Speculation: GPU Performance Comparisons of 2020 *Spawn*

[pjbliverpool]

Do they? Is simultaneous execution some extension of D3D? Extensions for Vulkan? Genuinely curious. I'd assumed this is something the driver would handle.

I think it was in the deep dive that went through those Wolfenstien frame time slides. It stuck with me because I would have thought like you that this would have been automatic. But apparently not.

It seems if fully utilised Ampere packs an enormous amount of potential. But many of it's capabilities require specific game support so it'll be interesting to see how much of its potential is able to shine.

It'd certainly be a monster in a console.

 

[PSman1700]

I think it was in the deep dive that went through those Wolfenstien frame time slides. It stuck with me because I would have thought like you that this would have been automatic. But apparently not.

It seems if fully utilised Ampere packs an enormous amount of potential. But many of it's capabilities require specific game support so it'll be interesting to see how much of its potential is able to shine.

It'd certainly be a monster in a console.

I think UE5/modern engines will do very well on those new GPUs, it's a generational shift towards this kind of rendering.

 

[DegustatoR]

I think it was in the deep dive that went through those Wolfenstien frame time slides. It stuck with me because I would have thought like you that this would have been automatic. But apparently not.

It's automatic in a sense that if you have two asynchronous overlapping workloads they will run simultaneously instead of running serially. But to actually take advantage of that your code must have such workloads and last time I've checked most games didn't use either RT or tensor cores simultaneously or otherwise.

 

[Jawed]

Doubling the 5700XT bars at 4K in 3080 reviews is ...

... fun.

 

[DavidGraham]

Doubling the 5700XT bars at 4K in 3080 reviews is ...

... fun.

Yeah, double the 5700XT lands exactly at the 3080 performance, according to computerbase.

https://www.computerbase.de/2020-09/geforce-rtx-3080-test/2/#abschnitt_benchmarks_in_3840__2160

 

[Scott_Arm]

Yeah, double the 5700XT lands exactly at the 3080 performance, according to computerbase.

https://www.computerbase.de/2020-09/geforce-rtx-3080-test/2/#abschnitt_benchmarks_in_3840__2160

With clock increases they can probably beat Nvidia in raster performance assuming scaling to double the CUs is 100% efficient. I have my doubts about how well they'll match up in ray tracing.

 

[Deleted member 13524]

Doubling the 5700XT bars at 4K in 3080 reviews is ...

... fun.

It's exactly what I've been doing. 2x Navi 10 is what I'm expecting for rasterization performance.

Though I'm guessing nvidia will release a 3080 20GB as soon as October 28th comes up, and in good nvidia fashion I bet they'll enable one more SM for the 20GB model, like they did for the GTX1060 3GB/6GB.

 

[Jawed]

Though I'm guessing nvidia will release a 3080 20GB as soon as October 28th comes up, and in good nvidia fashion I bet they'll enable one more SM for the 20GB model, like they did for the GTX1060 3GB/6GB.

I would expect the 20GB variant to be $200 more.

EDIT: GDDR6X module availability is a question though, isn't it? 2GB modules are not going to be available this year are they? So the possibilities for 3080 better than what's just been launched appear to be:

1. 3080 12GB
2. 3080 one more SM
3. 3080 one more SM and 12GB
4. 3080 20GB (this year?)
5. 3080 one more SM and 20GB (this year?)

AMD could price the 16GB model higher than RTX 2080 10GB, e.g. $800.

AMD isn't going to persuade die-hard NVidia/G-sync users anyway, so there's no point in being "aggressive" on price with a 16GB card. Most Pascal users aren't waiting to find out if they'll be buying AMD, they're waiting to find out the "real" price-performance of Ampere.

Obviously this presumes that the biggest Navi 21 is at least ~same performance as 3080.

 

[Deleted member 13524]

EDIT: GDDR6X module availability is a question though, isn't it? 2GB modules are not going to be available this year are they?

It doesn't really matter if there'll be 2GB chips only in 2021. nVidia's purpose come October 28th will be to rain on AMD's parade, even if the RTX 3080 20GB (3080 Ti? Super?) isn't coming until March.

 

[DegustatoR]

nVidia's purpose come October 28th will be to rain on AMD's parade, even if the RTX 3080 20GB (3080 Ti? Super?) isn't coming until March.

Yeah, this isn't going to happen. NV won't announce anything which won't be available in a week or two from the announcement date.

 

[Jawed]

Hang on, aren't 2GB modules required for 3090, which is releasing this year?

 

[DegustatoR]

Hang on, aren't 2GB modules required for 3090, which is releasing this year?

It'll have 24 chips on both sides of the PCB. Probably one of the reasons why it will be $1500.

 

[Rootax]

I wonder how they will solve the bandwitdh problem. I mean, amd/nvidia can have big increases in raw performances, but memory bandwitdh doesn't evolve a lot (for a given price)... Have we seen some patents about bandwitdh saving recently ?

 

[Deleted member 13524]

Yeah, this isn't going to happen. NV won't announce anything which won't be available in a week or two from the announcement date.

Except for the RTX 3090 that was announced 3 and a half weeks ago and is only being made available tomorrow?

I wonder how they will solve the bandwitdh problem. I mean, amd/nvidia can have big increases in raw performances, but memory bandwitdh doesn't evolve a lot (for a given price)... Have we seen some patents about bandwitdh saving recently ?

Well there's rumors of RDNA2 GPUs coming with relatively narrow GDDR6 buses (256bit on Big Navi?), and AMD will increase effective bandwidth by using large amounts of on-chip cache.

 

[T2098]

I would expect the 20GB variant to be $200 more.

EDIT: GDDR6X module availability is a question though, isn't it? 2GB modules are not going to be available this year are they? So the possibilities for 3080 better than what's just been launched appear to be:

1. 3080 12GB
2. 3080 one more SM
3. 3080 one more SM and 12GB
4. 3080 20GB (this year?)
5. 3080 one more SM and 20GB (this year?)

AMD could price the 16GB model higher than RTX 2080 10GB, e.g. $800.

AMD isn't going to persuade die-hard NVidia/G-sync users anyway, so there's no point in being "aggressive" on price with a 16GB card. Most Pascal users aren't waiting to find out if they'll be buying AMD, they're waiting to find out the "real" price-performance of Ampere.

Obviously this presumes that the biggest Navi 21 is at least ~same performance as 3080.

My vote is actually for a 6th option. 3080 with one more SM and 352-bit bus with 11GB.
Makes the 1080Ti and 2080Ti owners able to heave a sigh of relief and not buy a card with less VRAM than their existing one, and still lets them harvest chips with 1 memory controller disabled (and splits the difference on VRAM costs as well)

 

[DegustatoR]

Except for the RTX 3090 that was announced 3 and a half weeks ago and is only being made available tomorrow?

A different model of the same card as 3080. If you were looking for a nitpick 3070 would be a much better option.

Well there's rumors of RDNA2 GPUs coming with relatively narrow GDDR6 buses

I'm about 90% sure at the moment that these rumors are based on ROP numbers - which aren't tied to MCs in RDNA2.

 

[Jawed]

It's performing "very well" if Navi 21 can only reach performance parity with RTX 3070Ti (cut-down, double memory GA102 or GA103?). If Navi 21 matches or exceeds 3080, then that looks like a fail.

We shall see...

Remember, AMD hasn't just switched to an entirely new node, Navi 21 is a tweak of an existing chip on a tweaked node.

Is what now?

Added RT and presumably fixed shitty performance per watt (5700XT has worse PPW than 2080Ti).

Do you think it's more than that?

 

[Kaotik]

Added RT and presumably fixed shitty performance per watt (5700XT has worse PPW than 2080Ti).

Do you think it's more than that?

Well, I wouldn't call little over 10% worse perf per watt "shitty", especially when other variants of the same chip offer as good PPW as NVIDIAs best Turings.
As for the Navi21 itself, while the architecture hasn't apparently gone through as major changes as RDNA did over GCN, it's still far cry from being just "tweaked chip" when it has twice the compute resources and who knows what other changes (we know at least that the chip supports both HBM and GDDR memories and that 256-bit GDDR6 will never be anywhere near enough fast enough for it unless they found the holy grail of memory bandwidth savings, so there has to be something major going on there)

 

[trinibwoy]

Well, I wouldn't call little over 10% worse perf per watt "shitty", especially when other variants of the same chip offer as good PPW as NVIDIAs best Turings.
As for the Navi21 itself, while the architecture hasn't apparently gone through as major changes as RDNA did over GCN, it's still far cry from being just "tweaked chip" when it has twice the compute resources and who knows what other changes (we know at least that the chip supports both HBM and GDDR memories and that 256-bit GDDR6 will never be anywhere near enough fast enough for it unless they found the holy grail of memory bandwidth savings, so there has to be something major going on there)

Are you expecting some secret sauce in discrete RDNA2 that’s not present in the consoles? I would bet the answer is simply that the 256-bit rumor is nonsense.

 

[Kaotik]

Are you expecting some secret sauce in discrete RDNA2 that’s not present in the consoles? I would bet the answer is simply that the 256-bit rumor is nonsense.

I don't really know what to expect.
I mean, it's highly unlikely that AMD would list false information in their drivers this close to release, which would mean it has both GDDR and HBM support, which is so far unheard of (to the point where they did two practically identical chips with just different memory controllers, Navi 10 & 12).
Then there's the 16 L2 cache lines, which suggest the 256-bit GDDR memory controllers (I think only couple Xbox SoCs have taken a different route on this with crossbar?), which we know can't be enough for 80CU RDNA(2) chip (unless they found the holy grail, which I think we can agree is quite unlikely)
So there needs to be another explanation - either it's right in front of us (using both GDDR & HBM) no matter how unconventional it sounds, or it has to be some as of yet unknown solution (512-bit? 384-bit?) + crossbar to deal with the L2 cache lines, or some third explanation I can't come up with.