Firstly, we more or less know that 12nm FFN doesn't bring any great area improvements (vs 16nm FinFET).
Therefore any anticipated +30% improvements with Volta have to come mostly from either 1) architectural changes, 2) higher clocks, 3) using a large die.
For 1. unless the change is as large as Kepler<>Maxwell it is unlike to be enough. And at this stage it must be hard for Nvidia to come up with a major architectural efficiency gain.
Point 2. is unknown. So that leaves point 3. - a larger die. That seems the most likely, but it won't necessarily do that well in terms of TDP. Of course Nvidia don't like lower margins, but if die size is their only solution they may have no choice.

 

[Qwertilot]

Actually for 1 we basically know they've done it from the very large power efficiency improvement(s) from GV100 vs GP100.

It amazes me a bit too to be honest - its hardly like Pascal was subcontracted out to a bunch of chimpanzees to design!

They'll obviously then leverage that into bigger dies for the performance uplift & rely on the more mature process to keep their margins roughly where they like them.

 

[beginner99]

Actually for 1 we basically know they've done it from the very large power efficiency improvement(s) from GV100 vs GP100.

It amazes me a bit too to be honest - its hardly like Pascal was subcontracted out to a bunch of chimpanzees to design!

They'll obviously then leverage that into bigger dies for the performance uplift & rely on the more mature process to keep their margins roughly where they like them.

How do we know they have done that? GV100 is just a huge die and hence performance is expected to be better. Clock it a bit lower so that efficiency is ok and done. I don't expect much gains to be honest at same die size. They have to give you more for the money so you get better performance/$ achieved by a larger GPU mostly. Pascal was all about going from 28 to 14nm. This won't happen this time. Kepler->Maxwell was pretty much a 1 time thing. the same change that will happen with AMD Fiji -> vega.
So even if vega competes with Volta, AMD isn't dommed.

 

[leoneazzurro]

Actually for 1 we basically know they've done it from the very large power efficiency improvement(s) from GV100 vs GP100.

It amazes me a bit too to be honest - its hardly like Pascal was subcontracted out to a bunch of chimpanzees to design!

They'll obviously then leverage that into bigger dies for the performance uplift & rely on the more mature process to keep their margins roughly where they like them.

They claimed increase in Teraflops/W. Which also includes the "tensor" cores which actually AFAIK do nothing in graphical workloads. And anyway Teraflop can be directly related to performance only if the architecture stays the same.

 

[CatMerc]

They claim 50% perf/watt improvement per SM in FP32 and FP64.

For it to be Maxwell like in nature, it would have to be the average improvement. If the average is more like 10%-30% and the 50% is under specific edge cases, then it won't be enough for another Maxwell.

 

[Qwertilot]

The other 'easy' way to think about it is that they've gone from 610mm2 and 15.3bn transistors to 815mm2 and just over 21bn in the same power budget.

So its very definitely non trivial. Quite how large/how well it'll translate into gaming we'll see soon enough.

 

[nathanddrews]

The other 'easy' way to think about it is that they've gone from 610mm2 and 15.3bn transistors to 815mm2 and just over 21bn in the same power budget. So its very definitely non trivial. Quite how large/how well it'll translate into gaming we'll see soon enough.

Ballpark, that's a 25% increase in die size with a 27% increase in transistors, so that's looking more like an average upgrade. Of course, they'll likely iterate and optimize upon how they arrange and gate their CUDAs and caches, so that usually nets a few more efficiency points. I don't really see any way for Vega to be 30% better than 1080Ti. Now, if AMD is able to abstract two Vega GPUs as one on the same PCB (sans crossfire)... Vega RX2?

 

[Valantar]

They claim 50% perf/watt improvement per SM in FP32 and FP64.

For it to be Maxwell like in nature, it would have to be the average improvement. If the average is more like 10%-30% and the 50% is under specific edge cases, then it won't be enough for another Maxwell.

The other 'easy' way to think about it is that they've gone from 610mm2 and 15.3bn transistors to 815mm2 and just over 21bn in the same power budget.

So its very definitely non trivial. Quite how large/how well it'll translate into gaming we'll see soon enough.

How many of those transistors are due to the tensor cores? The increase in CUDA cores is 40%, while the die size increases by 33%. Which in and of itself speaks to some area savings with the new process, even without the tensor cores. On the other hand, I was under the impression that they take up a not insignificant area on the die (at least according to the diagrams in AT's V100 article). So, might the 12nm process actually be delivering significant area savings? On the other hand, what does the area and transistors used for tensor cores say about power usage? After all, these cores are for very specific workloads, and I have the impression that when they are in use the rest of the GPU is likely not under heavy load. As such, with proper power gating, it would have pretty much zero impact on TDP. Then again, a 40% CUDA core increase for the same power is, again, a big jump. Could they manage that just by lowering clocks slightly? Probably, yes. But some architectural gains would seem likely

 

[turtile]

I don't see why everyone thinks Vega will perform badly. With a different architecture, they should be more similar:

Vega for the professional market (out first) = 12.5TF (we don't know the die size yet)
Volta for the professional market = 15TF (huge die and 12nm)

So we have around a 20% difference.

If Vega 11 is around 60% of Vega 10, that's around 7.5 TF and the 1070 is around 6.5TF. This is assuming that they keep the clockspeed low which I think is less likely on an improved 14nm and a gaming chip compared to a professional chip.

 

[T1beriu]

Vega [...] (we don't know the die size yet)

After AMD released the package shot of Vega I did the pixel counting myself using HBM2 SK Hynix dimensions as reference: 7.75 mm × 11.87 mm.

Vega (10) die size: ±489 mm², so just under 500 like Raja said.

It the VRAM is Samsung's then the measurements are not correct.

 

[tential]

I don't see why everyone thinks Vega will perform badly. With a different architecture, they should be more similar:

Vega for the professional market (out first) = 12.5TF (we don't know the die size yet)
Volta for the professional market = 15TF (huge die and 12nm)

So we have around a 20% difference.

If Vega 11 is around 60% of Vega 10, that's around 7.5 TF and the 1070 is around 6.5TF. This is assuming that they keep the clockspeed low which I think is less likely on an improved 14nm and a gaming chip compared to a professional chip.

It's not whether Vega will perform badly. It's since Vega is so late how will it fare against Nvidias upcoming Volta architecture. In all manners. Not just gaming. Anything.
These cards are engineered to do more than just game on need to just focus ourselves in that aspect alone. Gpus are used for a lot of cool things so yes it is interesting to talk about the improvements in the Vega architecture in relation to Volta.

To me it's even more interesting that Vega 11 will have more features. Essentially, amd is releasing an improved chip every time now. They're not releasing a full line up at all but releasing a chip, an improved chip, a new architecture, improved, chip etc....

It'll be interesting to see how this new strategy fares.

 

[Valantar]

It's not whether Vega will perform badly. It's since Vega is so late how will it fare against Nvidias upcoming Volta architecture. In all manners. Not just gaming. Anything.
These cards are engineered to do more than just game on need to just focus ourselves in that aspect alone. Gpus are used for a lot of cool things so yes it is interesting to talk about the improvements in the Vega architecture in relation to Volta.

To me it's even more interesting that Vega 11 will have more features. Essentially, amd is releasing an improved chip every time now. They're not releasing a full line up at all but releasing a chip, an improved chip, a new architecture, improved, chip etc....

It'll be interesting to see how this new strategy fares.

How do we know that Vega 11 will have features that Vega 10 won't? Has AMD said so?

 

[turtile]

It's not whether Vega will perform badly. It's since Vega is so late how will it fare against Nvidias upcoming Volta architecture. In all manners. Not just gaming. Anything.
These cards are engineered to do more than just game on need to just focus ourselves in that aspect alone. Gpus are used for a lot of cool things so yes it is interesting to talk about the improvements in the Vega architecture in relation to Volta.

I don't think that Vega is late. I think that due to limited R&D, they skipped over another major revision in order to add features to compete with Volta before it came out. AMD is focused on the professional market and they need to make sure they get chips in before it's too late. As you've seen, both Vega and Volta have a heavy enthuses on packed math.

 

[tviceman]

It's not whether Vega will perform badly. It's since Vega is so late how will it fare against Nvidias upcoming Volta architecture. In all manners. Not just gaming. Anything.
These cards are engineered to do more than just game on need to just focus ourselves in that aspect alone. Gpus are used for a lot of cool things so yes it is interesting to talk about the improvements in the Vega architecture in relation to Volta.

To me it's even more interesting that Vega 11 will have more features. Essentially, amd is releasing an improved chip every time now. They're not releasing a full line up at all but releasing a chip, an improved chip, a new architecture, improved, chip etc....

It'll be interesting to see how this new strategy fares.

If Vega can't beat Pascal in perf/w and/or outright performance, then we already know how it will do against Volta. Assuming Vega is 50% more efficient than RX 480, it'll have to consume 270-280 watts on average to match the 1080 TI. That won't bode well at all against Volta.