NVIDIA GeForce RTX 2080 Ti features 4352 CUDA cores, and new 20 series Card From Gigabyte pictured!

[jasonvp]

20 minutes ago, Taf the Ghost said:

What it looks like is that Nvidia has moved the Titan up-market, again, so they can sell off their supply of partially defective Quadros into a somewhat larger market.

I'm still going to bet against that.  I don't think the Titan V was an indicator of where they're taking the "Titan" brand name; rather it was a one-off.

 

Bear in mind we haven't seen a Tesla card yet, which should also be based on the TU100.  I kind of expect the "partially defective" TU100s to end up in the Quadros, while the binned ones end up on the yet-to-be-announced Teslas.

 

That'll leave an unlocked TU102 for the Titan.

That's my speculation, anyway.

[jasonvp]

1 minute ago, leadeater said:

without an announced Tesla card with HBM

Dammit: we were posting at the same time. :-D

[Taf the Ghost]

1 minute ago, leadeater said:

I'm still a little dubious that the RTX 8000 is actually using the biggest Turing die, without an announced Tesla card with HBM I'd say be cautious in assuming any products are actually using a TU100. Not that it's a great indicator but the given Turing die specs it's which are smaller than Volta and that it's using the same process there could be a different die configuration for the Tesla cards, not that it matters greatly if that is the case because if true no gaming card will be based off of it.

I imagine the Tesla stays as Volta (so they can get their money back) until they launch the first 7nm Compute card next year. I think the next real design we see from Nvidia will be the GA100 in mid-2019 on TSMC's 7nm.

[leadeater]

1 minute ago, Taf the Ghost said:

I imagine the Tesla stays as Volta (so they can get their money back) until they launch the first 7nm Compute card next year. I think the next real design we see from Nvidia will be the GA100 in mid-2019 on TSMC's 7nm.

I don't think that is a problem, they have sold a heck of a lot of V100's and we're constantly unable to buy them. Even getting one is hard, had one research team ask for 10 and they ended up having to buy them from Amazon and also drop down to some Pascal cards as well.

 

Delaying it a bit longer makes sense but I'm still expecting to see a Turing Tesla card, maybe March next year ish.

[jasonvp]

Just now, leadeater said:

Delaying it a bit longer makes sense but I'm still expecting to see a Turing Tesla card, maybe March next year ish.

Yep.  I'm going to throw out a guess that: Titan (Turing) is announced in 2-3 months, ready for Christmas.  That'll follow NVidia's previous pattern of sitting on the Titan after the new GPU is launched for about 2-3 months.

 

Tesla follows on its heels.

[Taf the Ghost]

2 minutes ago, leadeater said:

I don't think that is a problem, they have sold a heck of a lot of V100's and we're constantly unable to buy them. Even getting one is hard, had one research team ask for 10 and they ended up having to buy them from Amazon and also drop down to some Pascal cards as well.

 

Delaying it a bit longer makes sense but I'm still expecting to see a Turing Tesla card, maybe March next year ish.

Would make sense. It's also extremely possible Nvidia has been hedging on the Node timing, so it's not decided which is going to happen yet.

[leadeater]

7 minutes ago, Taf the Ghost said:

Would make sense. It's also extremely possible Nvidia has been hedging on the Node timing, so it's not decided which is going to happen yet.

Could be a 7nm milestone target that is the deciding factor, if TSMC meets certain milestones then no Turing Tesla and straight to 7nm. I imagine it wouldn't be hard to push out a Turing Tesla design if required, of course that also depends if there is even demand for the unique features of Turing as well, in certain ways it's lower spec than Volta but without benchmarks who knows. Maybe this is the start of the Intel design method, consumer on new arch first then followed by enterprise on refined larger arch.

[Taf the Ghost]

8 minutes ago, leadeater said:

I don't think that is a problem, they have sold a heck of a lot of V100's and we're constantly unable to buy them. Even getting one is hard, had one research team ask for 10 and they ended up having to buy them from Amazon and also drop down to some Pascal cards as well.

 

Delaying it a bit longer makes sense but I'm still expecting to see a Turing Tesla card, maybe March next year ish.

Other thing is HBM2. While the TU100 adds some fun new stuff, does HBM2 work better for the types of workloads that the Tesla product line is mostly focused at?

 

I agree there's another Tesla on the way, but do they move to 7nm with it? I guess we'll know by about March.

[Taf the Ghost]

3 minutes ago, leadeater said:

Could be a 7nm milestone target that is the deciding factor, if TSMC meets certain milestones then no Turing Tesla and straight to 7nm. I imagine it wouldn't be hard to push out a Turing Tesla design if required, of course that also depends if there is even demand for the unique features of Turing as well, in certain ways it's lower spec than Volta but without benchmarks who knows. Maybe this is the start of the Intel design method, consumer on new arch first then followed by enterprise on refined larger arch.

Considering they're having to add unique units to keep pushing the Gaming Graphics, I think it's safe to assume we're going to get a full split between Gaming & HPC pretty soon.

[pas008]

1 hour ago, GoldenLag said:

So dead due to dx12?

 

SLI and Crossfire is dead in the way it scales poorly and creates extra issues to worry about. Not to mention the minority who buy high end cards and the minority of the minority who runs them im SLI.

 

Polaris was meant to be good in SLI, but you dont see people running dual 480s or 580s do you?

Are Dx 12 titles running sli? Ok then

Why do you think they have changed the bridge twice in a few yrs? Oh cause it's obviously dead

 

 

 

[leadeater]

1 minute ago, Taf the Ghost said:

Other thing is HBM2. While the TU100 adds some fun new stuff, does HBM2 work better for the types of workloads that the Tesla product line is mostly focused at?

Memory bandwidth is nearly 300GB/s faster with HBM2 on Volta. Higher clocked GDDR6 will close that gap, HBM as a technology is still much more capable of increasing bandwidth by large margins though.

[S w a t s o n]

45 minutes ago, Taf the Ghost said:

 They've pushed the Ti up a die size as well. That's what happened with Titan X(p) -> 1080 Ti -> Titan XP. 1080 Ti was just a he time frame they'd normally launch another generation.

xx80Ti was always the 100/102 die size

[Taf the Ghost]

5 minutes ago, leadeater said:

Memory bandwidth is nearly 300GB/s faster with HBM2 on Volta. Higher clocked GDDR6 will close that gap, HBM as a technology is still much more capable of increasing bandwidth by large margins though.

HBM3 will be on 7nm, so it's probably 2020 before we see it in volume. (It'll ramp up with DDR5, I imagine.) If HBM3 was near, I'd expect the 7nm Tesla with HBM3 to drop next year. That's probably the 2021 Tesla. Actually, you can probably book that device right now, though maybe late 2020.

 

Other aspect is that Nvidia probably doesn't want to compete against themselves too much, which leaves a lot of things up in the air.

[schwellmo92]

1 hour ago, Taf the Ghost said:

Other thing is HBM2. While the TU100 adds some fun new stuff, does HBM2 work better for the types of workloads that the Tesla product line is mostly focused at?

 

I agree there's another Tesla on the way, but do they move to 7nm with it? I guess we'll know by about March.

HBM(2) is basically better at everything, including emptying your wallet.

[schwellmo92]

51 minutes ago, Taf the Ghost said:

HBM3 will be on 7nm, so it's probably 2020 before we see it in volume. (It'll ramp up with DDR5, I imagine.) If HBM3 was near, I'd expect the 7nm Tesla with HBM3 to drop next year. That's probably the 2021 Tesla. Actually, you can probably book that device right now, though maybe late 2020.

 

Other aspect is that Nvidia probably doesn't want to compete against themselves too much, which leaves a lot of things up in the air.

We don’t even need HBM3. HBM2 > GDDR6 and even GDDR7. We have barely scratched the surface in terms of what it can deliver.

 

As an example, Vega has 2x 4GB 1.6Gbps 1.2v stacks for effective bandwidth of 400GB/s. They could swap that out to 4x 2GB 2.4Gbps 1.2v stacks for an effective bandwidth of 1200GB/s, and then add even more stacks.

[Taf the Ghost]

1 minute ago, schwellmo92 said:

HBM(2) is basically better at everything, including emptying your wallet.

I was more talking in terms of significant technical aspects of the HPC market. The type of workloads that HBM vs GDDR might end up being something stupid like 50% faster for HBM because of certain differences.

 

But I do like the joke.

[Princess Luna]

Can't wait to compare the RTX 2080 Ti with the TITAN V personally

[Tiber1337]

I hope these will come into laptops soon

[schwellmo92]

18 minutes ago, Taf the Ghost said:

I was more talking in terms of significant technical aspects of the HPC market. The type of workloads that HBM vs GDDR might end up being something stupid like 50% faster for HBM because of certain differences.

 

But I do like the joke.

I’m only speculating here but there may be workloads that run better on HBM. For example if there’s 1GB of data in VRAM that you need to access, let’s say you’re using GDDR6 and that entire 1GB is on a single chip, you’d only be getting 14Gbps x 32bit bus-width = 56GB/s of memory bandwidth, whereas if that 1GB was sitting on a HBM2 stack at 2.4Gbps x 1024bit bus-width = 307GB/s. I would assume the GDDR6 solution would have some smarts to split up data across as many chips as possible so I may be hypothesising an edge case.

[Taf the Ghost]

2 minutes ago, schwellmo92 said:

I’m only speculating here but there may be workloads that run better on HBM. For example if there’s 1GB of data in VRAM that you need to access, let’s say you’re using GDDR6 and that entire 1GB is on a single chip, you’d only be getting 14Gbps x 32bit bus-width = 56GB/s of memory bandwidth, whereas if that 1GB was sitting on a HBM2 stack at 2.4Gbps x 1024bit bus-width = 307GB/s. I would assume the GDDR6 solution would have some smarts to split up data across as many chips as possible so I may be hypothesising an edge case.

I was thinking more in terms of latency, as HBM is just physically, significantly closer to the die. The latency, combined with clocks, is literally the gaming performance difference between Ryzen & Coffee Lake, so we deal with these issues all of the time. And many server applications operate on pretty similar principles. Being HPC and Compute stuff doesn't normally crop up in much public data, for performance analysis, I don't really know if it matters.

 

It might, depending on the workload. Nvidia knows, and it'll dictate a lot of what they do.

[Amazonsucks]

5 hours ago, Taf the Ghost said:

I was thinking more in terms of latency, as HBM is just physically, significantly closer to the die. The latency, combined with clocks, is literally the gaming performance difference between Ryzen & Coffee Lake, so we deal with these issues all of the time. And many server applications operate on pretty similar principles. Being HPC and Compute stuff doesn't normally crop up in much public data, for performance analysis, I don't really know if it matters.

 

It might, depending on the workload. Nvidia knows, and it'll dictate a lot of what they do.

Matters for certain workloads. Similarly to how 3DS DDR4 DRAM which is stacked DRAM using TSVs on DIMMs improves performance as well as doubling the capacity per DIMM.

 

https://blogs.synopsys.com/committedtomemory/2015/07/08/samsung-ddr4-3ds-3d-stacked-dimms-using-through-silicon-vias-tsv/

[i_build_nanosuits]

11 hours ago, Tiber1337 said:

I hope these will come into laptops soon

Such a huge chip would melt on your knees in a laptop..

But i get what you mean, a smaller mobile version of this will have epic performance.

[jasonvp]

I've been thinking through a few theories regarding the latest round of leaks.  I considered three different possibilities:

 

1.  The AIB folks all screwed up somehow.  That means PNY, EVGA, MSI, et al all goofed in or around the same time, and images of boxes, cards, etc got out before Monday.

 

2.  The AIB folks were asked to leak the images and then quickly get rid of them.  All in the hopes of generating buzz (it worked!).  And all sponsored by Team Green themselves.

 

3.  It's a ruse.  Team Green gave the AIBs a list of phony specs and info and asked them to push that data out, to throw folks' expectations off a bit, leading to Gamescom.

 

I suspect the truth lies somewhere between 2 and 3.  I don't, for a moment, actually believe that each of those aforementioned AIBs "just goofed" at the same time.  Misinformation is a thing, though, and it's a useful thing prior to a big announcement.

[i_build_nanosuits]

1 hour ago, jasonvp said:

3.  It's a ruse.  Team Green gave the AIBs a list of phony specs and info and asked them to push that data out, to throw folks' expectations off a bit, leading to Gamescom.

so you're saying it could potentially be even greater than those specs?

[TheRandomness]

On 17/08/2018 at 10:25 PM, Locutus494 said:

That's a Quadro, not a GeForce. See the NVLink connector at the top?

Then you realise the NVlink connector on there is backwards, and there’s only one of them.