And that was across the board for a range of current modern games. Some games saw a 100% or more improvement.
Nvidia Ampere Discussion [2020-05-14]
- Thread starter Man from Atlantis
- Start date
-
- Tags
- nvidia
And that was across the board for a range of current modern games. Some games saw a 100% or more improvement.
Yeah, which means that they can save power when there's enough INT32 code in a warp to do just 16 FP32 and 16 INT32. When there isn't going with 32 FP32 will net you anywhere from 3 to 100% more performance.
You can't go with FP32+FP32+INT32 since warps are 32 wide - there aren't enough data to fill 48 SIMD lanes. Which is why there were always two choices for Ampere: either double width FP32 SIMD instead of 16-wide FP32+INT32 of Turing or what we seemingly got in the form of 16-wide SIMDs capable of both FP32 and INT32 but only 32 lanes total per clock - as 32 FP32 or 16 FP32 + 16 INT32. H/w can handle this in three ways really: either 2 SIMDs capable of FP32/INT32 or one FP32 and one FP32/INT32 or 3 SIMDs with INT32 still being a dedicated one. The latter option is IMO the closest to your typical NV h/w design choices.
You are comparing a 3080 having vastly more CUDA cores with a 2080.
A comparison of a RTX 3070 with a RTX 2080 would be more interesting (2944 vs 2944 FP32). There, the rasterization performance gain is a lot smaller.
Not sure if I follow you here. Weren't you the one who brought up FP32+FP32+INT32? And what's warp width got to with it? Each warp goes into one block in it's entirety until the instruction is completed, which normally is two clocks for SIMDs are 16-wide and warps have 32 items.
Since it's also supporting the Sparsity thingie, it's highly likely IMHO.
Nvidia is doing these comparisons, not me. It's the only data point from an official source so far. Additionally, I compare not the products, but the perf increase theoretical (2.7x FP32) vs. what's seen in games (~1.6-1.7x).
Benetanegia
Regular
That's not entirely correct because it's 36 INT per 100 FP, and it's not - 72 either because that would be if the total number was 200+72. For 200 total, the actual mix is 147 and 53, you can use the rule of three to confirm that it's the same percentage as 100 36.
And I guessed it easily because I actually kind of arrived at that several pages ago, in my own investigation. Basically both Turing and Ampere can do 200 ops over same period of cycles, but Turing only actually does 136 in your typical game. So 200 / 136 = 1,47.
As a h/w config, yeah, not as what gets scheduled per clock as this can't be more than 32 lanes due to warp width being 32. Two 16 wide units can be scheduled each clock because they are running 32 wide warps so the scheduler can fill them both each next cycle in turn while the other unit is executing what was scheduled in the previous clock. With three 16 wide units you have a scheduling issue as one of them will be idle any clock cycle with 32 wide warps. Hence why it's 32 FP32 or 16 FP32 + 16 INT32 and not 16+16+16.
Would be interesting to see if they'll be able to fit something gaming (graphics) related onto it.
You're right, thanks for the correction! So it's roughly 12% above that 1.47, which means other factors like additional ROPs and more L1 bandwith play their roles too.
Never argued that point, in fact, I said it as well: FP32+[FP32 or INT32] with 16-wide each of course.
Lenovo may have leaked the existence of an RTX 3070 Super/Ti SKU with 16GB of RAM:
https://www.tweaktown.com/news/7491...rce-rtx-3070-ti-rocks-16gb-of-vram/index.html
It wouldn't surprise me if Nvidia are also prepping a 20GB equipped RTX 3080 SKU just in case RDNA2 performs better than expected.
https://www.tweaktown.com/news/7491...rce-rtx-3070-ti-rocks-16gb-of-vram/index.html
It wouldn't surprise me if Nvidia are also prepping a 20GB equipped RTX 3080 SKU just in case RDNA2 performs better than expected.
Frenetic Pony
Veteran
And there we are, could've done it from the get go but no, gotta eke out that little bit of extra profit from the earliest adopters. Better thing to wait too, I'd expect the 6700xt or whatever to be around the same performance as a 3070 for $400-500, depending on just how competitive it is.
Do wonder if we'll see a 3080ti or whatever as well, another 8gb stick in. Since I'd expect "Big Navi" at lest in the lower end scenario, to compete with a 3080 with at least 12gb of ram, if not 24, it'd look better for Nvidia to have the ram numbers.
There is good evidence that there are 7 GPCs see my previous post.
That would mean 7 * 16 ROPs = 112 ROPs for a full GA102
I'm still very curious to the number of TMUs.
As L1 cache bandwidth has doubled, there is a good chance also TMUs have doubled from 4 to 8 per SM.
That would mean a whopping 8*84 = 672 TMUs.
this for real ? lol wow thats big
3080 is a 28cm card, so looks right. RTX2080 and GTX1080 were 26cm long.
sigh i have a vega 56. I hope a 3080 fits my case lol
Nvidia has a trend to increase the number of multi-processors per GPC since Kepler (or keep the GPC unit count hard capped at six), so GA102's config is more likely 6*14 with 2 MPs disabled for the RTX3090 SKU.
Vega 56 is almost a 27cm card, so 3080 is at worst 2 cm longer.
Edit: Yah Vega 56 is 26.8cm and 3080 is 28.5 cm.
Last edited:
well thank you. I was going to go get my measuring tape and take a gander but i don't have to now.
Only thing worrying me with the 3080 is the 10gigs.
my guess is if AMD is competitive with navi 2 we might see a 3080ti with 20 gigs ?
