NVIDIA GeForce 20 Series (Volta) to be released later this year - GV100 announced

[Sweepr]

With this basically being same clocks and density and with all the Volta changes announced as deep learning specific.. Are we just expecting Gv104 and GV102 just to have slightly higher shaders at about the same clocks? Maybe AMD was on to something when they teased "Poor Volta".

A GTX 1080 Ti is already 72% faster than Fury X with 3584 SPs at 1440P according to TPU. Vega 10 with its 4096 SPs has zero chance against a gaming optimized 5120 SPs GV102. Too big of gap for them to close. NVIDIA doesn't even need to improve Pascal shader performance to dominate, which I'm sure they did as well.

 

[Arachnotronic]

Aha, just i thought. gv100 runs a bit slower in clocks. But has way more cuda cores. Yet, if you compare the theoretical FP32 throughput of both, gv100 is actually a bit slower than gp100 at the same clock speed.

Check your math?

 

[tamz_msc]

I believe the L0 instruction cache might not make it in GV104. This thing seems to exist purely to aid the different types of cores within each SM of GV100. Wont bode well for power efficiency in a consumer part where it's all FP32.

 

[William Gaatjes]

Check your math?

Check it per cuda core.
If you compare the amount of cores with the FP32 performance while maintaining same clock.
I could be wrong.
Then again, compare gp100 with 1080 ti at same clock for the idea what gv might do.

 

[Sweepr]

The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope. New Tensor Cores designed specifically for deep learning deliver up to 12x higher peak TFLOPs for training. With independent, parallel integer and floating point datapaths, the Volta SM is also much more efficient on workloads with a mix of computation and addressing calculations. Volta’s new independent thread scheduling capability enables finer-grain synchronization and cooperation between parallel threads. Finally, a new combined L1 Data Cache and Shared Memory subsystem significantly improves performance while also simplifying programming.

 

[Arachnotronic]

Yep, this is another Maxwell.

 

[Glo.]

Yep, this is another Maxwell.

Compared to GP10X(consumer Pascal) chips? Yes.
Compared to GP100 chip(HPC Pascal)? No.

We are talking about core for core, clock for clock situation. You have to bare in mind we are comparing 600mm2 vs 835mm2(!) die size GPUs.

 

[tamz_msc]

A GTX 1080 Ti is already 72% faster than Fury X with 3584 SPs at 1440P according to TPU. Vega 10 with its 4096 SPs has zero chance against a gaming optimized 5120 SPs GV102. Too big of gap for them to close. NVIDIA doesn't even need to improve Pascal shader performance to dominate, which I'm sure they did as well.

A full GV102 Volta is irrelevant to even the biggest Vega. Of course a 500mm^2 chip can only be competitive against a 600mm^2 chip up to a limit.

 

[Sweepr]

Yep, this is another Maxwell.

I wonder what's next for gamers. GV104 with 3584 CC, 240 TMUs, 128 ROPs and 16GB GDDR5X / GDDR6 @ 12-16 Gbps later this year? Better performance, more efficient than GP102 and probably smaller die size without all the compute stuff, months after Vega 10. Also Hynix hinting at a 384-bit GDDR6 product in early 2018, could be the GV102 monster.

Ps: OP updated!

 

[Arachnotronic]

I wonder what's next for gamers. GV104 with 3584 CC, 240 TMUs, 128 ROPs and 16GB GDDR5X / GDDR6 @ 12-16 Gbps later this year? Better performance, more efficient than GP102 and probably smaller die size without all the compute stuff, only a few months after Vega 10. Also Hynix hinting at a 384-bit GDDR6 product in early 2018, could be the GV102 monster.

The Hynix product definitely smells like a GV102. 768GB/sec of bandwidth is pretty darn close to what this HBM2-powered Volta gets, so I wouldn't be at all surprised to see it basically include 5120 Volta CUDA cores (Titan Xv) at higher frequencies than what the GV100 is getting.

Now we don't have all the info on Volta, just the changes to the SM. We still need to know what changes NVIDIA made to the "graphics" specific parts of the GPU; given that the changes were pretty modest in Pascal, I think they'll be substantial with GV104/GV102/etc.

Exciting times, I am glad I went with a G-Sync monitor.

 

[Head1985]

I wonder what's next for gamers. GV104 with 3584 CC, 240 TMUs, 128 ROPs and 16GB GDDR5X / GDDR6 @ 12-16 Gbps later this year? Better performance, more efficient than GP102 and probably smaller die size without all the compute stuff, months after Vega 10. Also Hynix hinting at a 384-bit GDDR6 product in early 2018, could be the GV102 monster.

Ps: OP updated!

128rops?lol no way.96 or 88 just like 1080TI.Or still 64 if they improve them.
16Gb Vram nope..12GB probably

 

[Sweepr]

Now we don't have all the info on Volta, just the changes to the SM. We still need to know what changes NVIDIA made to the "graphics" specific parts of the GPU; given that the changes were pretty modest in Pascal, I think they'll be substantial with GV104/GV102/etc.

If we believe the rumors from the usual chinese sources, they are. Going to be a long wait till gaming Volta's announcement.

 

[swilli89]

A GTX 1080 Ti is already 72% faster than Fury X with 3584 SPs at 1440P according to TPU. Vega 10 with its 4096 SPs has zero chance against a gaming optimized 5120 SPs GV102. Too big of gap for them to close. NVIDIA doesn't even need to improve Pascal shader performance to dominate, which I'm sure they did as well.

What are you talking about? You're comparing a 16nm 2017 product to a 28nm 2015 product.. Who brought up AMD?

Just our information we have now shows Volta doesn't really change much besides just upping the die size for each respective performance bracket which in turn will increase costs. Density remains about the same as well.

 

[tamz_msc]

I wonder what's next for gamers. GV104 with 3584 CC, 240 TMUs, 128 ROPs and 16GB GDDR5X / GDDR6 @ 12-16 Gbps later this year? Better performance, more efficient than GP102 and probably smaller die size without all the compute stuff, months after Vega 10. Also Hynix hinting at a 384-bit GDDR6 product in early 2018, could be the GV102 monster.

Ps: OP updated!

What rubbish, ROPs=Bus Width/4. 384 bit means 96 ROPs, just like GP102.

 

[swilli89]

Yep, this is another Maxwell.

Amazing you are able to draw that conclusion in regards to gaming performance based only on marketing material of a big data chip all from Nvidia itself.

 

[Guru]

Clearly a niche card, we'll have to wait and see how much they change it up for gaming, as I seriously doubt Nvidia can be competitive in Vulkan and DX12 without a significant redesign, which they've kind of done it seems, but its specifically optimized for computing and AI, so really doubt it will be useful for gaming.

That said AMD is clearly very late with Vega, it should have launched April and we are mid May and no word on a release date. The wait better be good, or as much as I like and root for AMD, they will loose big if its not at the very least competitive with the 1080TI.

 

[Arachnotronic]

Amazing you are able to draw that conclusion in regards to gaming performance based only on marketing material of a big data chip all from Nvidia itself.

Well, NVIDIA has earned its credibility. Pascal delivered what NVIDIA said, Maxwell delivered what NVIDIA said, and Kepler delivered what NVIDIA said.

All solid releases, and it has only gotten better with each generation.

 

[w3rd]

Those are server chips, which means very little to the consumer market, or games. So it all depends how nvidia cuts these down, to function at a consumer level.

And if they will hold some value, when Volta GeForce is eventually released for gamers.

 

[Glo.]

Well, NVIDIA has earned its credibility. Pascal delivered what NVIDIA said, Maxwell delivered what NVIDIA said, and Kepler delivered what NVIDIA said.

All solid releases, and it has only gotten better with each generation.

The problem is that today Nvidia demoed ONLY Deep Learning Stuff. I genuinely suggest not getting dragged out by Nvidia marketing, and brand perception/deception, especially when it goes for gaming.

 

[swilli89]

Well, NVIDIA has earned its credibility. Pascal delivered what NVIDIA said, Maxwell delivered what NVIDIA said, and Kepler delivered what NVIDIA said.

All solid releases, and it has only gotten better with each generation.

Lol quote of the year! Nvidia stretches the truth just like any other company that self promotes. Its very unscientific and a bit odd to take them at their word when they have had so many issues in the past. They straight up lied about Aysnc compute and Dx12 with Maxwell and to even a certain extent, Pascal.

 

[Bacon1]

Well, NVIDIA has earned its credibility.

Since when exactly?

Pascal delivered what NVIDIA said

Hmm I clearly remember them saying that 1080 was 2x faster than Titan X but don't remember that actually being the case.

Also that it was "cool as a cucumber" while @ 2100Mhz with the FE Fan.

Also "crazy overclockability" on the FE... which throttled out of the box... I seem to remember it being a worse OCer than the 980 Ti.

Maxwell still lacks Async Compute even though it was sold as supporting it. 970 had the memory 3.5 + .5...

 

[ElFenix]

almost the size of an image sensor in a "full frame" SLR (36x24 or 864mm^2)

 

[JDG1980]

A GTX 1080 Ti is already 72% faster than Fury X with 3584 SPs at 1440P according to TPU.

In large part that is because it runs at far higher clock speeds. TPU calculated that the reference card ran at average clock speed of 1777 MHz during testing. That compares to only 1050 MHz for Fury X.

One of the big questions with both Vega and Volta is where clock speeds will end up. That is going to play a major role in performance.

 

[JDG1980]

Yep, this is another Maxwell.

Note that all of the statements made by Nvidia about Volta so far are regarding GPGPU. We have no information on how perf/TFlop will be affected in ordinary gaming.

GV100 is said by Nvidia to do 15 TFlops FP32. That is better than what AMD will be offering with Vega MI25 (12.5 TFlops) but only a 20% advantage. If the GV100 solution costs several times as much then Vega may still gain market share by providing far better perf/$. Nvidia will still have an advantage from CUDA, and the new matrix multiplication hardware will appeal to some users but only those who are willing and able to do software rewrites.

 

[SlowSpyder]

5120 cores with 1455MHz clock, HBM2? Delicious.