[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[maddie]

The idea that Navi won't achieve 2080 levels of Performance strikes me as being silly. Nvidia and AMD knew the approximate levels of Performance would exist at this time years ago when Navi's engineering began. Unless AMD completely cucked it up, their targets would have been at least that level of Performance. As someone else mentioned, AMD already reached that neighbourhood with Radeon VII on the generation prior to Navi. Not equaling the Radeon VII with Navi would be an epic fail. Could happen, I suppose, but I seriously doubt it.

That said, I don't know if AMD was expecting Ray Tracing to exist at this time, for Gaming anyways. If that's the case, I wouldn't be surprised if AMD can't match the 2080 for that task.

The latest PS5 info quite clearly claims Navi tech and RT tech, so RT was definitely in the design briefs for Navi as I can't imagine a custom APU having it exclusively, in light of the fundamental nature of RT to the rendering pipeline.

 

[Head1985]

The idea that Navi won't achieve 2080 levels of Performance strikes me as being silly. Nvidia and AMD knew the approximate levels of Performance would exist at this time years ago when Navi's engineering began. Unless AMD completely cucked it up, their targets would have been at least that level of Performance. As someone else mentioned, AMD already reached that neighbourhood with Radeon VII on the generation prior to Navi. Not equaling the Radeon VII with Navi would be an epic fail. Could happen, I suppose, but I seriously doubt it.

That said, I don't know if AMD was expecting Ray Tracing to exist at this time, for Gaming anyways. If that's the case, I wouldn't be surprised if AMD can't match the 2080 for that task.

Did polaris matched GTX1080?Polaris matched AMD 390 at launch and was 33% slower than furyx at 1440p.It will be similar with navi-vega 56(maybe 64 if its really good) performance and slower than radeon7.
https://www.computerbase.de/2016-06...hnitt_benchmarks_in_1920__1080_und_2560__1440

 

[Dribble]

The latest PS5 info quite clearly claims Navi tech and RT tech, so RT was definitely in the design briefs for Navi as I can't imagine a custom APU having it exclusively, in light of the fundamental nature of RT to the rendering pipeline.

Not convinced, no one thought Nvidia would actually make ray tracing work until they launched cards with hardware support, but the moment they did it was obvious the next gen consoles would have to have it as Sony/MS couldn't risk the other one adding RT support and them missing out.
Hence I think it's more likely AMD has had to add something late on in the design process. Pretty likely it won't be as ground up a design as Nvidia - more just adding enough support to normal shader cores to make the RT calculations much faster, and not even attempting the tensor cores (instead relying on shaders to denoise).

 

[maddie]

Did polaris matched GTX1080?Polaris matched AMD 390 at launch and was 33% slower than furyx at 1440p.It will be similar with navi-vega 56(maybe 64 if its really good) performance and slower than radeon7.
https://www.computerbase.de/2016-06...hnitt_benchmarks_in_1920__1080_und_2560__1440

I remember the pre-Polaris launch hype. Koduri always said that Polaris would satisfy the entry requirements for VR, which was a 290 class product.

Vega appeared to be a bust for gaming as they clearly expected much more than delivered.

This time Lisa Sui says that AMD will compete up to the high end. Does this mean x080 or x080Ti products, I don't know, but the minimum is the x080.

They do state their intentions in broad terms but we do not always hear.

 

[maddie]

Not convinced, no one thought Nvidia would actually make ray tracing work until they launched cards with hardware support, but the moment they did it was obvious the next gen consoles would have to have it as Sony/MS couldn't risk the other one adding RT support and them missing out.
Hence I think it's more likely AMD has had to add something late on in the design process. Pretty likely it won't be as ground up a design as Nvidia - more just adding enough support to normal shader cores to make the RT calculations much faster, and not even attempting the tensor cores (instead relying on shaders to denoise).

If you think that RT can be added in such a short period of time then so be it.

 

[sandorski]

Did polaris matched GTX1080?Polaris matched AMD 390 at launch and was 33% slower than furyx at 1440p.It will be similar with navi-vega 56(maybe 64 if its really good) performance and slower than radeon7.
https://www.computerbase.de/2016-06...hnitt_benchmarks_in_1920__1080_und_2560__1440

No it didn't, but that was a one time event. Generation after Generation AMD has beat, matched, or just fallen short of Nvidia's top chip. When Polaris and Vega were being developed AMD was in dire circumstances, unable to invest enough to release Vega alongside Polaris for a full lineup.

 

[coercitiv]

The idea that Navi won't achieve 2080 levels of Performance strikes me as being silly.

The question isn't if Navi arch can scale to produce a chip to match or best TU104, but whether AMD decided to make a 300+ mm2 chip as well as a 200+ mm2 one for this summer launch. This question is anything but silly considering AMD is undoubtedly executing on a strategy that puts business consumers first in a bid to earn trust over time. We can always hope they had the resources this time though.

Hence I think it's more likely AMD has had to add something late on in the design process. Pretty likely it won't be as ground up a design as Nvidia - more just adding enough support to normal shader cores to make the RT calculations much faster, and not even attempting the tensor cores (instead relying on shaders to denoise).

Yup, they wrote the specs the day Turing launched, probably based on the keynote slides. That was also the day an AMD marketing rep googled DXR.

Luckily AMD, fearing the unknown years ago, etched in the essence of Navi new FPGA-like circuitry based on proprietary technology called FUD - Formal Ubiquitous Design. So fear not, FUD will save the day: they can fundamentally change a GPU design in less than 6 months! /s

 

[tajoh111]

The question isn't if Navi arch can scale to produce a chip to match or best TU104, but whether AMD decided to make a 300+ mm2 chip as well as a 200+ mm2 one for this summer launch. This question is anything but silly considering AMD is undoubtedly executing on a strategy that puts business consumers first in a bid to earn trust over time. We can always hope they had the resources this time though.


Yup, they wrote the specs the day Turing launched, probably based on the keynote slides. That was also the day an AMD marketing rep googled DXR.

Luckily AMD, fearing the unknown years ago, etched in the essence of Navi new FPGA-like circuitry based on proprietary technology called FUD - Formal Ubiquitous Design. So fear not, FUD will save the day: they can fundamentally change a GPU design in less than 6 months! /s

AMD does not have the cashflow to release and develop so many 7nm at one time.

AMD concerns with 7nm chip development is getting the most bang for their buck with their investment.

Ryzen 2 comes first and GPU come away as a distant second.

For GPU, this means developing a chip to cover the largest possible market which means a polaris 10 replacement and after that, something in the low end market for the laptop/desktop/all in one market because this market AMD needs a replacement badly. They have lost almost all the laptops aside from Apple and their presence in the prebuilt desktop market has vanished. True high end is AMD last concern at this point.

We won't get a top to bottom line up of 7nm in 2019 as they do not have the money or man power to do it.

Why do people keep thinking that Radeon VII will stay as the top performing card? This is ludicrous. Why on earth would AMD purposely hold back their latest and greatest card just so that a low production stop gap can stay as the top card?

Oh, and I expect mid range cards to launch first. Top tier may be announced as well, but for later availability.

AMD will be highly limited by bandwidth which Radeon VII has 1tb/sec of. This is a luxury navi won't have. Getting more than this performance on a 200mm2 chip will be be difficult and getting significantly more(than Vega 20) with a 300mm2 chip will be difficult without using HBM2.

AMD will be likely limited to something along the lines of 448gb/sec -512gb/sec of bandwidth that a 256 bit DDR6 can provide. HBM2 at this point only makes sense on professional chips.

Since this is still GCN, that bandwidth is going to limit AMD potential performance.

AMD will only target the high end when it can simultaneously replace Vega 20 at the same time with a high end chip because large chips take more resources to develop and AMD needs to target the largest possible market with it's high end. A pure gaming chip for high end is financially irresponsible for AMD considering how much more money in terms of return on investment when it is spent on CPU's.

AMD will spend money in order of highest return on investment which means high end is getting least priority much like Vega. Smallest chips first because easiest development and largest target market. High end later when development and cost per wafer shrinks.

 

[maddie]

nm

 

[maddie]

The question isn't if Navi arch can scale to produce a chip to match or best TU104, but whether AMD decided to make a 300+ mm2 chip as well as a 200+ mm2 one for this summer launch. This question is anything but silly considering AMD is undoubtedly executing on a strategy that puts business consumers first in a bid to earn trust over time. We can always hope they had the resources this time though.


Yup, they wrote the specs the day Turing launched, probably based on the keynote slides. That was also the day an AMD marketing rep googled DXR.

Luckily AMD, fearing the unknown years ago, etched in the essence of Navi new FPGA-like circuitry based on proprietary technology called FUD - Formal Ubiquitous Design. So fear not, FUD will save the day: they can fundamentally change a GPU design in less than 6 months! /s

Should have put /sssss. You're dripping.

 

[maddie]

AMD does not have the cashflow to release and develop so many 7nm at one time.

AMD concerns with 7nm chip development is getting the most bang for their buck with their investment.

Ryzen 2 comes first and GPU come away as a distant second.

For GPU, this means developing a chip to cover the largest possible market which means a polaris 10 replacement and after that, something in the low end market for the laptop/desktop/all in one market because this market AMD needs a replacement badly. They have lost almost all the laptops aside from Apple and their presence in the prebuilt desktop market has vanished. True high end is AMD last concern at this point.

We won't get a top to bottom line up of 7nm in 2019 as they do not have the money or man power to do it.



AMD will be highly limited by bandwidth which Radeon VII has 1tb/sec of. This is a luxury navi won't have. Getting more than this performance on a 200mm2 chip will be be difficult and getting significantly more(than Vega 20) with a 300mm2 chip will be difficult without using HBM2.

AMD will be likely limited to something along the lines of 448gb/sec -512gb/sec of bandwidth that a 256 bit DDR6 can provide. HBM2 at this point only makes sense on professional chips.

Since this is still GCN, that bandwidth is going to limit AMD potential performance.

AMD will only target the high end when it can simultaneously replace Vega 20 at the same time with a high end chip because large chips take more resources to develop and AMD needs to target the largest possible market with it's high end. A pure gaming chip for high end is financially irresponsible for AMD considering how much more money in terms of return on investment when it is spent on CPU's.

AMD will spend money in order of highest return on investment which means high end is getting least priority much like Vega. Smallest chips first because easiest development and largest target market. High end later when development and cost per wafer shrinks.

You're assuming that each new design costs as much as the previous design. Take the shader units for example, once the work has been done for 7nm, using that in several GPUs adds minimal relative costs. This applies to the other blocks. Video output, memory controllers cache, etc.

A 7nm die might cost $100M+ but 2 and 3 die sharing the same IP blocks are a lot less than 200% of the initial one. It is definitely not a linear relationship.

 

[Stuka87]

Not convinced, no one thought Nvidia would actually make ray tracing work until they launched cards with hardware support, but the moment they did it was obvious the next gen consoles would have to have it as Sony/MS couldn't risk the other one adding RT support and them missing out.
Hence I think it's more likely AMD has had to add something late on in the design process. Pretty likely it won't be as ground up a design as Nvidia - more just adding enough support to normal shader cores to make the RT calculations much faster, and not even attempting the tensor cores (instead relying on shaders to denoise).

You can't be serious. You think that AMD just threw away what they had been working on for the last 3+ years and developed a new chip in a matter of months? There is no "adding" to a GPU years after its development started.

And just because the public didn't know about raytracing, doesn't mean AMD didn't. Both nVidia and AMD work with MS on DirectX, including DXR. nVidia, AMD, and MS knew about this years ago. If AMD wants hardware RT support for Navi, it will have it, as part of its day one design. If they decided not to, then it won't. But that doesn't limit out other implementations, such as Crytek's demonstration last month.

 

[DrMrLordX]

DXR works on hardware without dedicated RT cores (or similar). Neither AMD nor Sony had to add this in 6 months ago (or whenever). They can make it work now, depending on what quality levels they want.

I am not at all surprised to hear that Navi will be in the PS5. I have strongly suspected that Navi's development was tilted heavily in favor of the console makers despite others insisting that I am wrong.

 

[Stuka87]

DXR works on hardware without dedicated RT cores (or similar). Neither AMD nor Sony had to add this in 6 months ago (or whenever). They can make it work now, depending on what quality levels they want.

I am not at all surprised to hear that Navi will be in the PS5. I have strongly suspected that Navi's development was tilted heavily in favor of the console makers despite others insisting that I am wrong.

AMD themselves said that Navi was funded by "OEMs", which they seamed to state infer is Apple/Sony/MS. So zero surprise here IMO.

And when I said hardware, I meant dedicated cores for it, as opposed to more generalized compute cores.

 

[railven]

DXR works on hardware without dedicated RT cores (or similar). Neither AMD nor Sony had to add this in 6 months ago (or whenever). They can make it work now, depending on what quality levels they want.

I am not at all surprised to hear that Navi will be in the PS5. I have strongly suspected that Navi's development was tilted heavily in favor of the console makers despite others insisting that I am wrong.

Yeah, I don't understand why people are going to extremes. All it has to do is support a portion/part of it. It doesn't mean we'll get it out the ying-yang. It's marketing speak.

Remember when the PS3 was able to do tessellation? I do. It looked marvelous! But it wasn't tessellating every concrete barricade ala Crysis 2. Nope, it tessellated something you'd barely notice unless told before hand, but they could still tick off the check box.

I wouldn't be surprised if this is the extent we get for RT'ing. Fixed modes for screen shots. I'm playing through DMC5 right now, and it's in game cutscenes look GORGEOUS! Sucks they disable some of the effects for actual gameplay.

 

[tajoh111]

You're assuming that each new design costs as much as the previous design. Take the shader units for example, once the work has been done for 7nm, using that in several GPUs adds minimal relative costs. This applies to the other blocks. Video output, memory controllers cache, etc.

A 7nm die might cost $100M+ but 2 and 3 die sharing the same IP blocks are a lot less than 200% of the initial one. It is definitely not a linear relationship.

All cost scales upwards compared to 12/14/16nm. Meaning concurrent chip design goes up more and more as well. This is why we saw such a tepid release of chips on 16nm from AMD even though they vastly reduced salary and thus R and D cost by moving things to China(engineers make 1/5th the amount) according to glassdoor for AMD shanghai, this was offset by the cost in terms of hours and the man power.

So although there is some savings on chips with the same architecture the cost are still significant. Since costs have tripled going from 16nm, Do you think AMD is in a position, where they can release more than 2 Navi variants in 2019? No way. Lets wait until the end of this year and you can quote me on there being 2 or less new chips based on Navi released for discrete graphics.

But between PS5, Navi, Ryzen 2, compromise has to come somewhere because the headcount at AMD is still at 2013 levels. Since Navi discrete has the least guaranteed economic benefit of the three, it's where some compromise was likely made.

 

[Dribble]

But between PS5, Navi, Ryzen 2, compromise has to come somewhere because the headcount at AMD is still at 2013 levels. Since Navi discrete has the least guaranteed economic benefit of the three, it's where some compromise was likely made.

You can bet the compromise is allocate nearly everything to console development as that's where they have a contract with a deadline they must deliver on. Desktop gpu dev will get the left overs until that is done.

 

[psolord]

The prices tell you what the performance is going to be. If it is $250 it will have 1660Ti performance.

That's not how Ryzen was priced...!

 

[ozzy702]

That's not how Ryzen was priced...!

Ryzen didn't have the same performance as Intel, were cheap to produce and AMD cpus had an even worse black eye and public perception than AMD gpus currently have. I'm sure they will be priced less than comparable NVIDIA gpus but not by much.

 

[DrMrLordX]

Ryzen didn't have the same performance as Intel

Eh?

Compare an 1800x to a 6800k in March 2017.

 

[ozzy702]

Eh?

Compare an 1800x to a 6800k in March 2017.

March 2017 the 6800k was already an outdated CPU, especially for gaming. The 6700k was almost 2 years old at that point and to this day is a better gaming CPU than anything AMD has brought to the table.

Yes, it brought more cores to the landscape and had a great price/performance ratio, but even the 2700x is outmatched by Intel CPUs from 2015 in games.

Ryzen didn't bring apples to apples across the board parity with Intel offerings so yes, it's completely different than the consumer GPU landscape and pricing.

 

[Elfear]

March 2017 the 6800k was already an outdated CPU, especially for gaming. The 6700k was almost 2 years old at that point and to this day is a better gaming CPU than anything AMD has brought to the table.

Yes, it brought more cores to the landscape and had a great price/performance ratio, but even the 2700x is outmatched by Intel CPUs from 2015 in games.

Ryzen didn't bring apples to apples across the board parity with Intel offerings so yes, it's completely different than the consumer GPU landscape and pricing.

Huh?

Seems like a 2700X does very well against a 7700k which is faster than a 6700k. Maybe "outmatched" is a little disingenuous?

https://www.tomshardware.com/reviews/cpu-hierarchy,4312.html

Gamers Nexus
"For gaming or lightweight production tasks, like occasional gameplay edits, the 7700K is still fully acceptable. The R7 2700(X) and the i7-9700K both feel like lateral upgrades, or “sidegrades,” "




Sorry for the OT. Here's to hoping AMD sends it with Navi so we can get some more competition in the rather boring GPU space.

 

[DrMrLordX]

March 2017 the 6800k was already an outdated CPU, especially for gaming.

It was retailing for $999 at the time.

Here's to hoping AMD sends it with Navi so we can get some more competition in the rather boring GPU space.

I wish them the best, and maybe this will be great for midrange shoppers, but if anyone's looking for something to challenge the 2080Ti then I think they'll be left wanting.

 

[AtenRa]

A 200-220mm2 chip with performance between RTX2060 and RTX2070 , with 150W TDP and MSRP at $275-$300 would be nice.

 

[tajoh111]

That's not how Ryzen was priced...!

Ryzen was only well priced because of how bad Intel was ripping us off over the years.

We were paying 350 dollars for dies about 114mm2 with skylake 4 cores with no board components, no memory(aside from cache), no cooler(in the case of K series pricessor), no board partner cut, additional milking on motherboard chipsets. We used to be able to get 300mm2 die for this price with lynnfield and this was pure CPU. Not 50mm2 of CPU and 64mm2 of integrated GPU.

There was a tonne of profit in processors because Intel used all the nodal improvements on die shrinkage which vastly increase yields per wafer.

Intel margins would be even more insane if they they did spend so much money on contra revenue trying to get into market where they take billion dollar losses. Arm mobility market, Mobility modem market(Intel stock jumped big yesterday after they announced they were leaving this market) are markets where Intel has spent billions, perhaps even tens of billion when you include the R and D which lower their margins.

The videocard market is different because your selling a relatively big die, where board partners in the MSRP still have to pay rest of the board components and generate a margin for themselves as well to be able to pay for RMA's and support which Videocards need more of.

There's a reason why Ryzen 8 core can drop from 499 and a better version only a year later drop to $299 without a sweat with the old 1800x selling for $199.

At it's original price of $499, the original price of ryzen 1800x still had a tonne of profit because it was only a 198mm2 die.

Videocards are different. AMD is continuous struggling to selling Polaris under 200 by themselves and they don't want to sell it below 150 because they start making no money. That is part of the reason they rebrand. Get the old SKU off the market and reset pricing back to their original levels. Right now they are forced to sell certain polaris cheap because of the over supply of Polaris chips and even now, they have been trying to resist prices changes by keeping retail RX pricing high by not doing any official price drops. This is because your getting alot of hardware, silicon, components with videocards.

In other words, they can't afford to undercut Nvidia by huge amounts because the margins are not there.

We cringe at prices of RTX 2080 ti and RTX titans, but AMD sells chips with such margins already and Intel does as well.

You don't think a $1000 dollar 2950x which has 2x 198mm2 dies does not have a crazy margin or a 32 core 2990x which consist of the above x2 doesn't have a crazy margin.

Chips like the RTX 2080ti and 2080 consist of huge monolithic chips, particularly the former which is a single chip at 751mm2. This die likely has terrible yields which is why even with it's high price, it is the only chip that has not been below MSRP even during sales because the quantity available is the low which is no surprise.

In comparison, the 9980xe which is a 18 core chip only has a die around 400mm2 but again a 2000+ price point.

If AMD has the performance, they will charge us more even for a 200mm2 die. Polaris pricing initially was the result of being 50% slower than a gtx 1070 which was supposedly priced at $399. When AMD had an upper hand even for a moment in recent history, they were selling similar die(7870) for 350 dollars which was only 213mm2, an even bigger rip off in pricing than we have today since it price to performance just matched the gtx 570 initially(techpowerup).

AMD will price Navi as high as it can initially because their products have long shelf life's and will need to start high to make subsequent rebrands look like better products than they are when the rebrands sell for lower prices. AMD is no savior when it comes to videocard pricing as they arguably accelerated/started it with 7970 and 7870 initial pricing which allowed Nvidia to increase the price of GK104 which had a cascade effect on all videocard pricing afterwards.