I'm not sure how they are producing the big Volta chip - TSMC supports multiple different products per wafer and I believe to a higher level of complexity than other foundries - part of why they are only getting "one" per wafer may be due to having a low number of the big core per wafer and filling out the extra space with smaller products depending on what would be optimal use of the space.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
NVIDIA Volta with GDDR6 in early 2018?
- Thread starter Kaapstad
- Start date
I'm not sure how they are producing the big Volta chip - TSMC supports multiple different products per wafer and I believe to a higher level of complexity than other foundries - part of why they are only getting "one" per wafer may be due to having a low number of the big core per wafer and filling out the extra space with smaller products depending on what would be optimal use of the space.
I've never actually checked to be fair, but I've never heard/seen a wafer with different chips on it. Is that possible?
I'm not really familiar with nVidia's manufacturing but I've worked at places that use TSMC's shuttle services and looked into using them myself for personal projects for a power management circuit I use a lot and build with discretes (but way too high cost unless I had a guaranteed market).
Geek.com said:
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
In bullet point format:
- For the foreseeable future I plan on staying 1440p as I'd rather turn up all the eye candy than lower it somewhat at 4K (but perhaps Volta/?/? will do everything at 4K).
- Properly utilized CPU cores + threads: hopefully but it seems a reasonable long ways out (as does the MCM arch, kinda, so they might hit simultaneously) with the occasional and indeed increasing exception(s). It needs to start happening rather sooner than later though.
- Shuttle services: mighty interesting and I never knew that so thanks! There are some knowledgeable peeps in here, far more knowledgeable than me.
- 1 Volta chip per wafer is shockingly low! How many Pascal per wafer, does anyone know?
- For the foreseeable future I plan on staying 1440p as I'd rather turn up all the eye candy than lower it somewhat at 4K (but perhaps Volta/?/? will do everything at 4K).
- Properly utilized CPU cores + threads: hopefully but it seems a reasonable long ways out (as does the MCM arch, kinda, so they might hit simultaneously) with the occasional and indeed increasing exception(s). It needs to start happening rather sooner than later though.
- Shuttle services: mighty interesting and I never knew that so thanks! There are some knowledgeable peeps in here, far more knowledgeable than me.
- 1 Volta chip per wafer is shockingly low! How many Pascal per wafer, does anyone know?
Don't think there are exact figures out there - TSMC has described their 16nm FF+ production as being "mature" which usually means above 70% yields.
EDIT: Given that Pascal are potentially operating around 1.9-2GHz at a die sizes where you normally wouldn't aim for more than ~1.86GHz as the top end frequency on a 16nm process I'd say yields are probably quite good on Pascal.
The top Volta card on 12nm, 1080 Ti successor, should be ~50% faster than the 1080 Ti.
This is since the V100 card is 5120 FP32 cores, 2560 FP64 cores, and 640 Tensor cores. Much like with P100 vs the 1080 Ti, they'll likely strip out the FP64 and Tensor cores to make it much smaller but retain the FP32 cores.
This would give 5120 cores x 2 x 2 GHz = 20.5 Tflops, which is ~50% more than a 1080 Ti.
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
I would sleep very well at night and with a grin on my face with such a massive increase.
Agreed, they'll strip the FP64 and tensor cores from the GeForce line as you don't need those for gaming. I do wonder what they'll do with a putative Titan XV; I can see them leaving (quite a few) FP64 cores and (a few) tensor cores in - unless they won't and kinda force people that need those cores onto a Quadro, which wouldn't be the first time if memory serves me well.
I imagine they'll strip out FP64 and Tensor for the Titan too. But what they MIGHT do is turn on double-FP16 for the Titan line.
Since AMD are having 2xFP16 for Vega, but there's currently no 'normal person' use for that, they may use it as a super cheap differentiating factor for the Titan line vs the normal Geforce line.
An interesting consideration - hadn't thought of that.
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
An interesting idea indeed, I can see them doing that. Also ensures 100% they won't be undercutting their professional line.
Also worth noting the FP64 cores and the Tensor cores are very very likely tied together, in the sense they'll have their V100 cars with all 3 types of core, and then the Titan and 2080 Ti will be a different chip.
Now this chip could either just have the FP64 cores stripped out, or both FP64 and Tensor. I don't think it makes sense for them to have FP32 and Tensor on a chip they're basically just using for gaming and 'normal' professional workflows. And they only have to compete with 25 Tflops of FP16 performance in the 'cheap' space.
So I think it'll be only FP32, with the capability for 2xFP16 enabling in the drivers. Bearing in mind they're not getting a 'real' die shrink (I think the area reduction of 12nm is less than 5%, it's mainly power and performance improvements) so they can't waste space on their gaming cards if they want them to be comfortably faster than the GTX10 series.
If they did want to have an FP32 + Tensor (but not FP64) chip, that would mean they'd be making 3 large chips instead of the regular 2. Nvidia's R&D budget is obviously very large, and they could increase it even further with the amazing last financial year they've had, but I still don't think they'd make 3 different monolithic large-configurations. I imagine they'll wait for MCM chips before they offer many many different configurations of FP16, FP32, FP64, and Tensor.
Depends on the approach IIRC "12nm" gives around 10% performance increase or 25% power reduction as your normal two directions but through a mixture of techniques can potentially get a ~33% performance increase or a bigger than nominal area reduction (not sure actual figures for it as its mostly covered in buzzwords) using other approaches.
EDIT: "12nm also offers a 20% area reduction with 6T Libraries versus 7.5T or 9T." but I'm not sure what that 20% is in context to - it might be 16nm FF rather than 16nm FF+.
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
Who knows... (wish I did)
Soldato
- Joined
- 19 Oct 2008
- Posts
- 6,060
Think they were second last time, 1080 came out first or at least there was availability of 1080's to buy before the 1070 - think it was another month or so for the 1070's.
1080's been with us for over a year now.Shame AMD have had a missed generation - even today still nothing out there to compete.
Come 2019 a Volta Ti will do everything at 4K but it will be a fairly large chip and will be expensive.
Soldato
- Joined
- 19 Oct 2008
- Posts
- 6,060
Yeah, not like the cheap Titan P
Unfortunately Volta is nowhere near powerful enough to max out 4K.
Caporegime
- Joined
- 30 Jul 2013
- Posts
- 30,017
Your version of 'max out' usually includes Super Sampling it to 8k though Kaap
1080Ti/Pascal Xp are nearly there, one more generational leap should do it.
1080Ti/Pascal Xp are nearly there, one more generational leap should do it.
For 3440x1440, I am more than happy with the 1080Ti and copes with pretty much all settings up. The occasional poorly optimised game does need some settings toned down but for the majority of games, they run very very well and look amazing.