Teraflops: How Important is this Unit of Measure?

[fiendcode]

GPU OEM's do not use FLOP's because it is a waste of time. Let us just say that.


We had GFLOP's prior to this generation. So thinking that the evolution to teraflops is a terrible metric performance is clearly neither here nor there. If anything, the original Xbox had 7.3GLOPs, the PS2 had 6.2GLOP's.

Clearly that performance metric mattered then Sony and Toshiba came up with Emotion Engine and when Microsoft went to Nvidia for a card.

Click to shrink...

Actually the OG Xbox was 20 GFLOPS and Gamecube was 9.4. Also PS2's 6.2 GFLOPS figure is combining the CPU+GPU while GC/Xbox are only their GPUs. For later consoles in this general GFLOPS range Wii was 14.1, 3DS was 6.4 and Vita was 14.2 (or 28.4 in boost mode).

 

[Utherellus]

Correlation does not imply causation. Beginning of the gen was very weird, some games (PES for example) having a performance difference larger than even the Tflops implied.

Click to shrink...

Sure. Its not 100% accurate representative. But again, on the same architecture it works in 90% of cases.

 

[Fredrik]

With the same game and developer on the same architecture - pretty important.

Different game and developer on different architecture - not important at all.

CU count, clock speed, etc, could matter as well if everything else is the same as with the top example.

 

[Albert Penello]

You're wrong. The PS4 gpu was widely understood to be "40% better" than the Xb1 gpu. In fact, it was such a prevalent topic of discussion that a Microsoft employee tried to downplay the TF difference.

Click to shrink...

I find the interpretation of my statements to always be interesting now that we're, like, 7 years on. It's sort of funny to look at what people are posting now and see hints of what I posted years ago.

I actually never tried to downplay the TFLOPS difference - that was a mathematical fact. What I was trying to point out was that things like fast ESRAM, and the fact that they had a higher-clocked CPU and GPU, as well as some efficiencies in DX, would close some of the difference in raw TFLOPS. I was suggesting that TFLOPS wasn't the only measurement of performance and when people talk about PS4 vs. XBO they always neglect any of the benefits that helped close the gap.

Turns out I was wrong. They didn't get the utilization of the ESRAM that was expected, and the CPU and GPU upclock likely *did* help make things better than they would have been otherwise. Meaning, without those improvements the differences would have been even greater.

I learned two big lessons from back then. First, was maybe be more cautious in how bold I make proclamations because the context of my statements get lost in time and I've said lots of times that quote didn't age well relative to what I was trying to say. It made people mad and I get that.

The second was that while "secret sauce" can help, the raw compute numbers ended up being super important because you can't expect every developer to utilize platform-specific features. There is a "lowest common denominator" effect that happens a lot when you're asking devs to build 8+ ports of a game.

The spirit of those lessons informed a lot of the Xbox One X. Instead of relying on cleverness and making developers use tricks to extract perf, they relied on a big GPU, lots of fast RAM, and continued improvements in DirectX calls. There were *also* secret sauce improvements but the biggest gains you see come from simply giving developers a lot of performance.

 

[vivftp]

I find the interpretation of my statements to always be interesting now that we're, like, 7 years on. It's sort of funny to look at what people are posting now and see hints of what I posted years ago.

I actually never tried to downplay the TFLOPS difference - that was a mathematical fact. What I was trying to point out was that things like fast ESRAM, and the fact that they had a higher-clocked CPU and GPU, as well as some efficiencies in DX, would close some of the difference in raw TFLOPS. I was suggesting that TFLOPS wasn't the only measurement of performance and when people talk about PS4 vs. XBO they always neglect any of the benefits that helped close the gap.

Turns out I was wrong. They didn't get the utilization of the ESRAM that was expected, and the CPU and GPU upclock likely *did* help make things better than they would have been otherwise. Meaning, without those improvements the differences would have been even greater.

I learned two big lessons from back then. First, was maybe be more cautious in how bold I make proclamations because the context of my statements get lost in time and I've said lots of times that quote didn't age well relative to what I was trying to say. It made people mad and I get that.

The second was that while "secret sauce" can help, the raw compute numbers ended up being super important because you can't expect every developer to utilize platform-specific features. There is a "lowest common denominator" effect that happens a lot when you're asking devs to build 8+ ports of a game.

The spirit of those lessons informed a lot of the Xbox One X. Instead of relying on cleverness and making developers use tricks to extract perf, they relied on a big GPU, lots of fast RAM, and continued improvements in DirectX calls. There were *also* secret sauce improvements but the biggest gains you see come from simply giving developers a lot of performance.

Click to shrink...

Nifty, thanks for the insights Albert :)

Great thread folks, lots of interesting info posted so far.

 

[Deleted member 224]

I find the interpretation of my statements to always be interesting now that we're, like, 7 years on. It's sort of funny to look at what people are posting now and see hints of what I posted years ago.

I actually never tried to downplay the TFLOPS difference - that was a mathematical fact. What I was trying to point out was that things like fast ESRAM, and the fact that they had a higher-clocked CPU and GPU, as well as some efficiencies in DX, would close some of the difference in raw TFLOPS. I was suggesting that TFLOPS wasn't the only measurement of performance and when people talk about PS4 vs. XBO they always neglect any of the benefits that helped close the gap.

Turns out I was wrong. They didn't get the utilization of the ESRAM that was expected, and the CPU and GPU upclock likely *did* help make things better than they would have been otherwise. Meaning, without those improvements the differences would have been even greater.

I learned two big lessons from back then. First, was maybe be more cautious in how bold I make proclamations because the context of my statements get lost in time and I've said lots of times that quote didn't age well relative to what I was trying to say. It made people mad and I get that.

The second was that while "secret sauce" can help, the raw compute numbers ended up being super important because you can't expect every developer to utilize platform-specific features. There is a "lowest common denominator" effect that happens a lot when you're asking devs to build 8+ ports of a game.

The spirit of those lessons informed a lot of the Xbox One X. Instead of relying on cleverness and making developers use tricks to extract perf, they relied on a big GPU, lots of fast RAM, and continued improvements in DirectX calls. There were *also* secret sauce improvements but the biggest gains you see come from simply giving developers a lot of performance.

Click to shrink...

I wasn't misinterpreting what you said. You said that other things (ESRAM, cpu clocks, gpu upclock, etc.) would close the gap.

I used your quote for 2 reasons:
1. You directly addressed the TF difference, which shows that if was a huge talking point back in 2013.
2. Despite what you said, that 40% TF difference was seen. We're seeing some of the exact same arguments you used this time, only now they're in favor of the PS5. The difference is smaller (20%), but it's undeniably there.

 

[z0m3le]

Small adendum: it would be possible to emulate operations with more bits by splitting and doing it with more calculations, but you would loose a lot of compute power.

so instead of calculating 1 value @ 16 bit (65 536‬) you could calculate 2x 8 bit , and would aditionally need to interpretthese 2 values together.

so from 8 (256)->16 (65 536) and from 16->32 (4 294 967 296‬) were huge. From 32 -> 64 (18 446 744 073 709 551 616) not so much but still relevant, and with 64 we have enough values for most variable you could need in game development, and for the view that would need more precision, these view you can do with multiple operations. 128 will come, but wohnt be as relevant (‭3,4028236692093846346337460743177e+38‬), and with 256... i honestly dont even expect it to come to the mainstream, except maybe if the aim is to work with 2 values parrallel on one core.



Ok, quoting wikipedia:

"256-bit processors could be used for addressing directly up to 2^256 bytes. Already 2^128 ( 128-bit) would greatly exceed the total data stored on Earth as of 2010, which has been estimated to be around 1.2 zettabytes (over 2^70 bytes).[1]"

256-bit computing - Wikipedia

en.wikipedia.org

With 128 there seem to be at least some uses:

128-bit computing - Wikipedia

en.wikipedia.org

Click to shrink...

Absolutely. I tried to highlight the move from 8bit through 32bit, as you say the limitations matter less, especially for gaming afterwards, and while 64bit or extentions mattered for more usable ram, that was one of the last hold outs for our modern computing and computers have gone 64bit for the last 2 decades (well since 2003).

How does one measure these "teraflops", through a flopometer?

Click to shrink...

ALUs*2*clock/1000, so for instance Switch has 256 ALU at 768mhz = 393gflops. Which is the number of floating point operations the GPU can do a second.

If you had 8GB of 100GB/s RAM and doubled the CPU clock to 2GHz, and overclocked the GPU to 2.5GHz, you'd be able to pretty much match the PS4 with the superior architecture in Switch. Of course those clocks just don't exist for the Switch, it would fail to even turn on, blow capacitors and whatnot.

Despite what you said, that 40% TF difference was seen. We're seeing some of the exact same arguments you used this time, only now they're in favor of the PS5. The difference is smaller (20%), but it's undeniably there.

Click to shrink...

Microsoft upped the clocks on the XB1 before launch, they are running 1825MHz, but if they pushed it to say 2100MHz instead, they would have 14TFLOPs, a similar 40% advantage.

If PS5 can push the architecture to 2.3GHz, there shouldn't be a reason why XSX couldn't increase their clocks, in fact they do say that the thing is designed to run cool and quiet even in the desert, but implementing the same AMD power solution to their SoC should allow the console to handle a much higher GPU clock as well.

 

[Albert Penello]

I wasn't misinterpreting what you said. You said that other things (ESRAM, cpu clocks, gpu upclock, etc.) would close the gap.

I used your quote for 2 reasons:
1. You directly addressed the TF difference, which shows that if was a huge talking point back in 2013.
2. Despite what you said, that 40% TF difference was seen. We're seeing some of the exact same arguments you used this time, only now they're in favor of the PS5. The difference is smaller (20%), but it's undeniably there.

Click to shrink...

Sorry bad context on my part. It was a general statement not directed towards you.

You're right to say that TFLOPS was a discussion point way back in 2013.

For the second - I think most people would agree that the majority of games were 1080p vs. 900p so you're right there was about 30% difference in pixel count. However, there were also games that were 1080p on both, and games that were 900p or less on both. So yes, for the majority of games there was 30% difference in pixel resolution, but given there was a 50% difference in TFLOPS and a huge difference in memory bandwidth, I think it's fair to also say that the XBO did better relative to what the specs would have suggested.

Now, for the record, I'm also of the mind that the difference between 10.3 vs. 12 will not be *that* enormous either assuming that the power-management across CPU and GPU works with no unexpected downsides.