But Jensen, how fast is it ?

[ToOo]

Seeing as there were no relative performance figures for the 20xx vs the 10xx series in non raytracing applications it's a relatively safe bet there arent many architecture improvements on the pixel shader side and the improved 12nm process isnt likely to let them clock much faster. Assuming that how fast are the 20xx cards?

For a while the FLOP calculation for Nvidia GPU's has been
CUDA cores x Clock speed x 2

TITAN V: GPU Cores 5120
1455 MHz (Turbo)= 14.9 TFLOP
1900MHz (overclock) = 19.5 TFLOP

RTX 2080 Ti : GPU Cores 4352
1545 MHz (Turbo)= 13.4 TFLOP
2000MHz (overclock ?) = 17.4 TFLOP

GTX1080 Ti : GPU Cores 3584
1582 MHz (Turbo) = 11.3 TFLOP
2000MHz (overclock) = 14.3 TFLOP

RTX 2080 : GPU Cores 2994
1710 MHz (Turbo)= 10.2 TFLOP
2000MHz (overclock ?) = 12 TFLOP

GTX1080 : GPU Cores 2560
1733 MHz (Turbo) = 8.9 TFLOP
2000MHz (overclock) = 10.2 TFLOP


Keep in mind there's some assumptions going on here but FLOPS generally translate pretty linearly to frame rate within the same general architecture. If i'm right RTX2080 looks like a pretty terrible deal while RTX 2080Ti isnt actually that bad, despite the initial sticker shock.

 

[ToOo]

Titan V overclocks a bit better than you have there and actually performs much the same as Pascal for clockspeed.

I did lowball the clock on the Titan V but only because i've not seen them consistently hit 2GHz and counting 50MHz jumps seems a bit mealy.

The OCUK timespy results are actually pretty revealing, taking the top overclock GPU score for each of the cards on OCUK
TitanV @1972/1000, GFX Score 14598 (2.85 points per cuda core)
TitanXP @2050/3106, GFX Score 11405 (2.97 points per cuda core)
1080 Ti @2151/3000, GFX Score 11267 (3.14 points per cuda core)
1080 @2190/2850, GFX Score 8552 (3.34 points per cuda core)

There's an 11% difference in clock speed between the Titan V and the GTX1080 and there's 17% difference in points per cuda core. In the scheme of things that 6% difference is pretty small compared to the overall gain in performance. Like i said there are assumptions here which i know are flawed, i'm just trying to "guess with the benefit of experience"

 

[ToOo]

You really need to discount the 1080 and 1080 Ti scores above as they were done on waterblocks with exceptional cards. Normal 1080 and 1080 Ti cards hover around 2000mhz on air much the same as the Titan V and Titan Xp do.

With respect you probably are correct but I dont see why that's really relevant to the core argument of my post, you're getting a little to hung up on the detail. The point is that RTX2080 is a total dog at that price point and nothing from RTX2080TI will get anywhere close to TITAN V.

p.s. my 1080 is on air and matches the other ones score, anecdote is anecdote but still

 

[ToOo]

That GTX1080 score is with the GTX1080 11gbps card. The normal GTX1080 came with 10gbs ram and at 2190/2762 does 8349

https://www.3dmark.com/spy/397173

(yes the benchmark is mine and last time checked the Firestrike GPU scores of that card weren't beaten by any other 1080)

This is still not relevant

My personal 1080 is hitting 8552 GPU score in timespy on air and i gaurantee you it's an original 10Gbps card

https://www.3dmark.com/3dm/22224310

 

[ToOo]

True but they went where the money was, we can't really blame them for that. AMD have in the past had the best cards both for performance and for the money (i'm thinking the 290x) and still lost to NVIDIA mind share at the high end, we really have no one to blame but ourselves.

 

[ToOo]

2080TI should be in the ball park of 150% 1080 performance assuming both are overclocked, that's a pretty sizable jump and as you say there isnt really another option right now.

 

[ToOo]

No.

Please don't post unsubstantiated and unrealistic claims. Stick to Reddit for that.

If you read the post at the start of the thread you'll find that it's not entirely unsubstantiated . Speculation yes to be sure, but not without some consideration.

 

[ToOo]

I will change my signature to say :
JediFragger "What Are you ?"
ToOo: "An idiot sandwich"

If the top RTX2080TI GPU score on the OCUK forum in time spy is anything less than 150% of the best 1080 by the end of October (i.e 12828 GPU at least)

 

[ToOo]

Oh dear. Surely if that was the case then benchmarks and graphs would be plastered all over the event?

"NEW RAYTRACING SYSTEMS AND OVER DOUBLE THE PERFORMANCE OF OUR PREVIOUS GENERATION!!"

But no. Nothing. That doesn't strike you as odd? I'll be surprised if we get a 15% bump.

Sorry but you seem to be arguing and agreeing with me at the same time...

I'm talking 2080ti to 1080. Not 1080ti. If the 2080ti is 150% of 1080 then I worked it out as around 112% of 1080ti or another way of saying that is 12% faster than the 1080ti.

 

[ToOo]

Very much suspicious. There's plausible room for 2080ti to be a decent overclocker, it's a big chip with lots of cuda cores and gobs of memory bandwidth. But I just can't see how 2080 is going to be anything but a dog. So many fewer cuda cores than 1080ti means it will need to hit insane clock speed just to match it unless they've made massive improvement to the architecture, and since they're already telling us it's an evolution on Volta I'd bet it's really just Maxwell rev 4.

 

[ToOo]

Could be some architecture improvements or there might be a way to use those tensor cores for regular raster graphics in the new driver. But failing that the only answer i see is clock speed and the published base and boost clocks would seem to discount any massive gains on that side.

 

[ToOo]

But 980 wasn't all that much faster than 780ti which had been discounted to compete against 290x so there was limited room for Nvidia to charge more than about £550.