Transistor count increases

[jpiniero]

I was curious about the transistor count increases for this gen, and sure enough nVidia increased counts by a (typical?) margin while AMD didn't do much. At the same time though the prices of nVidia's cards have gone up quite a bit while AMD kept the 480 in the same price range as the 380(X), or at least attempted to.

Tonga (5) -> Polaris 10 (5.7) 14% (btw Grenada is 6.2)

GM206 (2.9) -> GP106 (4.4) 51%
GM204 (5.2) -> GP104 (7.2) 38%
GM200 (8) -> GP102 (12) 50%

Don't have a number for Polaris 11.

 

[ShintaiDK]

Remember utilization went down as well. So the increase isn't as big as it seems.

 

[alcoholbob]

TSMC price per transistor supposedly is either about the same or only slightly lower per transistor at "16nm FF" (aka 20nm with 16nm Finfets...) than 28nm.

It's either a volume issue or a poor engineering issue, since Intel claims they are on track for standard cost per transistor decreases from 22nm all the way down to 10nm and below.

Nvidia's basically charging you for all the research they did into massively upping clock speeds, which is how they managed to make a GTX 1080 with only 25% more transistors than a RX 480 and yet is nearly double the speed. In terms of how cut down the GTX 1070 is, it's basically identical in silicon footprint to the RX 480 yet is about 70-80% faster.

 

[ShintaiDK]

It's either a volume issue or a poor engineering issue, since Intel claims they are on track for standard cost per transistor decreases from 22nm all the way down to 10nm and below.

A bit of both. Its due to the IC design and cost at the lower nodes. Volume and cost saving per unit is about if you ever get the ROI back. Its not a TSMC issue, its a customer issue so to say. Intels utilization rate on the transistors are higher than AMD and NVidia for example and any other in the world. Simply because they are able and willing to spend much more on the better IC design to get the cheaper transistors. But to regain that cost, you need the volume for it.

 

[lehtv]

In terms of how cut down the GTX 1070 is, it's basically identical in silicon footprint to the RX 480 yet is about 70-80% faster.

It's nowhere near that much faster. It's 50% faster in DX11 games, depending on resolution. Less in DX12 / Vulkan games since RX 480 is being held back by DX11 driver overhead. In terms of TFLOPs, it's only 6.46 / 5.5 = 1.174545... -> 17.5% faster.

 

[ShintaiDK]

It's nowhere near that much faster. It's 50% faster in DX11 games, depending on resolution. Less in DX12 / Vulkan games since RX 480 is being held back by DX11 driver overhead. In terms of TFLOPs, it's only 6.46 / 5.5 = 1.174545... -> 17.5% faster.

Vulkan using GCN Shader Extenstions while using SPIR-V for Nvidia isn't exactly a proper compare. Nor is sponsored titles with low performing/sabotaged paths.

1070 is 69% better in perf/watt than RX480.

 

[Flapdrol1337]

Nvidia's basically charging you for all the research they did into massively upping clock speeds, which is how they managed to make a GTX 1080 with only 25% more transistors than a RX 480 and yet is nearly double the speed. In terms of how cut down the GTX 1070 is, it's basically identical in silicon footprint to the RX 480 yet is about 70-80% faster.

I don't think you can really say a 1070 has the same footprint as a 480. A 1060 is only 15% smaller than a 480 and performs about the same, the 1060 has much better powerconsumption figures though.

 

[Lepton87]

ShintaiDK so you just casually switched from outright performance to performance per watt to defend your grossly inflated figure? Wouldn't it be just easier to admit that you slipped by a "few" % or are you a narcissist who can't admit to being wrong.


UPDATE:
Sorry, I didn't see that it wasn't you who said that 70-80% faster.

 

[alcoholbob]

The exact percentage is irrelevant. Point is AMD has a massive hill to climb in terms of performance per transistor which really points to a big die Vega being more of a GTX 1080 competitor than against GP102, unless they magically get 50% higher clock speeds out of fabbing from TSMC instead of GloFlo.

 

[jpiniero]

Big Vega (well it should) be faster than the 1080 if it does have 4096 cores and HBM2. Oh and fabbed at TSMC.

 

[lehtv]

The exact percentage is irrelevant. Point is AMD has a massive hill to climb in terms of performance per transistor which really points to a big die Vega being more of a GTX 1080 competitor than against GP102, unless they magically get 50% higher clock speeds out of fabbing from TSMC instead of GloFlo.

If the exact percentage is irrelevant, then maybe you shouldn't post exact percentages? The least you could do is own it and admit your numbers were very off.

 

[lehtv]

Vulkan using GCN Shader Extenstions while using SPIR-V for Nvidia isn't exactly a proper compare.

We are talking about actual performance, not about where that performance comes from. It's not a "proper compare" either that in DX11, NVIDIA cards use efficient NVIDIA drivers while AMD cards use high CPU overhead AMD drivers, but that's what we have to work with.

Nor is sponsored titles with low performing/sabotaged paths.

Again, that's real world performance for you. If you or alcoholbob want to support the idea that GTX 1070 is 70-80% faster then specify exactly in what situation that claim holds, because it certainly doesn't hold in actual real world gaming situations.

1070 is 69% better in perf/watt than RX480.

And that is 100% irrelevant

 

[alcoholbob]

If the exact percentage is irrelevant, then maybe you shouldn't post exact percentages? The least you could do is own it and admit your numbers were very off.

I'm not sure why you are so puritanical. I used some numbers based on a review of an aftermarket card which was boosting to nearly 1.9GHz, obviously every card varies out of the box. There's no absolute performance delta since every card, even reference cards, can vary as much as 150+MHz in boost clocks out of the box.

Secondly, whether its a 53% delta (as per Techpowerup), or a 70% delta (of a card with a higher boost clock out of the box), the point still stands that a 25% cut down GP104 (which has functionally 5.4 billion transistors) is outperforming the RX480 by a wide margin, which has 5.7 billion transistors.

You are tied up on semantics. The point I'm trying to make is what this means for Vega.

After all, AMD still is limited by the laws of manufacturing, die size limits, and of course TDP. They have to make gargantuan improvements in perf per watt and clock speed to make Vega relevant at a 300W TDP. Now it might be a 375W card on water at 1600MHz, I dunno. But from the looks of it that would be the least it would take to fall in somewhere between 1080 and Titan X. I'm not sure they would want to go that route since I imagine profit margin would be extremely low on such a SKU whereas Nvidia is making a killing without having to push their chips that far or have to build expensive power delivery on their reference board just to compete.

If that's the case AMD might decide to have Vega compete with the mid end, not high end with Nvidia. What that means for the consumer is possibly no competitive downward pressure on prices for Nvidia GPUs (particularly no incentive to have a 1080 Ti priced like the 980 Ti, or cause much if any drop in the price of 1070/1080) until the next generation of cards in 2H 2017.

 

[lehtv]

*snip*

So now it's 70% instead of 70-80%? Maybe next it's 60-70%, then 50-60% right? If you're trying to back up your original claim about the performance gap then please be transparent about it and post the source, otherwise as I said, maybe you shouldn't post exact percentages.

Whether there's an absolute performance delta is irrelevant. Unless otherwise specified, generalized claims like "GTX 1070 is 70-80% faster than RX 480" are implicitly about averages, not about absolutes and certainly not about individual sample cards.

Semantics is concerned with the meaning of words, i.e. definitions. This discussion has absolutely nothing to do with semantics as far as I can tell.

I really don't care what other points you're trying to make, I've never disputed those. I pointed out your 70-80% performance gap was untrue, and you're trying to avoid having to admit it was a mistake or an exaggeration by changing the topic. Why can't you just say "You're right, it's not 70-80% faster, my mistake." Is that really so hard?

 

[tviceman]

I was curious about the transistor count increases for this gen, and sure enough nVidia increased counts by a (typical?) margin while AMD didn't do much. At the same time though the prices of nVidia's cards have gone up quite a bit while AMD kept the 480 in the same price range as the 380(X), or at least attempted to.

Tonga (5) -> Polaris 10 (5.7) 14% (btw Grenada is 6.2)

GM206 (2.9) -> GP106 (4.4) 51%
GM204 (5.2) -> GP104 (7.2) 38%
GM200 (8) -> GP102 (12) 50%

Don't have a number for Polaris 11.

Compare transistor/mm2. AMD's increase on a percentage basis is close to the same as Nvidia's. Saying RX480 is Tonga's replacement isn't entirely accurate because Polaris is a "Pitcarin" class chip in die size and product positioning among current gen (finfet) products. Tonga was a Tahiti replacement with added features.

 

[Erenhardt]

the point still stands that a 25% cut down GP104 (which has functionally 5.4 billion transistors) is outperforming the RX480 by a wide margin, which has 5.7 billion transistors.

So the point is wrong again.
How many of those 25% transistors you slashed come from memory controllers and IO?

Anyway, this is pointless actually. There is more to performance than just gaming. How is FP64 going with 1070?

 

[alcoholbob]

I really don't care what other points you're trying to make, I've never disputed those. I pointed out your 70-80% performance gap was untrue, and you're trying to avoid having to admit it was a mistake or an exaggeration by changing the topic. Why can't you just say "You're right, it's not 70-80% faster, my mistake." Is that really so hard?

So you just want acknowledgement on the internet for self esteem purposes? You keep quoting an irrelevant minor point in my post because you want an apology so badly. Seriously, there's bigger goals to go for in this world than a gold star for being acknowledged by somebody on their internet.