itsmydamnation
Veteran
Actually its completely the other way around, Design costs are skyrocketing. Vega is competing across consumer gfx, pro gfx and compute that how it goes up against GP100,102,104. As disappointing as it is Gaming is its weakest link at this stage.
NV is taking huge bets with some of those chips. AMD hasn't had the financial resources to take those risks and in a way they dont even have to, something like p47 could do very well and AMD took almost 0 risk to do it. How much risk is NV taking with GV100? Some start up makes a TPUv2 class asic and then what(not saying this will happen)?
Some not-really-a-startup-anymore is making a TPUv2-class ASIC called TPUv2 and GV100's prospects still seem OK, so… I dunno, I think NVIDIA will be fine, at least for now.
TPUv2 isn't really available for everyone though, even if you can rent time on it from Googles cloud, you can't actually get your own machines with it AFAIK. That hypothetical startup would obviously be selling it on open markets.
NV has the advantage of being a big, reliable, established player with a lot of software/API support coming automatically with doing business with them. Both directly from NV themselves, and from all the other commercial and open-source developers using NV hardware.
They're a quite safe bet, which undoubtedly helps them a lot, kind of being the IBM of their times.
Now, the fortunes of companies can always change of course (in no small part depending on how arrogant and complacent they become), and NV isn't exactly a super huge corp either, but for now they seem to have worked out a good niche for themselves where nobody is really in a position to threaten them, due to their being well established, and the advanced, specialized nature of their products.
I wouldn't worry too much about Google beating nVidia at their own game, also given their inconistent quality of their products/ideeas ( golang, Google Plus, the latest `feminist` scandal)
Perhaps these TPU fullfil a specific need for Google, hence their mass production
While I of course agree that design costs are skyrocketing, every company tries to keep those costs in check. Looking at mentioned Nvidia, they have SM-building blocks which they re-use to a large degree. It's not like they started GV100 from scratch. Maybe not even the Tensor Cores, but probably they are fairly new. But re-using functional blocks obviously is much cheaper than re-designing everything anew. Especially if you had the foresight AMD is now installing to make the IP blocks more and more independent, so you could scale and renew different parts of the chip as required with multiple teams working asynchronously.
That said, yes, there is the risk, that Googles TPUv2 is taking a big bite out of Volta's cake, but the risk of a newly formed startup achieving that is far less I would think. For one, good and able engineers do not grow on trees, and for another, startup possibly don't have the ability to throw that much ressources at a problem as google. And with the TPU (v1 more so than v2 it seems) being a VERY highly specialized processor, chances are low that startups would take such a massive risk.
Plus what others have said.
I guess his point was that it definitely does not compete with GP102. As it stands its GP102 is 30% ahead, plus it also costs way more. The products are not in the same market tier, therefore not competing.
leoneazzurro
Regular
It is competing with the PRO version of the GP102, this is the sense. And it competes quite well, indeed, with the Vega FE often outperforming the P6000 for a fraction of the price.
Yes, but Titan XP now has PRO drivers as well so it is performing the same as P6000 so there's really no need to pull P6000 price point into the discussion. And yes I agree we have AMD to thank for to jerking NV into reacting.
Well that actually reinforces the "is competing with GP102" argument since Titan XP price is closer to Vega FE's. I forgot about the non-gaming use cases when answering him.
DavidGraham
Veteran
WX line of Vega is underclocked to 1200Mhz, So they will fall behind Quadros in performance if you need driver certification.
Vega FE line is directed toward developers who game and design traditional games and VR, it may tie TitanXp in some productivity apps, but it falls behind it significantly in gaming/VR performance. So it's not really competing in that use case.
Vega FE line is directed toward developers who game and design traditional games and VR, it may tie TitanXp in some productivity apps, but it falls behind it significantly in gaming/VR performance. So it's not really competing in that use case.
leoneazzurro
Regular
OK, that's also a point but it was only an example for better understanding that yes, in certain markets Vega IS competing quite well with GP102.
leoneazzurro
Regular
That's the base clock. Boost clock is 1,5 Ghz and being the Vega FE quite limited in reaching the max boost clock of 1,6 GHz (practically hovering at 1,45 Ghz at least on the air version) by thermal throttling, I wouldn't expect a big difference in actual performance between the two.
To my understanding Titan Xp still doesn't have "full blown Pro-driver" but rather very specific set of optimizations to most popular (most benchmarked?) prosoftware used
They included the Pro application certified optimizations to the Geforce drivers instead of only having those optimizations in the Quadro/Tesla drivers. Geforce drivers lack the complete certification found in Quadro drivers.
There are still extra features there for Quadro yes. And so there are extra features in WX 9100 that are not there in Vega FE.
itsmydamnation
Veteran
I dont agree, that's a hind sight 20/20 position, to lead a market you have to start your planning and execution before anyone else does.
When you start 3-4 years out designing a 800mm chip on a custom leading edge process who's targeting an emerging market that may or may not be their in the future and your product many or may not be competitive ( because you have so much silicon dedicated to other functions), its of higher risk. Compare it to GP104, that thing basically doesn't have a single feature that isn't main stream today, it even lacks features of low end parts, compare the risk profile of the two.
Nv has the financial position to take a 200-500 million hit if GV100 completely floops, AMD of the last couple of years has not had that luxury, thus you see a more middle of the road approach.
Anarchist4000
Veteran
Middle of the road, but more versatile and AMD don't segment the market as much. Specialized chips may make less sense in a competitive market. Tensors aren't difficult to accelerate, but TPU may not exist with more reasonable prices.
Wasn't it 2.5B for Pascal development? If tensor cores prove useless in the face of TPUs along with that monster die size and the R&D for it that may not be far off. Hard to see any R&D as a loss though as they could make 800mm2 parts in the future.
