Support NeoGAF

[Ascend]

They are just reporting the basic mathematics, as does everyone.

What is left to be seen is how efficiently the architecture utilizes the stream processors. Historically, nVidia has had the edge in this area with AMD having higher sp counts (and TF numbers) but being on the short end of performance.

Not really. Their SM in Ampere can do either 128 FP32 OR 64 FP32 + 64 INT32. The only time you will reach the advertised TF is when there are ZERO INT32 operations. At any other time, the TF is literally cut in half for the SM doing the INT operation.

 

[DaGwaphics]

It would be fine if they just said FLOPS, but they made it about CUs.
I'd call it lying.
CU is more than just number cruncher, just that it could doe more ops per clock doesn't turn it into 2 CUs.

With Turing, each actual shader contained 1 FP32 pipeline and 1 INT-32 pipeline. With Ampere, they're shifting to 1 FP32 pipeline and a second pipeline capable of both INT and FP32 - but only one at a time.

True, but you know AMD would also be going with the biggest number they could in the same scenario.

 

[DaGwaphics]

Not really. Their SM in Ampere can do either 128 FP32 OR 64 FP32 + 64 INT32. The only time you will reach the advertised TF is when there are ZERO INT32 operations. At any other time, the TF is literally cut in half for the SM doing the INT operation.

I'm assuming they are looking to push INT32 almost entirely on to Tensors now, or at least that's what I would expect from looking at the slides. Or is that not a possible workflow, this is all above my level of understanding of gpu functionality. LOL

 

[Ascend]

True, but you know AMD would also be going with the biggest number they could in the same scenario.

True. Still doesn't make it right. Just to give an idea what this actually is. It is the equivalent of Microsoft saying their XSX has 920GB/s of bandwidth.

 

[llien]

Not, it was a bad idea even when released. Just drop the price of Xbox One X? They should've done what Sony did, regular version, and disk-less version of XSX. Makes more sense.

44% more CUs, bigger chip. XSeX is essentially "pro" version.
System with one third of its power, aimed at 1080p rendering kinda makes sense, if very cheap.

 

[iHaunter]

44% more CUs, bigger chip. XSeX is essentially "pro" version.
System with one third of its power, aimed at 1080p rendering kinda makes sense, if very cheap.

I guess we can't say yet until we know the price.

 

[Krisprolls]

I'm assuming they are looking to push INT32 almost entirely on to Tensors now, or at least that's what I would expect from looking at the slides. Or is that not a possible workflow, this is all above my level of understanding of gpu functionality. LOL

The other explanation is they just wanted to show a big Teraflops number regardless of reality, just for marketing purpose. But no... that would be too simple.

 

[DaGwaphics]

True. Still doesn't make it right. Just to give an idea what this actually is. It is the equivalent of Microsoft saying their XSX has 920GB/s of bandwidth.

That's not a good comparison, IMO. The RTX could run a benchmark/burn test utilizing only fp calculations and hit those marks. Is that indicative of anything real-world, maybe not, but it could be done.

560GB/s is the most that's ever coming out of the XSX memory chips, no matter how stilted you made the tests.

 

[WatDat]

Well... the Switch is still selling like hot cakes so I would say yes. There’s a market for it for sure.

Just wanna top this golden reply. #SSD

 

[TheGreatWhiteShark]

hmmmm OP, let's calculate together.....
30tf thingy targets the game at 8k
4tf thingy targets the game at 1080p

what do you think ?

 

[Ascend]

That's not a good comparison, IMO. The RTX could run a benchmark/burn test utilizing only fp calculations and hit those marks. Is that indicative of anything real-world, maybe not, but it could be done.

560GB/s is the most that's ever coming out of the XSX memory chips, no matter how stilted you made the tests.

Fair enough. But hopefully you understand my point. Maybe, an "up to" would have been appropriate to the TF specification.

 

[DaGwaphics]

Fair enough. But hopefully you understand my point. Maybe, an "up to" would have been appropriate to the TF specification.

I hear you. But, TF has always been a theoretical peak that is almost impossible to reach (if not impossible). Useful for comparing GPUs in the same lineage, but not much else. With that in mind, an "up to" should really be included with all the sec sheets, instead it's just taken to be understood that these are theoretical peaks.

 

[tkscz]

30TF GPU

This figure is highly misleading.
E.g. 3070 is "20TF GPU" but is touted as merely "=2080Ti", which is, wait for it, 13.7TF.

nVidia did do this before with the Xavier models, combing the CUDA TFs with the Tensor TFs to make them seem a lot faster.

However that isn't the case as they separate the CUDA and Tensor TFs

Also, same image says faster than, not equal to.

 

[RayHell]

nVidia did do this before with the Xavier models, combing the CUDA TFs with the Tensor TFs to make them seem a lot faster.

However that isn't the case as they separate the CUDA and Tensor TFs

Also, same image says faster than, not equal to.

Jensen himself said the flagship RTX 3080 offers “twice the performance of the 2080,”
2x 10.1TF
But marketed as 30TF

 

[AGRacing]

So many of us screaming from the rooftops for months that TFs dont compare between architectures, generations of architectures or brands... Apples to Oranges to Watermelons.... And so many of you are so completely deaf to this..... It's getting really old.

Advice to anyone trying to get this across... Just give up.


One X marketing team.... This is your legacy...

6TF4LIFE would be my license plate if I designed that system.

 

[RayHell]

So many of us screaming from the rooftops for months that TFs dont compare between architectures, generations of architectures or brands... Apples to Oranges to Watermelons.... And so many of you are so completely deaf to this..... It's getting really old.

Advice to anyone trying to get this across... Just give up.


One X marketing team.... This is your legacy...

6TF4LIFE would be my license plate if I designed that system.

Remind me of the flat earthers. There's no way to educate those people.
"Jensen said it's 2x time the performance." No It'S ThReE TiMe tHe PeRFoRMAnCe !!!
¯\_(ツ)_/¯

 

[Ascend]

So many of us screaming from the rooftops for months that TFs dont compare between architectures, generations of architectures or brands... Apples to Oranges to Watermelons.... And so many of you are so completely deaf to this..... It's getting really old.

Advice to anyone trying to get this across... Just give up.


One X marketing team.... This is your legacy...

6TF4LIFE would be my license plate if I designed that system.

All true. But generally, you expect the IPC ("and thus FLOP efficiency" if you will) to increase with subsequent products from the same vendor. Right now, there is at least a 33% decrease in IPC when you compare Ampere to Turing. Nobody expects that. It makes it quite deceiving. I mean, a card that is advertised as 30TF but actually performs like a 20TF... It should be brought to light. We can't expect everyone to know what a TF is, let alone understand it. If these things are talked about, the layman has a higher chance of bumping into these discussions and know better what he's getting.

I know that I will be pissed off if a lower priced more efficient Navi with 25TF beats the 30TF nVidia equivalent by 25%, but people buy the nVidia card because of the paper specs. Wouldn't be the first time. Why it would piss me off? Because market share influences developer resource allocation. And if AMD's market share stays low despite having the superior card, everyone that is knowledgeable and did go for the actual better performing card with lower TFLOPS, will get the short end of the stick.

 

[Ten_Fold]

If I recall the ps4 got like 1.5TF or something an was underpowered when it launched, but we got games looking great like U4, GoT. Having a bunch of power is cool but most casuals could careless. The games will look great for the next 5-7years.

 

[vaibhavpisal]

If I recall the ps4 got like 1.5TF or something an was underpowered when it launched, but we got games looking great like U4, GoT. Having a bunch of power is cool but most casuals could careless. The games will look great for the next 5-7years.

I like this post.

Its the development efforts that count. That's why Nintendo games evoke such passion/reactions.

You can have face melting graphics for a brief period but it tends to get old with in a month. Then upgrade your card again to get similar high. Its endless cycle.