WTF is a Generation? *PissOff from Scorpio Thread*

[Shifty Geezer]

Generations are probably over.

Iterations are actually the new hotness.

That's my belief. Hence instead of offspring, I'm working on cloning myself with genetic enhancements each time. Currently on Shifty Geezer 2.3, although 2.4 is nearly ready with 5% better visual acuity, so I'll throw this model off a cliff and upgrade when it's done.

 

[Technewsreader]

they are done with generations. Thats what they said. Its too powerful to be considered a mid-gen upgrade. Also if Supersampling to 1080P will be pushed like on pro i'd not view it as "Xb1 for 4k owners" There is more to that console.

 

[iroboto]

This is not true. AMD GFLOPS != Nvidia GFLOPS. According to AMD's numbers, the R9 390X is a 6 TFLOPS card, and according to benchmarks it performs roughly like a GTX 980. The Ti version is about 25% faster.

Oh for the love of it. FLOPS are FLOPS. Whoever came up with this concept that a flop is not a flop is probably working for trump, cause he somehow made you believe a meter is not a meter if someone else made it.

If you want to talk about architectural differences or driver differences refer to the hardware or driver in question. Clearly VEGA hardware will completely out perform GeForce 1 performance at the same FLOP count. Pascal performance will crush GeForce 1 performance at the same FLOP count so Nvidia flops != nvidia FLOPS???

Lol oh man. Let's stop with this false generalization.

 

[Deleted member 11852]

Oh for the love of it. FLOPS are FLOPS.

When it comes to compute, this is not the case at all. I've mentioned before that we invested moderately in both AMD and Nvidia clusters to explore what we can do with compute instead of traditional general processing and there can be quite significant difference. I can't see why this should be different if that same hardware is being tasked with more traditional graphics tasks.

Not all floating point instructions are equal.

 

[iroboto]

When it comes to compute, this is not the case at all. I've mentioned before that we invested moderately in both AMD and Nvidia clusters to explore what we can do with compute instead of traditional general processing and there can be quite significant difference. I can't see why this should be different if that same hardware is being tasked with more traditional graphics tasks.

Not all floating point instructions are equal.

ehh, I guess it's more a technicality I have an issue with.
FLOPS are derived from formula, much like how we calculate velocity and acceleration.
The best way I can describe the behaviour you are modelling after is that 2 objects dropped from height will receive the same amount of gravity (FLOPS) but their force will vary on impact (performance) due to their weight.

I don't mind people calling the performance/saturation being different (which will vary for each architecture). But FLOP is a derived equation, we haven't changed it since it was developed back when Floating Points were introduced.

 

[Shifty Geezer]

It's not the FLOP that's different - multiply accumulate does the same thing in a cycle no matter the GPU. It's the other stuff controlling flow that makes the difference, which is iroboto's point. Thus the summary that nVidia FLOPs are different to AMD FLOPs is inaccurate, yet a usable short-hand to say that different efficiencies between architectures can yield more functional performance for a given raw FLOP throughput.

Both viewpoints are valid. People can carry on beating each other over the head over their preferred interpretation if they want. Or perhaps we should come up with a better, more accurate shorthand?

This is not true. AMD utilisation != Nvidia utilisation. According to AMD's numbers, the R9 390X is a 6 TFLOPS card, and according to benchmarks it performs roughly like a GTX 980. The Ti version is about 25% faster.

Job done?

 

[Deleted member 11852]

It's not the FLOP that's different - multiply accumulate does the same thing in a cycle no matter the GPU.

And if you limit yourself to common floating point operations you can take that view but the reality is when you move beyond IEEE 754 you'll find both vendors have instructions with extended operands that result in a single floating point operation but for which there is no equivalent on the competitor so comparing FLOPS increasingly becomes meaningless. In part it depends on what you are doing.

 

[iroboto]

And if you limit yourself to common floating point operations you can take that view but the reality is when you move beyond IEEE 754 you'll find both vendors have instructions with extended operands that result in a single floating point operation but for which there is no equivalent on the competitor so comparing FLOPS increasingly becomes meaningless. In part it depends on what you are doing.

I think we're in agreement then, it's meaningless to compare FLOPS

 

[Shifty Geezer]

And if you limit yourself to common floating point operations you can take that view but the reality is when you move beyond IEEE 754 you'll find both vendors have instructions with extended operands that result in a single floating point operation but for which there is no equivalent on the competitor so comparing FLOPS increasingly becomes meaningless. In part it depends on what you are doing.

When a GPU vendor reports the number of FLOPS of their GPU, what instructions do they count?

 

[Deleted member 11852]

When a GPU vendor reports the number of FLOPS of their GPU, what instructions do they count?

You expect GPU manufacturers to be transparent about that?

 

[Shifty Geezer]

Clearly they're going to express the shortest 1-cycle maximum operations for the biggest number. Ergo I think it safe to say they are counting the vector FMADDs as iroboto said ("it's an equation," units * SIMD * fmadd * clock sort of thing). I'd say the amount achieved in one actual operation as requested is part of the efficiency/utilisation consideration. The FLOP counts we use are the raw calculation of peak maximum attainable in one calculation based on the ALUs performing one FMADD per cycle (plus any pesky extra little bits).

 

[Deleted member 11852]

Nvidia and AMD GPU's have single cycle floating point operations? What are they?

 

[ProspectorPete]

That's my belief. Hence instead of offspring, I'm working on cloning myself with genetic enhancements each time. Currently on Shifty Geezer 2.3, although 2.4 is nearly ready with 5% better visual acuity, so I'll throw this model off a cliff and upgrade when it's done.

Throw off a cliff with each enhanced version? Why not keep 1 in each (bed)room? I mean, you at least play Netflix right

 

[London Geezer]

A new gen is when a new Zelda comes out.

 

[iroboto]

Nvidia and AMD GPU's have single cycle floating point operations? What are they?

I think we can use this for GPU FLOP calculation:
http://uniquecalculator.com/calculation/FLOPS-calculator.php

edit no. Maybe not.

just use the classic method:

You just need to figure out the FLOPS/cycle
https://en.wikipedia.org/wiki/FLOPS

provided for CPUs in the wikipedia, we'll need to dig around for GPU.

For Pascal:
The theoretical single-precision processing power of a Pascal GPU in GFLOPS is computed as 2 (operations per FMA instruction per CUDA core per cycle) × number of CUDA cores × core clock speed (in GHz).

The theoretical double-precision processing power of a Pascal GPU is 1/2 of the single precision performance on GP100, and 1/32 on GP102, GP104, GP106 & GP107.

The theoretical half-precision processing power of a Pascal GPU is 2× of the single precision performance on GP100[8] and 1/64 on GP102, GP104, GP106 & GP107.[14]

GCN?
Sounds the same.
# SIMD in a CU * # of CU * 2 FMA * Clock Speed = FLOPS

 

[Deleted member 11852]

That works when all things are equal and AMD and Nvidia floating point operations are not equal and flops/cycle are not equal. This is as dumb as the people who tried to compare Motorola and Intel CPUs based on clock speed.

 

[AlNom]

You expect GPU manufacturers to be transparent about that?

 

[London Geezer]

That's beautiful. Tremendous.

 

[Cyan]

Would you still think that if it uses Jaguar, or do you regard PS4 Pro as a "next generation full-fledged console" as well?

what do you think? PS4 Pro shares the same base name as the PS4. Sony never intended to abandon the path of PS4. On the contrary, Xbox One name is going to be avoided like the plague,I guess.

 

[ProspectorPete]

what do you think? PS4 Pro shares the same base name as the PS4. Sony never intended to abandon the path of PS4. On the contrary, Xbox One name is going to be avoided like the plague,I guess.

Sorry I believe I misunderstood your post.
MS is going to present it as a clean slate, which will be backwards compatible/share all games with Xbox One. Even if it has a Jaguar as a CPU. Is that in line with what you meant?