NVIDIA GeForce GTX 1180 Bare PCB Pictured

[FordGT90Concept]

Why not? Not everyone can afford a $2000+ card but they still want RSX. For all we know, GV102 could have two stacks of HBM2 and GV104 could have one forming a complete, professional product stack. Deep learning and neural processing require very different silicon from display tasks. It makes sense to divorce the two at some point.

 

[efikkan]

Why not? Not everyone can afford a $2000+ card but they still want RSX. For all we know, GV102 could have two stacks of HBM2 and GV104 could have one forming a complete, professional product stack. Deep learning and neural processing require very different silicon from display tasks. It makes sense to divorce the two at some point.

We've all seen the prototype board and this new "1180" board, one with 384-bit GDDR6 and one with 256-bit GDDR6. No HBM on either of these.

 

[FordGT90Concept]

This PCB could easily be for a Turing or Ampere GPU. What are gamers going to do with tensor cores?

 

[efikkan]

This PCB could easily be for a Turing or Ampere GPU.

I don't care about the speculation about name, wccftech, videocardz & co changes their mind every other week about that stuff.

What I do know is that "GV102" and "GV104" was taped out before last summer, and that Nvidia is about to release a lineup of new GeForce cards which will be "Volta based". I don't know what product names and "architecture" names Nvidia will use for branding, and frankly I don't care.

What are gamers going to do with tensor cores?

Many assume the RTX technology uses these. RTX is primarily designed for consumers.
Anyway, it's very common that Nvidia reduce or eliminate features on their lower models. Like GP102 reduces fp64 and removes fp16 support vs. GP102, along with a different memory controller, etc.

 

[FordGT90Concept]

Turing may have been branding for mining cards which may or may not have been scraped. Ampere is an architecture with the pilot GPU bearing the GA100 model number.

You may be right that Volta-based GPUs are being announced in a few weeks and, excepting the GV100 (with Tensor cores) are just 12nm rehashes of Pascal. We'll find out in due time.

 

[efikkan]

Turing may have been branding for mining cards which may or may not have been scraped. Ampere is an architecture with the pilot GPU bearing the GA100 model number.

You may be right that Volta-based GPUs are being announced in a few weeks and, excepting the GV100 (with Tensor cores) are just 12nm rehashes of Pascal. We'll find out in due time.

GV100 is not a rehash of Pascal. Pascal was the stepping stone between Maxwell and Volta, introduced due to the delays of Volta. Pascal is Maxwell with some of the design features from Volta.

I recommend reading the White paper from Nvidia, or some of the articles from Anandtech.
Especially page 10-12 from the white paper is an interesting read, some snippets:

Similar to Pascal GP100, the GV100 SM incorporates 64 FP32 cores and 32 FP64 cores per SM. However, the GV100 SM uses a new partitioning method to improve SM utilization and overall performance.

Integration within the shared memory block ensures the Volta GV100 L1 cache has much lower latency and higher bandwidth than the L1 caches in past NVIDIA GPUs.

Unlike Pascal GPUs, which could not execute FP32 and INT32 instructions simultaneously, the Volta GV100 SM includes separate FP32 and INT32 cores, allowing simultaneous execution of FP32 and INT32 operations at full throughput, while also increasing instruction issue throughput.

And so much more. I highly recommend reading it, since it's a good insight into what's going to trickle down to the consumers.

 

[jabbadap]

GV100 is not a rehash of Pascal. Pascal was the stepping stone between Maxwell and Volta, introduced due to the delays of Volta. Pascal is Maxwell with some of the design features from Volta.

I recommend reading the White paper from Nvidia, or some of the articles from Anandtech.
Especially page 10-12 from the white paper is an interesting read, some snippets:

And so much more. I highly recommend reading it, since it's a good insight into what's going to trickle down to the consumers.

Volta was not exactly delayed, it had target to meet with Summit Supercomputer(Was always to be 16nm product, as technically it still is. 12nm FFN is closer to 16nm FF+ rather than 12nm FFC). TSMC 20nm manufacturing process were not suitable for GPUs, thus nvidia was forced to strip FP64 out of Maxwell and wait for 16nm to put it back on Pascal. You could say gp100 was very short lived King for HPC, and I would even say that Nvidia would have wanted to delay Volta just to sell gp100 products for a longer time... But yeah Volta is vastly different arch than Pascal/Maxwell.

This PCB could easily be for a Turing or Ampere GPU. What are gamers going to do with tensor cores?

Gameworks Denoisers might have use for Tensor cores. I don't think RTX itself needs them though. And like Efikkan said: Volta can do int32 and fp32 math simultaneously, while Pascal or earlier ones can't(i.e. using int8 to inference denoiser and fp32/fp16 for rasterizing). Nvidia might allow full speed fp16 too, now that expensive pro cards have Tensor cores do such tasks.

 

[efikkan]

Volta was not exactly delayed, it had target to meet with Summit Supercomputer(Was always to be 16nm product, as technically it still is. 12nm FFN is closer to 16nm FF+ rather than 12nm FFC). TSMC 20nm manufacturing process were not suitable for GPUs, thus nvidia was forced to strip FP64 out of Maxwell and wait for 16nm to put it back on Pascal. You could say gp100 was very short lived King for HPC, and I would even say that Nvidia would have wanted to delay Volta just to sell gp100 products for a longer time... But yeah Volta is vastly different arch than Pascal/Maxwell.

If you look at past roadmaps, Volta was scheduled for 2016, but was postponed for late 2017 and Pascal shoved in-between.
But as you are saying, TSMC "12nm FFN" is not any denser than 16nm at all, it's just a refinement.