Review: Nvidia Turing architecture - GeForce RTX 2080 and 2080 Ti

#5584549 Posted on: 09/14/2018 05:41 PM

---

I can do ~2100 with my 1080 if I want to, and running at 1930'ish out of the box in the games that I play.

2200 is unimpressive at all (but expected, since this is basically the same node, just slightly refined), and it will probably consume HUGE amounts of power at that frequency, given that Turing is a bigger chip with all those fancy semi-useless raytracing corez.

Really curious what the benchmarks will show compared to SIMILARLY PRICED Pascal:
So 2070 results vs 1080 Ti, and not vs 1070.

#5584550 Posted on: 09/14/2018 05:44 PM

---

Can't wait to see testing results. I have this sneaking suspicion that the improvements are not that great unless games are making use of new features. Just going off the onslaught of pre-launch marketing that we have seen.

#5584551 Posted on: 09/14/2018 05:45 PM

---

Great job Hilbert!

I know that everybody is justified to talk about the prices, but the tech itself looks crazy, I am genuinely impressed, especially considering that Nvidia didn't have to innovate at all. Pascal with +40% Shader engines would be enough for everyone to be satisfied, and much cheaper to produce.
Applause for the leather jacket.

I expect the initial benchmarks to be underwhelming, as this is a new architecture, but if Nvidia has a good driver for this at a point, it should be night and day with Pascal. The potential of the Tensor cores alone is insane.

Haven't seen you in a bit - how are things going?

@Hilbert Hagedoorn

Thanks for the article.

Can you post the full photo of DLSS comparison?

Also can you comment on this

"Another important change with NVLink SLI is that now each GPU can access the other's memory in a cache-coherent way, which lets them combine framebuffer sizes- something that was not possible with SLI before. The underlying reason is that the old SLI link was used to only transfer the final rendered frames to the master GPU, which would then combine them with its own frame data and then output the combined image on the monitor. In framebuffer-combined mode, each GPU will automatically route memory requests to the correct card no matter which GPU is asking for which chunk of memory."

You said that RTX cannot share memory.
Is this a software limitation where shared memory-access is limited to 'prosumer' cards?

I think that its more limited to prosumer workloads. The latency access across NVLink is too slow to complete rasterization tasks in the time required for realtime frame rendering. With certain workloads like data modeling and stuff, the card doesn't have to spit out a frame in X(ms).. it can take longer, in which case this can be useful. There may be upcoming graphics techniques where this can be applied but at the moment mGPU/SLI will continue to just share the frame buffer at the end.

#5584558 Posted on: 09/14/2018 05:54 PM

---

Haven't seen you in a bit - how are things going?



I think that its more limited to prosumer workloads. The latency access across NVLink is too slow to complete rasterization tasks in the time required for realtime frame rendering. With certain workloads like data modeling and stuff, the card doesn't have to spit out a frame in X(ms).. it can take longer, in which case this can be useful. There may be upcoming graphics techniques where this can be applied but at the moment mGPU/SLI will continue to just share the frame buffer at the end.
Hey man, just fine, but a different job role, I have missed you guys

I actually believe that AMD will compete with all this very soon, just not in the ultra high end. 750+mm wafers are an insane risk.

#5584559 Posted on: 09/14/2018 05:55 PM

---

This is just an appetizer! Let the crow eating begin:




https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf

#5584566 Posted on: 09/14/2018 06:07 PM

---

It all depends on the quality of the model training. Nvidia exaggerates always, so hold a smaller basket.

It is still impressive and the only logical way forward. Both the machine learning and the partial ray tracing. If AMD was more liquid, we would have had these a bit earlier.

People also forget that we will have an Intel GPU in a couple of years. Things are bound to be more interesting then.

#5584584 Posted on: 09/14/2018 06:56 PM

---

That's a lot of new tech pack into this new video cards , great review @Hilbert Hagedoorn , it looks good but that price it is what turns off many people. Nevertheless great to see new technology moving forward.

#5584587 Posted on: 09/14/2018 06:58 PM

---

Any new RTX tech/process adopted by competitors will need to be different so as not to infringe on Nvidia's hardware and software IP. In that aspect it may take longer to develop as opposed to licensing.

#5584613 Posted on: 09/14/2018 07:27 PM

---

I personally think these first RTX cards will be a flop. They will sell but people have already realized that these cards with RT On won't be pushing out the performance they expected.

Real-Time Ray Tracing might be a cool new feature but the hardware to run that type of tech is not here yet.

I don't care too much about Ray Tracing if you can only hit 30-60fps at 1080p.

#5584614 Posted on: 09/14/2018 07:28 PM

---

Strictly aesthetically speaking, I really dislike the MSI Gaming this time around, the design is a mess. I wonder how much better the cooling is compared to the Duke, which I find to look a lot better. And of course I'm looking forward to all the Turing reviews like everybody else, despite the fact that I'll most likely skip this generation. Nice preview boss.

#5584617 Posted on: 09/14/2018 07:37 PM

---

"The lines up top marked “1 Turing frame” are the key. For Nvidia’s other GPUs, it would consist of just one portion—the yellow FP32 line. But it’s more complicated in Turing.

While that standard FP32 shader processing takes place, the dedicated RT cores and integer pipeline are executing their own specialized tasks at the same time. Once all that’s done, everything’s handed off to the tensor cores for the final 20 percent of a frame, which perform their machine learning-enhanced magic—such as denoising ray traced images or applying Deep Learning Super Sampling—in games that utilize Nvidia’s RTX technology stack."

Source: PC WORLD

#5584620 Posted on: 09/14/2018 07:39 PM

---

I cant afford this. Im poor *sad, crying face*

edit: also, good stuff Hilbert, thanks a lot :´)

#5584623 Posted on: 09/14/2018 07:48 PM

---

ops:...when RTX will be out .... i will try to mod the BIOS - Tensor cores -

#5584624 Posted on: 09/14/2018 07:48 PM

---

Somehow I feel that all the effort will be thrown on the backs of game developers, which need to recode their engines specifically for Turing to take advantage of the new stuff.
And because the market for these will be so tiny in the first year or so... they won't bother, unless NV pays them big to do it.

Without all the new fancy mumbo-jumbo into the game code, the mighty RTX cards will probably be, as all estimates show so far, only 15-25% faster than their Pascal counterparts, from the raw number of FP32 cores, and that only in 4K. Lower resolutions will see even less gain.
Less than a week left, we'll see.

#5584626 Posted on: 09/14/2018 07:49 PM

---

Added, just click the links under the image - be advised though, images are provided by NVIDIA so I cannot vouch for them until I tested it myself in games. The live demos I played at the NV event looks pretty sweet though.