Well one of the reasons for leaving out such things and creating a big gap between the cards is so they can charge the prices they want. Compared to a 1080ti it's only 1gb less though so is that really a big deal?
It has to do with the GDDR6X memory they are using. It currently only has half the chip memory capacity of GDDR6 memory. This will likely be rectified as time goes on, but it made it difficult for Nvidia to put large memory sizes on their new GPUs. Also, Nvidia usually differentiates their professional products from their gaming products with memory capacity, so that is likely another reason.
Yeah, that wouldn't be worth the cost. You'd be better off just pocketing the $200 and putting it towards a 4080 or equivalent in a couple years when 10GB might start becoming restrictive.
Hardwareluxx too made a similar diagram:OK I have created Ampere SM approximation to see the difference of the new Cuda Cores partition
Inside the red square is the new Cuda Core partition that can execute one 16x FP32 or one 16x INT32 per cycle.
With the addition of the second 16x FP32 partition they can now execute 16x FP32 + 16x FP32 per cycle or 16x FP32 + 16x INT32.
So now they can do 128x FP32 versus 64x FP32 per SM and thats the double throughput they get in Ampere.
Ampere GA102 SM Approximation
No offense to my statement when i say this...
One data path is capable of FP. The other datapath is capable of FP or INT, so you can't do FP+FP+INT all at the same time.
Correct as answered for point 1. The existence of the FP and INT compute cores aren't all that is necessary to do calculations. There are registers, dispatch units, etc. that are required to actually get the correct data to and from the FP and compute units. The second data path that can do INT or FP will share some of the circuitry to save on space and power.
That is what Turing did. Turing could do 1 INT + 1 FP. This is the next evolution where there is an additional data path but only for FP calculations which allows for either 1FP + 1FP or alternatively 1FP + 1INT.
You could do that but everything in engineering is a balancing game. You have size and power considerations, register pressure considerations, bandwidth considerations, etc. Basically Nvidia looked at what calculation mix modern games are making, what calculations they think modern games will be making in the near future, and planned the architecture accordingly. More specifically, Nvidia looked at the calculation mix they think Ampere will be asked to be performing for modern games and determined that it would be roughly 2/3 FP and 1/3 Int. So this was their solution to tweak Turing to get the most balanced and efficient architecture to meet that compute mix. Hopefully this makes sense.
Unless NVIDIA provides the number of chips actually shipped and sold, everybody will spin this just as they want, and none of them will be either right or wrong. It will be Shrödinger's supply and demand. Exactly how it always have been, no matter which brand or vendor we're talking about.
So this is why 30 series cards' gaming performance is so much lower than the max theoretical compute performance.
So they did it to specifically enable FP+FP. I suppose this is probably aimed at scientific calculations, etc? Are there any desktop use cases where this could come into play?
Is it possible to guess how much bigger the die would've been if they implemented three paths?
A big reason, yes. The quoted FLOPs is assuming you are doing FP+FP the entire time which is not realistic for gaming. I haven't studied all of Nvidia's marketing slides but usually there's additional FLOP quotes where they are counting tensor flops and everything together as well which is also unrealistic.
Pure compute should benefit obviously but I'm sure gaming will benefit as well because, as I said, games tend to be FP heavy so you'll get a decent speed boost out of it. No where near 2x mind you, but it should be a noticeable boost.
Not really, not without knowing the size of each CU and each data path inside the CU which I'm pretty sure no one can tell you outside of Nvidia but if someone can, I'd be happy to hear.
Depends on a lot of things. If we waive away power and size requirements, then most likely you could get a very noticeable speed bump. However, we do have size and power requirements so the real question is what do you have to sacrifice in order to put that 3rd path in. I'm guessing at 8 nm, you would do more harm than good putting in that third path.
Well with the benchmarks we have now it would seem that the 3070 with "5888 cuda cores" is indeed faster than the 4352 cores of the current 2080ti. Though until 3rd party benchmarks come out its not going to be exactly clear how much the new design scales in games and typical desktop applications. If we take the ampere reveal at face value then objectively it would seem that have truly doubled compute performance.
I game at 1440p so I'm pretty sure it will be just fine for quite some now. At this point I just want to upgrade - RTX 3080 would be absolutely massive upgrade and it would still offer 2 GB more VRAM. Heck, I don't care what comes out after.
Yep. It seems that Nvidia is allowing for some flexibility in the int/fp instructions ratio as the fixed ratio that they used for the 2xxx series can only be rarely optimal for any one game. This way they get closer to the best possible.
Facebook
Twitter