[anandtech] NVIDIA SIGGRAPH 2018 Keynote -RTX GPUs

[PeterScott]

https://www.anandtech.com/show/13215/nvidia-siggraph-2018-keynote-live-blog-4pm-pacific

Ongoing right now with new breakthrough ray-tracing card.

Any guesses on how many thousands of dollars?

 

[jpiniero]

Holy crap, 754 mm2? That has to be GT102.

 

[PeterScott]

Seeing the full specs, I am thinking its GT100, industry insider card for $10000+

 

[jpiniero]

https://videocardz.com/77260/nvidia-announces-quadro-rtx-6000-ray-tracing-gpu

Says top model is 10k. Also 6300 and 2300 models.

 

[PeterScott]

I think I prefer the eventual consumer GPUs to skip the RT HW. It takes away from die area better served with more CUDA units, and probably won't provide enough RT power to warrant having it on.

 

[alcoholbob]

Are these RT cores just rebadged tensor cores?

 

[PeterScott]

Are these RT cores just rebadged tensor cores?

RT Cores and Tensor Cores looks like different things:
https://images.anandtech.com/doci/13214/Pascal_vs_Turing.jpg

 

[JDG1980]

I wonder if the RTX 5000 is a die-salvaged part, or a smaller chip. It has a narrower memory bus (256-bit) so it could potentially be the "4"-series chip.

Since we know that the big chip has 4608 shaders and does 16 TFlops in single-precision, we can calculate that its clock speeds are around 1736 MHz. That's higher than the boost clocks on Nvidia's stock GP102 and GP100 cards, but not as high as some AIB gaming cards can go. It's probably a good assumption that Turing can clock somewhat higher than Pascal, but not by a huge margin.

 

[alcoholbob]

Hmm so top end Turing will have less shader cores than Volta, but will have these new RT cores. Wonder if how long before we'll have to wait to see games use these RT cores.

 

[HurleyBird]

Holy crap, 754 mm2? That has to be GT102.

Seeing the full specs, I am thinking its GT100, industry insider card for $10000+

Whatever it is, the timing versus Volta and the very similar die size seems a bit odd. One would also expect GT100 to carry HBM2. And one wouldn't expect a replacement for GV100 to come in now at 12nm when 7nm is so close. At the same time, it's hard to think of a theoretical GT102 with such a huge die.

It's a real headteacher. It could be the equivalent of GT100, or GT102, or even something entirely different. Whatever the case, I'm sure there's a *very* interesting story behind it. Was the part originally meant for 7nm? Were both Volta and Turning parts originally meant to come in much earlier than they did? One day, perhaps we'll know.

 

[alcoholbob]

Whatever it is, the timing versus Volta and the very similar die size seems a bit odd. One would also expect GT100 to carry HBM2. And one wouldn't expect a replacement for GV100 to come in now at 12nm when 7nm is so close. At the same time, it's hard to think of a theoretical GT102 with such a huge die.

It's a real headteacher. It could be the equivalent of GT100, or GT102, or even something entirely different. Whatever the case, I'm sure there's a *very* interesting story behind it. Was the part originally meant for 7nm? Were both Volta and Turning parts originally meant to come in much earlier than they did? One day, perhaps we'll know.

It's probably yields. Something GV100 sized would be over 500mm2 on 7nm. They probably figured its just not profitable enough yet and let AMD jump on the node first with an expensive mid-size die (Vega 20 will be ~300mm2 and will be a frontier edition only style card).

 

[tamz_msc]

Based on the Turing GPU die image and the specifications of GV100 my guess for the layout of the "big-Turing" GPU is

• Six GPCs, each having
  
  • six TPCs(2 SMs for each TPC)
  • 12 SMs
• 72 Turing SMs, each having
  • 64 FP32 units
  • ? INT units
  • ? FP64 units
  • 8 Tensor cores
  • (probably) Four texture units

So compared with Volta, NVIDIA has trimmed 2 SMs in each GPC to make room for the RT cores. The numbers and the image suggest that this is a fully enabled chip, unlike GV100 which had 4 out of 84 SMs disabled.

I'm also inclined to believe that the "small-Turing" chip featured in the Quadro RTX 5000 is also a fully enabled chip.

 

[Gideon]

One would also expect GT100 to carry HBM2. And one wouldn't expect a replacement for GV100 to come in now at 12nm when 7nm is so close.

I'm pretty sure it's not intended as a replacement. Nvidia simply has enough funds to design a different GPU for Quadro, going for the lucrative movie-industry, and probably a Volta replacement on 7nm for other GPGPU tasks.

 

[SPBHM]

Hmm so top end Turing will have less shader cores than Volta, but will have these new RT cores. Wonder if how long before we'll have to wait to see games use these RT cores.

it would probably have to be a sponsored title with some effect thrown on top after the game is done, like the GPU PhysX stuff was... because, I don't see this stuff scaling decently to mid range and consoles anytime soon, so it's kind of not really viable to make games with that in mind.

 

[jpiniero]

It's a real headteacher. It could be the equivalent of GT100, or GT102, or even something entirely different. Whatever the case, I'm sure there's a *very* interesting story behind it. Was the part originally meant for 7nm? Were both Volta and Turning parts originally meant to come in much earlier than they did? One day, perhaps we'll know.

There were rumors that nVidia was getting ready to launch new gaming chips right before mining took off. Would that have been GT104? Possibly. But now it looks like you are getting both GT102 and GT104 at the same time.

 

[Guru]

Well it was obvious that we won't get any new desktop parts in 2018, not to toot my horn, but told you so! It seems these new Turning RTX parts are late Q4 2018 with one new feature that is specific cores for ray-tracing.

|We don't know if these are 12nm yet, though they seem to be, as 7nm is exclusive to AMD first and we won't be getting it on Nvidia before 2019. So it seems 12nm Turning RTX graphics in late Q4 2018, probably 7nm very late Q1 or Q2 for desktop parts.

The die size is also way too big for these graphics to be able to translate well into desktop, so it seems to me the gaming parts will stick to the tried and true cores that we have right now, with obvious changes to improve DX12 and Vulkan performance, as Nvidia is lagging badly here, though they did bet really well again on DX11 with their 1000 series GPU's.

 

[PeterScott]

Well it was obvious that we won't get any new desktop parts in 2018, not to toot my horn, but told you so! It seems these new Turning RTX parts are late Q4 2018 with one new feature that is specific cores for ray-tracing.

I disagree. This Video from NVidia is full of easter eggs for a RTX 2080 Reveal on Aug 20:

https://www.youtube.com/watch?v=F7ElMOiAOBI

 

[LTC8K6]

WCCF link

https://wccftech.com/nvidia-quadro-rtx-turing-gpu-samsung-gddr6-memory/

Announcement ⋮ Hardware ⋮ Report
Samsung Powers NVIDIA Quadro RTX Graphics Cards With 16Gb GDDR6 Memory – Up To 14 Gbps Speeds, 672 GB/s Bandwidth For Professional Cards

 

[tviceman]

Well it was obvious that we won't get any new desktop parts in 2018, not to toot my horn, but told you so! It seems these new Turning RTX parts are late Q4 2018 with one new feature that is specific cores for ray-tracing.

Do you always jump the gun and enjoy being a debby downer? The gaming event is August 20th. Hold your horses.

 

[Cloudfire777]

RTX 5000 = GTX 2080
RTX 6000 = GTX 2080 Ti (Minus the compute performance)

There we have it.

And obviously the Geforce card will launch within September - October somewhere so shut up with your BS, Guru

 

[PeterScott]

Whatever it is, the timing versus Volta and the very similar die size seems a bit odd. One would also expect GT100 to carry HBM2. And one wouldn't expect a replacement for GV100 to come in now at 12nm when 7nm is so close. At the same time, it's hard to think of a theoretical GT102 with such a huge die.

It's a real headteacher. It could be the equivalent of GT100, or GT102, or even something entirely different. Whatever the case, I'm sure there's a *very* interesting story behind it. Was the part originally meant for 7nm? Were both Volta and Turning parts originally meant to come in much earlier than they did? One day, perhaps we'll know.

Not really a headscratcher. Just an aggressive move by NVidia, which has a pretty aggressive CEO.

This replaces GV100 in every way. It has more memory bandwidth, more compute performance, more Tensor OP performance, and bonus Ray tracing performance on top, on a slightly smaller dies, with less expensive memory technology.

It's Win-Win-Win-Win For RTX 8000(GT100) vs GV100. Why does GV100 need to exist anymore?

Sure it's only been a year, but NVidia has the money and talent to do this kind of high end part to win yet another computational market.

 

[PeterScott]

I'm also inclined to believe that the "small-Turing" chip featured in the Quadro RTX 5000 is also a fully enabled chip.

I believe the "small-Turing" is just a partially disabled "big-Turing". RTX-8000,6000,5000 are all on the same die.

 

[jpiniero]

I believe the "small-Turing" is just a partially disabled "big-Turing". RTX-8000,6000,5000 are all on the same die.

The 5000 has 50% less cores. That would be a pretty large cut. Possible, but I still think the 5000 is full GT104. I think what is throwing people off is that Quadro usually doesn't "launch" first.

I could see 2080/2070/2060 being GT104 while anything lower is Pascal rebrands.

 

[Headfoot]

GV100 will still have more NVLink capacity. Looks like Turing v Volta is an HPC vs Workstation card breakdown. I personally think nVidia has enough $$ and potential revenue to customize the same base level architecture for each niche a bit differently. Volta continues to be the hyperscale chip of choice?

 

[PeterScott]

The 5000 has 50% less cores. That would be a pretty large cut. Possible, but I still think the 5000 is full GT104. I think what is throwing people off is that Quadro usually doesn't "launch" first.

I could see 2080/2070/2060 being GT104 while anything lower is Pascal rebrands.

Math Nitpick: It's ~33% less cores, but yes, still a big cut.

Whats throwing me, is that if RTX 5000 is GT104 destined for consumer RTX 2080/2070, why would it still be full of Tensor cores? I know they will likely disable them, but it seems like a waste of silicon for what will be a volume consumer part, and probably an insignificant volume of Quadro parts. I am more inclined to believe it's just part cut for RTX 5000. But neither is a slam dunk IMO.

Rebrands I agree are very likely for the GTX part of the lineup. I already suggested before than GTX 1060 would make a nice GTX 2050/2050Ti. I definitely don't seen an RTX 2050 coming.

At least it's only 6 more days of waiting.