Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[jpiniero]

The same as what?

Turing.

 

[BenSkywalker]

You genuinely think that Ampere is going to have the same, or close, performance per watt as Turing?

 

[maddie]

Turing.

Is Samsung's process so poor?

 

[jpiniero]

Is Samsung's process so poor?

That's TBD, but you are talking about a design originally intended on a much better node. They are for sure pushing the frequencies hard, perhaps out of the efficiency range.

Could also be that the power that the memory draws is going up too, and that is playing a part in it.

 

[Thala]

Turing.

Thats totally unrealistic to assume. Even only just shrinking Turing would give you a huge perf/power boost without doing anything else.

 

[jpiniero]

Thats totally unrealistic to assume. Even only just shrinking Turing would give you a huge perf/power boost without doing anything else.

You would think so, but that appears to not be the case.

 

[Thala]

You would think so, but that appears to not be the case.

This is based on what information precisely? Do we have both reliable power and performance numbers of upcoming Ampere GPUs?
Common sense says the perf/Watt goes up, so you must have pretty reliable information to think otherwise?

 

[uzzi38]

Thats totally unrealistic to assume. Even only just shrinking Turing would give you a huge perf/power boost without doing anything else.

While I certainly don't think we'll see the same or worse perf/W - rather we should see a definite uptick - you have to realise that if a 350W power draw is legit it's completely possible that Ampere may be pushed past it's perf/W sweetspot in terms of clocks, and so the uptick in perf/W may not be as huge as you'd first assume.

 

[jpiniero]

This is based on what information precisely? Do we have both reliable power and performance numbers of upcoming Ampere GPUs?
Common sense says the perf/Watt goes up, so you must have pretty reliable information to think otherwise?

The rumors about the TDP. Seems the 3080 Ti is going to be roughly 30-40% faster in raster games at 320-340W TDP.

 

[Thala]

The rumors about the TDP. Seems the 3080 Ti is going to be roughly 30-40% faster in raster games at 320-340W TDP.

I would not go with rumors if they are in contradiction with common sense and prior evidence. NVidia must have put in serious flaws in Ampere in order to eat up all the power gain.
This 320-340W for 30-40% faster 3080 TI are both based on leaked numbers from within NVidia?

 

[jpiniero]

I would not go with rumors if they are in contradiction with common sense and prior evidence. NVidia must have put in serious flaws in Ampere in order to eat up all the power gain.
This 320-340W for 30-40% faster 3080 TI are both based on leaked numbers from within NVidia?

There was a 3dmark or something leak not too long ago. Plus that's roughly what you would expect from a gen-gen upgrade from nVidia.

Remember, this was supposed to be on Samsung 7 nm.

 

[uzzi38]

There was a 3dmark or something leak not too long ago. Plus that's roughly what you would expect from a gen-gen upgrade from nVidia.

Remember, this was supposed to be on Samsung 7 nm.

That 30% uplift over a stock 2080Ti FE was using 12Gbps memory instead of the final spec. The clocks were definitely reporting correctly. Time Spy Graphics test 2 is pretty memory heavy, so if I were to take a guess (and this is purely just a guess) a score more accurate to final perf would be roughly 40-45% over a stock 2080Ti FE, depending on how memory bound the uArch is. I am accounting for the reported clocks not being final clocks.

Either way though, a stock 2080Ti FE pulls around 275W on average from memory, not the rated 250W TBP, so even with all of that you're looking at around a 20% perf/W uplift, give or take a couple of percent.

If Ampere is actually on 8LPP and is clocked well past optimal V/f, I'd call that a win honestly.

 

[Thala]

There was a 3dmark or something leak not too long ago. Plus that's roughly what you would expect from a gen-gen upgrade from nVidia.
Remember, this was supposed to be on Samsung 7 nm.

Was the power number of precisely the card you saw on 3dmark leaked?
Note that i have no problem with the 30-40% faster assumption - thats what you get by utilizing the larger gate-count you have at your disposal. In this case you would get the full power advantage of the new node.
The only way you stay flat in perf/Watt is, when you try to get the performance not be increasing ALUs but by increasing frequency and voltage. Again very unlikely - because that would be a totally stupid approach.

Either way though, a stock 2080Ti FE pulls around 275W on average from memory, not the rated 250W TBP, so even with all of that you're looking at around a 20% perf/W uplift, give or take a couple of percent.

Indeed assuming a 20% perf/watt uplift is around the minimum i would expect for architecture on the new node.

 

[jpiniero]

The only way you stay flat in perf/Watt is, when you try to get the performance not be increasing ALUs but by increasing frequency and voltage. Again very unlikely - because that would be a totally stupid approach.

That's pretty close to what they are doing. The shader count is only 20% more because they also wanted to get the die size down (TU102 is 754 mm2). They would have hit their perf and perf/w target just fine had it been on SS7.

 

[Glo.]

The only way you stay flat in perf/Watt is, when you try to get the performance not be increasing ALUs but by increasing frequency and voltage. Again very unlikely - because that would be a totally stupid approach.

That is solely dependend on how your competition performs.

If AMD does have much better GPU arch, than Nvidia anticipated, then there is absolutely no strangeness to the idea, that Nvidia was forced to clock their GPUs way past the voltage/frequency curve. In this light think about why there are rumors about RTX 3080 SKU being on 102 die, not on 104, like it was in previous years.

 

[jpiniero]

In this light think about why there are rumors about RTX 3080 SKU being on 102 die, not on 104, like it was in previous years.

Actually the 3080 was originally rumored to be on a full? GA103 (3840 shaders). GA103 appears to be cancelled and they are now using GA102 instead (4352). GA104 appears to be 3072, which is actually the same as TU104.

IIRC there was at one time a GA101 too, but perhaps that was cancelled when SS7 got cancelled.

 

[BenSkywalker]

No-one has considered the other alternative? NVidia is out for blood and looking to annihilate AMD so they are pushing wattage higher?

"Poor Volta"

AMD's hype is saying 50% better performance per watt them RDNA. That would put a 300 watt card roughly on par with those 3DMark scores you are talking about which most people I saw thought was the 3080, not Ti/3090.

The problem is that AMD has been hysterically dishonest with their hype in the GPU segment in the past, they threw down poor Volta years ago and have since introduced three new tiers none of which has matched the hype they claimed.

This time they claimed "nVidia killer", perhaps they just ticked JHH off enough that he's out to humiliate them this generation and reinforce the widely accepted view that AMD just can't compete.

Do I think the above is going to happen? No, but it's far more likely than what you people are setting yourselves up for.

Go back and check the Twitter claims again, the same guy saying it's 100% SS8 was saying no SS8 all SS7 previously. Then he swapped to SS8 and now he comes out with this 600 watt power plug? Sounds like someone is mole hunting.

Even if we assume this is a SS 8nm, SS 10nm was comparable to TSMC's 10nm, and supposedly this is on 8nm EUV which is a very large step up over "12nm". That would still be a full node drop and simply looking at history a full node drop for nVidia has resulted in *massive* gains every time. Meanwhile AMD is claiming a 50% performance gain per watt on a slightly refined node- something they have never come close to demonstrating they could do.

Despite all of this, you guys are truly trying to convince yourselves that your scenario has a serious chance to end up right?

Upside to all of this, everyone has pages worth of posts so when these products hit we can all see exactly how good everyone is at speculating.

 

[Kenmitch]

Despite all of this, you guys are truly trying to convince yourselves that your scenario has a serious chance to end up right?

The excessive drinking of kool-aid tends to alter one's perception of reality.

Wasn't Raja the hype master?

 

[jpiniero]

Even if we assume this is a SS 8nm, SS 10nm was comparable to TSMC's 10nm, and supposedly this is on 8nm EUV which is a very large step up over "12nm". That would still be a full node drop and simply looking at history a full node drop for nVidia has resulted in *massive* gains every time. Meanwhile AMD is claiming a 50% performance gain per watt on a slightly refined node- something they have never come close to demonstrating they could do.

SS8 is not EUV.

We don't know what the competitive product AMD is targeting at with the biggest Big Navi. I am guessing that it's at least faster than the 2080 Ti in raster.

 

[BenSkywalker]

Normal SS8 is a LPP node, used for mobile and automotive chips, low power. This is the node you all think is going to create a 600 watt GPU......?

 

[jpiniero]

Normal SS8 is a LPP node, used for mobile and automotive chips, low power. This is the node you all think is going to create a 600 watt GPU......?

No... They will stick to the max of dual 8 pin (375 W)

 

[FaaR]

No-one has considered the other alternative? NVidia is out for blood and looking to annihilate AMD so they are pushing wattage higher?

Please... Nobody's going to be annihilating anyone in the GPU space. That's not how it works. Even when AMD was at their weakest a few years ago they weren't annihilated, and now that their fortunes are rising they certainly won't be.

Same as when NV stumbled with the overly large and expensive GTX 280 (which all ended up dying due to bumpgate later on btw) vs AMD's faster and wayyy cheaper Radeon HD 4890, NV wasn't annihilated.

It's not in gamers' interests any GPU manufacturer gets annihilated btw. Competition is the root to all prosperity in the consumer space. Otherwise all you'll get is slow and expensive crap, like it was in the Soviet Union.

 

[MrTeal]

Hm, pretty sure I saw one of the images (linked on TPU's news post page about this connector IIRC) stating 9A per pin? If all of the pins in the connector are dedicated to power delivery, that's almost 650W actually.

Current rating was taken from the datasheet. You can get 9A per pin with 16 gauge wire, but with 20 gauge it would be 4A with all circuits populated.
It appears FCPowerUp has updated their article though. It still says it's limited to 20 gauge, but has updated it with the images from TPU showing mandated 16 gauge wire. Maybe the original article didn't translate well?

 

[FaaR]

4A per pin doesn't make much sense. That's less power than two standard 8-pin today. It just creates confusion, and is inferior to boot. So why would they go with a connector like that?

Btw, totally unrelated... Gauge must be the dumbest measuring unit ever. It makes even less sense than the fahrenheit temperature scale. How do you get a thicker wire than 1 gauge; go negative...?

 

[Glo.]

Who cares about whether AMD or Nvidia wins?

All that matters is value of a product. Its price to performance ratio. Nothing defines a product more, than this. Increased competition from both companies, at least, gives hope that this generation we will see better products in price/performance ratio increases, than we had with previous generation.