'Turing mid-range GPU' (256-bit GDDR6 bus) still will beat it as far as I can predict good.
'Turing mid-range GPU' (256-bit GDDR6 bus) still will beat it as far as I can predict good.
If you're suggesting midrange Turing (that's the xx60) would beat 1080 Ti then I'm pretty sure you can't predict good.
If you're suggesting midrange Turing would beat Titan V you can't predict even badly.
ImSpartacus
Regular
By "mid-range", he probably means the hypothetical 104-type (GT104?) GPU that would show up in an 1180 and 1170. Historically, that "tier" of chip has a 256-bit bus.
And yes, that tier of chip probably will match or slightly beat the 1080 Ti. It'll also probably cost about $600-800.
xx70-xx80 range could indeed match or even slightly beat 1080 Ti, just can't think those as "midrange". No chance they would beat Titan V though really
ImSpartacus
Regular
Don't get hung up on the "mid-range" thing. That's a constant source of childish argument since the 104-type parts technically aren't at the top of Nvidia's lineup (that's 100/102-type), leaving them somewhere in the middle (i.e. "mid-range"?), but they are hardly slouches in the performance department, especially against the competition.
Concentrate on the 256-bit part. That correlates with memory bandwidth, which strongly correlates with general performance.
If the "old" 256-bit GPU used 10Gbps memory and the "new" 256-bit GPU used 14 Gbps memory, then that's a 40% bump in peak memory bandwidth.
It's not hard to see a scenarios where generalgeneral perf also increases by roughly 40%, which gets you to 1080 Ti territory and beyond.
Benetanegia
Regular
Well, I wouldn't concentrate on that much either. How much of V100 is devoted to units that improve gaming performance vs FP64, Tensor, NVLINK, etc? And isn't Titan V using a significantly cut down chip too? How much higher could the smaller chip clock, considering a not too different TDP, if they go that route?
I don't expect Titan V performance at all, but I don't think die size comparisons to an HPC chip is of so much value either.
Edit: Personally I think that ImSpartacus is a lot more spot on with his bandwidth comparison.
Benetanegia
Regular
Yeah, I appreciate how you ignore everything else I considered, besides die size. I don't need to be an expert to tell you that IF all that stuff was in fact 400mm^2 (which everybody knows is not the case), the 256-bit 400mm^2 consumer "Volta" chip would not only beat it, it would absolutely and utterly destroy the Titan V in gaming performance. And that is ironically a reality that you are ignoring.
I'll reiterate my opinion that you are looking at the wrong metric to compare. If we step aside the Titan V, for just a little bit, I'm afraid that the notion of a 256-bit GDDR6 400mm^2 12 FFN "Volta" chip beating a 384-bit 471mm^2 16FF+ Pascal chip is not something only posible in science fiction or fantasy, sir. And it's not something I'm taking for granted either, but the probability is high, I'd say. And at that point the Titan V is just a mere 25-30% off, in best case scenario. So really nowhere close to what a 2x die difference may suggest.
Benetanegia
Regular
What? Stop projecting, please. The only strawman is comparing performance of a severely cut down HPC card to a posible consumer focused part on a single metric and pretending that's "reality".
There's nothing strawman in the second paragraph on my post. And the first one is an hypothetical case entirely based on your claim. No one but you suggested that the "gaming portion" of V100 would be 400mm^2. It's doesn't need to be that small for a 400mm^2 gaming focused chip to utterly destroy it, because die size is not the only metric that governs performance.
Benetanegia
Regular
If you need to be remembered of your claims, yes, I will gladly do so.
No one came close to claiming 400mm^2 for the gaming portion of V100, besdes your post. Many glaring differences where pointed out tho, which you decided to ignore in favor of your strawman that "die size is reality, everything else doesn't matter".
EDIT: And BTW, which delusion? That the next Gx104 part will more than likely beat the current Gx100/102 part like it's been the case in every generation prior to this?
By all means quote them. Otherwise stop using logical fallacies (or lying), whichever you prefer.
What are you talking about. I never said anything about the amount of area dedicated to gaming hardware vs non. This all bullshit you have fabricated. The 400mm^2 figure is simple math from the known die size of GV100 at ~800mm^2 and the rumored die size for GV104 of ~400mm^2. 800 - 400 = 400 derp.
GV100 is not a meaningful prior generation (that would be GP100), and there is no process shrink either. So I have no idea what you are smoking, but I don't want any of it. Bye.
Benetanegia
Regular
And where did I say that your claim was that? Your claim, whether you formulated it that way or not is that "the only way in which 400mm^2 GV104 (let's call it that for simplycity) could match GV100 in performance is if the gaming oriented parts of GV100 itself are 400mm^2". Because you ignored anything but die area, while I asked about die area, but also about how much of that area is actually active (Titan V is a severely cut down GV100 in case you didn't notice yet), and I talked about clocks, and many other things that would impact the performance of the Titan V vs the hypothetical GV104.
Again, who said anything about GV100 in that comparison? Do you even read the posts you reply to? Me, as well as ImSpartacus, the only two people you replied to, are talking about Gx104 beating GP102. We never said anything about matching GV100. Of course GV100 is not too far off from GP102, which is only 470mm^2 in 16FF+ BTW. Meaning that if GV104 beats GP102* it wouldn't be too far from Titan V either.
*Which is reasonable, considering a somewhat better node, similar die size, narrower memory config that can help make the already small size difference even less relevant, architectural improvements, etc.
EDIT: BTW if you have any weapon other than ad hominem in your arsenal, it'd be much appreciated. The sum of all your contributions to the discussion are zero.
Benetanegia
Regular
LOL. No. Don't even try that, I mean, it's even in the same page we are now. It was a reply to this direct quote:
To which you responded:
And it's that very claim that I disputed. (Edit: not that it is twice the size, mind you, but that it correlates to performance difference)
D
Deleted member 2197
Guest
NVIDIA® Deep Learning Accelerator (NVDLA) project
Documentation for future reference.
http://nvdla.org/primer.html
Documentation for future reference.
http://nvdla.org/primer.html
Last edited by a moderator:
That part sounds very strange to me. I don't see how that could be.
Maybe they are talking about overall power consumption of needing to move that data around?
Dunno, doesn't seem like it would be true in the vast majority of the cases.
