Re-litigating Radeon market performance *spawn

[DavidGraham]

I always wonder how the 7000/GCN1 series would have looked if Nvidia didn't split up Kepler like they did and just released the full lineup in 2012

The 7970 appeared in Dec 2011.
GTX 680 appeared in March 2012.
Tesla K20 using the big Kepler die appeared in Nov 2012.
GTX Titan appeared in Feb 2013.
GTX 780 appeared in May 2013.
R9 290X appeared in Oct 2013.
GTX 780Ti appeared in Nov 2013.

NVIDIA was surprized the 7970 was only successful against their middle class die (GK104), so they only launched the high end die (GK110) 6 months later for their compute lineup in the form of Tesla K20, then they staggered the launch even more by waiting another 6 months and then releasing the GTX Titan using a slightly cut down die, then after 3 months they cut it even more and released the GTX 780, and only after the release of 290X did NVIDIA release the full GK110 die to consumers in the form of 780Ti and Titan Black.

 

[Bondrewd]

NVIDIA was surprized the 7970 was only successful against their middle class die (GK104), so they only launched the high end die (GK110) 6 months later for their compute lineup in the form of Tesla K20, then they staggered the launch even more by waiting another 6 months and then releasing the GTX Titan using a slightly cut down die,

that's bull, yielding chungus dies on early 28nm HKMG was just a pain.
Either way Hawaii was a tiny ass die for what it did (mere 438mm^2, I think).

 

[Seanspeed]

The 7970 appeared in Dec 2011.
GTX 680 appeared in March 2012.
Tesla K20 using the big Kepler die appeared in Nov 2012.
GTX Titan appeared in Feb 2013.
GTX 780 appeared in May 2013.
R9 290X appeared in Oct 2013.
GTX 780Ti appeared in Nov 2013.

NVIDIA was surprized the 7970 was only successful against their middle class die (GK104), so they only launched the high end die (GK110) 6 months later for their compute lineup in the form of Tesla K20, then they staggered the launch even more by waiting another 6 months and then releasing the GTX Titan using a slightly cut down die, then after 3 months they cut it even more and released the GTX 780, and only after the release of 290X did NVIDIA release the full GK110 die to consumers in the form of 780Ti and Titan Black.

That's kinda my point. Nvidia knew they had a winner on their hands, yet delayed that hand. I get the reasoning, especially because it allowed them to let 28nm mature before they released GK100, but I think it might have accelerated the 'AMD is behind' narrative that really only started with Fiji. Perception is everything, even if AMD were actually still pretty competitive in terms of performance per mm².

 

[Bondrewd]

yet delayed that hand

They didn't, yielding a chungus 561mm^2 die on early 28nm was PITA.
Which is why the OG K20 in the likes of Titan (the supercomputer, not the halo GPU named after it) was severely castrated relative to the full die.

but I think it might have accelerated the 'AMD is behind' narrative that really only started with Fiji.

It started with Vega.

 

[no-X]

NVIDIA was surprized the 7970 was only successful against their middle class die (GK104)…

Sorry, that's a complete nonsense. GK104 was projected as a high-end product and never was intended to be launched for mainstream price, as some "sources" suggested at the time of its launch. I have it confirmed personally by an Nvidia employee.

 

[Seanspeed]

They didn't, yielding a chungus 561mm^2 die on early 28nm was PITA.
Which is why the OG K20 in the likes of Titan (the supercomputer, not the halo GPU named after it) was severely castrated relative to the full die.

It started with Vega.

I doubt it was any hardship, just yielded worse.

And no, it was definitely Fiji. It had the unfortunate pleasure of releasing basically RIGHT as next gen games became the norm, so while it fared well enough in reviews because most of the games benched were earlier cross-gen titles, it very quickly started to suffer from its 4GB limit as any kind of high end part. Plus it still came a deal later than the Nvidia equivalent it was competing against(a recurring theme).

 

[Bondrewd]

I doubt it was any hardship, just yielded worse.

Yields equate to hardship.
You ship what you can yield and even zillion margin K20 was very chopped (minus tons of shaders and minus the 64b worth of memory interface).

And no, it was definitely Fiji. It had the unfortunate pleasure of releasing basically RIGHT as next gen games became the norm, so while it fared well enough in reviews because most of the games benched were earlier cross-gen titles, it very quickly started to suffer from its 4GB limit as any kind of high end part. Plus it still came a deal later than the Nvidia equivalent it was competing against(a recurring theme).

It was very much good enough to not damage the perception much.
But then Vega happened and now the reddit plebs expect every AMD GPU release to be another Vega.
Too bad!

 

[Seanspeed]

Yields equate to hardship.
You ship what you can yield and even zillion margin K20 was very chopped (minus tons of shaders and minus the 64b worth of memory interface).

It was very much good enough to not damage the perception much.
But then Vega happened and now the reddit plebs expect every AMD GPU release to be another Vega.
Too bad!

Yields hardship is cost.

You are fooling yourself if you dont think Radeon's reputation wasn't already taking a hit with Fiji. That absolutely happened, even if you want to memory hole it. Again, the timing was just really bad, given we were just starting to get true next gen titles taking advantage of the larger memory pools of the XB1/PS4. Fiji was still performant enough otherwise, but having to reduce settings for a supposed 'high end' part hurt it.

 

[Bondrewd]

Yields equate to cost, not hardship.

No they equate to what you can ship and at what frequency.

You are fooling yourself if you dont think Radeon's reputation wasn't already taking a hit with Fiji

Fiji was a nothingburger relative to Vega.

 

[DavidGraham]

Which is why the OG K20 in the likes of Titan (the supercomputer, not the halo GPU named after it) was severely castrated relative to the full die.

It wasn't, the K20c launched with 2496 cores out of 2880, and the K20x launched with 2688 out of 2880. Both launched on Nov 2012.

If NVIDIA wanted a GTX 780 and 780Ti would easily be available on Nov 2012 too, with Titan like prices of course.

 

[Seanspeed]

No they equate to what you can ship and at what frequency.

Fiji was a nothingburger relative to Vega.

By mid 2012, 28nm was in decent enough shape if Nvidia really wanted take the hit on yields.

And again, you can pretend Fiji wasn't largely considered lackluster by gamers back in 2015, but it was. It just was. We were all there and know this. It had some cool novelties like the Fury Nano and all, but it wasn't enough.

 

[DavidGraham]

Sorry, that's a complete nonsense. GK104 was projected as a high-end product and never was intended to be launched for mainstream price, as some "sources" suggested at the time of its launch. I have it confirmed personally by an Nvidia employee.

GK104 can and could never be a high end die for the simple fact that 6 months later NVIDIA had a much bigger die (GK110) available on the market.

NVIDIA was setting around doing nothing with that die, that they thought they could create a new consumer SKU with it, at a high price tag, and they called it "Titan", before they milked it again a few months later with the more cut down GTX 780 SKU.

NVIDIA practically squeezed the GK110 so hard, it lasted them two generations of AMD GPUs (HD 7970 and HD 290X).

 

[Bondrewd]

It wasn't, the K20c launched with 2496 cores out of 2880, and the K20x launched with 2688 out of 2880. Both launched on Nov 2012.

Yeah and guess what they used to cost and how they clocked.

If NVIDIA wanted a GTX 780 and 780Ti would easily be available on Nov 2012 too, with Titan like prices of course.

You need to clock these.

By mid 2012, 28nm was in decent enough shape if Nvidia really wanted take the hit on yields.

It wasn't.
Phone vendors weren't even using the HKMG variant till mid'13 for obvious reasons.
Like SD600 shipped 28LPm lmao.

And again, you can pretend Fiji wasn't largely considered lackluster by gamers back in 2015, but it was. It just was. We were all there and know this. It had some cool novelties like the Fury Nano and all, but it wasn't enough.

It was ok, Vega was the one that tanked the perception.
The public perception of stuff even in Polaris time was OK.
Vega was the one that kinda killed it (even if AMD itself is to blame for that given the marketing).

 

[Bondrewd]

also none of this is relevant to RDNA4. yay.

 

[Seanspeed]

Yeah and guess what they used to cost and how they clocked.

You need to clock these.

It wasn't.
Phone vendors weren't even using the HKMG variant till mid'13 for obvious reasons.
Like SD600 shipped 28LPm lmao.

It was ok, Vega was the one that tanked the perception.
This forum is a bubble, and a very dumb one at that.
The public perception of stuff even in Polaris time was OK.
Vega was the one that kinda killed it (even if AMD itself is to blame for that given the marketing).

AMD were using 28nm even with a midrange, 300mm²+ die in 2011, so no, it wasn't in such dire straits.

I'm also not talking about this forum. This forum is a bubble, that's the most correct thing you've ever said, but I'm telling you that this poor perception about Fiji was widespread.

You'd be right that Vega was the one that properly solidified the downward perception, but it didn't start it.

 

[Bondrewd]

AMD were using 28nm even with a midrange, 300mm²+ die in 2011, so no, it wasn't in such dire straits.

Yeah it was, which is why Tahiti was priced at a mighty fine $549 instead of the usual Cayman/Cypress sub 400 bucks.

but I'm telling you that this poor perception about Fiji was widespread.

Not poor in a way Radeon currently is.

You'd be right that Vega was the one that properly solidified the downward perception, but it didn't start it.

From that POV the downward perception started with Tahiti being priced at $549.
Or Hawaii ref being what it was (whirr).
Or Cayman being mid and stack shifting in weird ways.
Hard to pinpoint until Vega, which did a number.

 

[Seanspeed]

Yeah it was, which is why Tahiti was priced at a mighty fine $549 instead of the usual Cayman/Cypress sub 400 bucks.

Thanks for proving this was an issue of cost. lol

 

[Bondrewd]

Thanks for proving this was an issue of cost. lol

It was the issue of yields and clocks.
Which is why GHz edition exists.
Pre-2013 28nm being good is some funny revisionist history, misses on all the good sweat and tears phone SoC vendors had in 2012.

 

[trinibwoy]

I don’t remember Fiji being bad per se but more of a novelty that didn’t see widespread adoption. There was more hype for Vega and therefore more room for disappointment.

 

[del42sa]

I don’t remember Fiji being bad per se but more of a novelty that didn’t see widespread adoption. There was more hype for Vega and therefore more room for disappointment.

I remember it being quoted by Joe Macri as "overclock dream" moreover it barely clocked above 1Ghz