Speculation: GPU Performance Comparisons of 2020 *Spawn*

[trinibwoy]

GCN introduced the dedicated INT pipeline back in, erm, 2012 and it's still there in Navi.

Sorry, are you saying that "scalar" in AMD parlance means INT? Because you can certainly have SIMD INT instructions.

From the whitepaper:

"The scalar ALU accesses the scalar register file and performs basic 64-bit arithmetic operations. The scalar ALU is used for a variety of control purposes, such as calculating values that will be broadcast across an entire wavefront, managing predicates for the wavefront, and computing branch targets. In addition, the scalar ALU is used for address calculation when reading or writing data in the scalar data cache."

That matches my understanding that the scalar unit was meant for truly scalar operations, not for processing INT SIMD instructions.

 

[Bondrewd]

If performance did not match the thoretical output

It did, but it breaks at 4k and nothing helps it there, for some reason.

That matches my understanding that the scalar unit was meant for truly scalar operations

That's what int pipe mostly does in Turing.

 

[trinibwoy]

bro it has scalar units for that purpose (whole two, one per SIMD, for Navi, that is).
Since GCN1 pretty sure.

Nope. Both you and Jawed confused me for a sec but it looks like I'm not crazy

"The lanes have also been rearchitected for efficiently executing a wide variety of data types as illustrated in Figure 9. The execution units perform single-precision FMA, add, multiply, or other general FP operations at full rate as well as 24-bit/32-bit integer multiply-accumulates."

AMD's SIMDs run either FP32 or INT32. Which means that the 5700xt has to run both on the same hardware while Nvidia has a dedicated INT pipeline. So my earlier point stands. The 5700xt is keeping up without the benefit of the dedicated INT pipe.

 

[JoeJ]

So GCN and Navi don't execute INT at all on the SIMDs and it's all running through a single scalar ALU? That sounds quite strange.

Yeah, GCN scalar is used for integer data that is constant over the whole wave, e.g. execution mask or memory addresses. It's nice and can save on per thread registers, but regular integer math is done on the same SIMDs as floating point math. Scalar and SIMD execution can be concurrently, but that's all it has in common with Voltas concurrent and separated int / float ALUs.

Somebody here said Ampere no longer uses this separation, but not sure if that's true. I though it only changed the ratio, so instead having 1 float : 1 int units, they now have 2 float : 1 int units. So even more units than Volta. Correct me if i'm wrong.

It might well be AMD won't adopt this idea if it does not suit their architecture. Probably we can not say 'AMD lacks a feature' just because of this, and we would need some benchmarks (which we really lack).

 

[trinibwoy]

That's what int pipe mostly does in Turing.

Source? Why types of INT instructions is AMD running on the SIMDs?

 

[trinibwoy]

Somebody here said Ampere no longer uses this separation, but not sure if that's true. I though it only changed the ratio, so instead having 1 float : 1 int units, they now have 2 float : 1 int units. So even more units than Volta. Correct me if i'm wrong.

That's right.

 

[w0lfram]

They've demoed NV solution running through DML on NV h/w. Not sure if this qualifies as MS working on their solution for quite some time.

No, they demoes Microsoft's solution, and thanked NVidia for the modeling and use of their hardware for their demonstration.

Right?

 

[Jawed]

Sorry, are you saying that "scalar" in AMD parlance means INT? Because you can certainly have SIMD INT instructions.

I'm sorry, I was wrong to call AMD's scalar ALU the equivalent of NVidia's INT SIMD. You're right, AMD runs per-work-item INT on the same ALUs as floating point.

 

[Leoneazzurro5]

It did, but it breaks at 4k and nothing helps it there, for some reason.

Are you suggesting that shader core itself is "more capable" (Flops or CU IPC) in desktop RDNA2 respect to RDNA1 then? Because even under 4K it is on average slightly under the 2070 Super and yes it completely breaks at higher resolutions. But for competing with the 3000 series it needs quite probably a better shading output, also thinking that part of the RT calculations are done in the shaders.

 

[trinibwoy]

I'm sorry, I was wrong to call AMD's scalar ALU the equivalent of NVidia's INT SIMD. You're right, AMD runs per-work-item INT on the same ALUs as floating point.

No worries, all clear now!

 

[chris1515]

It did, but it breaks at 4k and nothing helps it there, for some reason.

That's what int pipe mostly does in Turing.

This is the only example of RDNA2 GPU but I was surprised to see so many native 4k 30 or 60 fps on PS5 maybe something improve on RDNA 2 side. Deathloop a third party title is announced native 4k 60 fps for example. We just need to wait RDNA 2 AMD GPU with much more power and memory bandwidth than a PS5.

 

[Bondrewd]

Are you suggesting that shader core itself is "more capable" (Flops or CU IPC) in desktop RDNA2 respect to RDNA1 then?

Nah.

Because even under 4K it is on average slightly under the 2070 Super and yes it completely breaks at higher resolutions.

Yes, there was something odd with it.

This is the only example of RDNA2 GPU but I was surprised to see so many native 4k 30 or 60 fps on PS5 maybe something improve on RDNA 2 side.

It's a console so not directly applicable.

 

[Leoneazzurro5]

Then , as said, if there is a big improvement in Perf/W (on the same process ) it may be due to better power management but also to bottleneck resolution...

 

[yuri]

High-end thingies are usually made for PR purposes.
Being the best in a set of metrics is very nice and it gives every other product in your lineup the prestigious halo of dominance.

Actually, the highend market is fucking huge. JPR estimates the highend segment makes 47% of the 2020 TAM. So no more sad stories such as Polaris.

 

[Bondrewd]

Actually, the highend market is fucking huge

Not $1500 paperweights, no.
The classic high-end lives at $450-700.

So no more sad stories such as Polaris.

?

 

[Benetanegia]

He explained it pretty clearly: G92 was priced and positioned in anticipation of RV670, while GTX 970/980 was launched to compete with the then old and discounted R9 290/290X.

Neither AMD nor Nvidia price their chips out of the kindness of their hearts. If anything, it's been Nvidia's that's been caught by surprise by AMD's performance/price ratios (e.g. RV770 causing refunds and pricing reversals of GT200 cards, and Hawaii disrupting Kepler Titan pricing).

I understood what he said perfectly. It is just false. The links I posted show it clearly. Better go gaslight someone else.

And no one said they did it out the kindness of their hearts. They sold like hotcakes and that was their intent.

 

[Leoneazzurro5]

In fact, 8800 GT was not hurried, this is why the first batch of cards on the market caused a major turmoil because of overheating and badly designed cooling system. And I know something about it, I owned one (luckily I was able to get RMA with an updated model). Lol.

 

[sonen]

Actually, the highend market is fucking huge. JPR estimates the highend segment makes 47% of the 2020 TAM.

Not $1500 paperweights, no.

Super high-end segment with the average price of 1236 euros per piece brought 14% of the entire Nvidia revenue from Mindfactory in Q1 2020. Which is almost the end of the life-cycle, so not hard to imagine contributing even more on average and especially during the initial WOW phase.
That's more revenue than Mainstream 100-250€ and easily even more profit. All this offset by higher purchasing power of Germans.

https://www.3dcenter.org/artikel/mindfactory-grafikkarten-retail-verkaufsreport-q12020

So no more sad stories such as Polaris.

?

Maybe he means... letting Nvidia to roam freely by ceding X% of the market right of the bat because "Is All About Giving Gamers A Sweet Performance Per Dollar" as Raja said.

 

[Bondrewd]

That's more revenue than Mainstream 100-250€ and easily even more profit.

200-500 still rules everything with iron but fair fist.
$300-350 sweet spot needs to be a fight once more.

Maybe he means... letting Nvidia to roam freely by ceding X% of the market right of the bat because

They're ceding it either way; kids just don't like AMD.

"Is All About Giving Gamers A Sweet Performance Per Dollar"

Didn't Jensen adress Pascal owners basically the same way during the recent keynote

 

[sonen]

Didn't Jensen adress Pascal owners basically the same way during the recent keynote

Maybe he did.
I am sure he didn't say he won't do business with you if you have more than $250