Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[eek2121]

Nah. Even Nvidia is deprecating the use of tensor cores on the sly.

Ray tracing denoising shaders are good examples that might benefit greatly from doubling FP32 throughput.

NVIDIA uses tensor cores for DLSS. Tensor cores aren’t going away any time soon.

 

[maddie]

NVIDIA uses tensor cores for DLSS. Tensor cores aren’t going away any time soon.

True, forgot.

 

[Veradun]

I hope you "heard" my comment in jest and not me trying to be a jerk. This isn't FB or heaven forbid twitter, and I just want to be clear that I thought it was a chance to be chuckle worthy. Apologies if it was taken any other way. We are all here to escape the rest of the world.

no worries

 

[A///]

You might see them in small quantities, pricier, or delayed until more mobile players move on to 5nm. If Ryzen is selling like hotcakes, those dies are far more profitable for AMD than GPUs, unless the GPU can be in the $700+ range.

My post was sarcasm. I believe GPUs are to remain on N7, while processors move to another N7 node, and the consoles are N7+... I find it very confusing. In any case, Sony began production earlier than June. Realistically, Zen3 was taped out months and month and months ago.

 

[Tup3x]

Nah. Even Nvidia is deprecating the use of tensor cores on the sly.

Ray tracing denoising shaders are good examples that might benefit greatly from doubling FP32 throughput.

I think that obvious. Game developers would write universal code instead of doing NVIDIA only version that uses tensor cores. If NVIDIA offers plug and play solution then that's a bit different story but still.

 

[Kenmitch]

Hype is one thing, but we have the XBox die shot to work with. If you just remove the CPU clusters you are left with ~ 317mm^2 for a 56CU + 320 bit GDDR6 interface + all the IO + multimedia circuitry.

Even if we do a simple ratio analysis (the worst one as it expands every part equally), we get 158% for a 500mm^2 die or ~ 88CU.

It strongly suggests that even with a GDDR6 512 bit memory bus, the 80CU @ ~ 500mm^2 die is wrong. We either have more than 80CU or a smaller die and HBM2 controllers will allow for even more CU.

Where am I so wrong in this?

You know how it goes....Internet get's the hype train going....AMD get's blamed for it if it doesn't turn out so good.

Nothing wrong with speculation or trying to second guess what's coming in the end.

 

[uzzi38]

Let's not get the hype train going too fast. After all there isn't even a conductor....Raja left for Intel and so far there hasn't been any volunteers.

More like Lisa would have their head on a stake if anybody even tried to create hype.

 

[Kenmitch]

More like Lisa would have their head on a stake if anybody even tried to create hype.

I'm leaning towards she'd make a human popsicle out of them and put them on display at the main entrance....Just don't visualize it.

 

[Paul98]

I am more optimistic that RDNA2 will be able to compete with NVidia at least through the 3080 after seeing the prices and where the different cards land. ( like the 3080 being a cut down 3090 ) Along with what we have seen from Xbox and PS specs.

I will finally be getting a new GPU by the end of the year, I just am waiting for both AMD and NVidia to release along with independent benchmarks. I am quite looking forward to this, hopefully in the near future we start seeing some more info on RDNA2

 

[eek2121]

You know how it goes....Internet get's the hype train going....AMD get's blamed for it if it doesn't turn out so good.

Nothing wrong with speculation or trying to second guess what's coming in the end.

I’ve long taken the position that we don’t have more information or a launch date because AMD doesn’t. They are using 100% of their capacity currently, and as soon as they get done with their run for the consoles they will be able to switch over a percentage of that capacity to Zen 3 and Navi2X. Basically: AMD doesn’t have a release date. (Maybe they do now, but they didn’t earlier)

 

[A///]

I’ve long taken the position that we don’t have more information or a launch date because AMD doesn’t. They are using 100% of their capacity currently, and as soon as they get done with their run for the consoles they will be able to switch over a percentage of that capacity to Zen 3 and Navi2X. Basically: AMD doesn’t have a release date. (Maybe they do now, but they didn’t earlier)

Or they'd rather not start another flurry of hype. The console contract with Sony alone for the initial run of 10M units goes to March 2021. You've basically implied that at least with Sony and their 10M units if not more will cause a delay in Zen 3 and Navi2X being launched well past March, as they'd need need many weeks from tapeout to finished product. Possibly seeing hardware in stores in May... Production began in June according to various reports, at least for Sony. No news on XSX. These are monolithic components and not chiplets. AMD also is delivering Milan in 2020. Now explain how AMD can just give up a very profitable sector only to supply consoles which they receive a very thin margin on per unit sold? Remember, these are being sold at a loss. Estimates have them pegged at $500, when the hardware inside is worth more.

We have not been told much since December! That was nearly a year ago. If AMD have sat on their hands since then and have yet to start production of Zen3 and RDNA2 hardware, then they are doomed. This is basically what you've implied. They would have had their last major release nearly 2 years ago by May 2021. If AMD do not announce a product within the next 2 weeks and keep waiting and waiting, people will buy Nvidia. That is the truth. No one is going to wait forever to see how RDNA2 stacks up 1 month after Nvidia launch, or 2 months, or in your example given end dates for initial orders, in June of next year.

 

[CastleBravo]

Hype is one thing, but we have the XBox die shot to work with. If you just remove the CPU clusters you are left with ~ 317mm^2 for a 56CU + 320 bit GDDR6 interface + all the IO + multimedia circuitry.

Even if we do a simple ratio analysis (the worst one as it expands every part equally), we get 158% for a 500mm^2 die or ~ 88CU.

It strongly suggests that even with a GDDR6 512 bit memory bus, the 80CU @ ~ 500mm^2 die is wrong. We either have more than 80CU or a smaller die and HBM2 controllers will allow for even more CU.

Where am I so wrong in this?

Maybe the die supports both GDDR and HBM to allow for a wider range of products?

 

[maddie]

Maybe the die supports both GDDR and HBM to allow for a wider range of products?

I've seen that mentioned, but It seems improbable to me, at least for the highest end card.

7nm wafers would have been known to be in demand a long time ago. Huawei was still around then.
A 512 bit GDDR6 memory system would be expensive in area and power. ~ 100mm^2

Why waste area in having two mutually exclusive controllers on the same die. Design the top one for HBM2 exclusively as HBM2 modules are in much higher production now compared to even 2 yrs ago.

 

[Mopetar]

Let's not get the hype train going too fast. After all there isn't even a conductor....Raja left for Intel and so far there hasn't been any volunteers.

Honestly I think it's an improvement. Raja was terrible at building hype and watching any of the events he presented at always left me less interested in whatever he was selling than before the event. He could make sleeping with a supermodel sound boring and unappealing.

The funny part about all of this is that even after the NVidia announcement there's nothing solid about mainstream cards. The 3080 looks like it's going to be a great value. Whether AMD can match them for a similar price doesn't matter, because $700 is out of the price range for most of the market. Someone needs to deliver a great card at $300 because that's what will drive sales. I don't think we've had a great card at that price since the 970.

 

[Kenmitch]

I’ve long taken the position that we don’t have more information or a launch date because AMD doesn’t. They are using 100% of their capacity currently, and as soon as they get done with their run for the consoles they will be able to switch over a percentage of that capacity to Zen 3 and Navi2X. Basically: AMD doesn’t have a release date. (Maybe they do now, but they didn’t earlier)

Really?

 

[moinmoin]

Why waste area in having two mutually exclusive controllers on the same die.

The MC for HBM2 doesn't use that much area but allows for value add that GDDR6 alone doesn't offer. Since datacenter chips are now completely separate with CDNA, not adding GDDR6 is no longer feasible regarding TAM.

But you mention an interesting point, we so far talk about "big navi" essentially as one chip. But we should expect at least two distinct dies to cover the whole range. This may be where AMD may still be able to surprise some.

Really?

Haven't you heard? AMD can only do serial business, and their business throughput is way too low. Clearly their current job is delivering all those console chips first, after that Sony and Microsoft can finally come out with the launch dates for their consoles, then AMD can finally start to worry about how and when to launch Zen 3 and RDNA2. /s

 

[CastleBravo]

I've seen that mentioned, but It seems improbable to me, at least for the highest end card.

7nm wafers would have been known to be in demand a long time ago. Huawei was still around then.
A 512 bit GDDR6 memory system would be expensive in area and power. ~ 100mm^2

Why waste area in having two mutually exclusive controllers on the same die. Design the top one for HBM2 exclusively as HBM2 modules are in much higher production now compared to even 2 yrs ago.

The GDDR6(x) controller would likely be 384bit or maybe less since it would be for the cut down version, and while it would make for a larger die, you won't get the usual drop in yields since the top card can still use dies with a defect in the GDDR controller.

 

[eek2121]

The MC for HBM2 doesn't use that much area but allows for value add that GDDR6 alone doesn't offer. Since datacenter chips are now completely separate with CDNA, not adding GDDR6 is no longer feasible regarding TAM.

But you mention an interesting point, we so far talk about "big navi" essentially as one chip. But we should expect at least two distinct dies to cover the whole range. This may be where AMD may still be able to surprise some.


Haven't you heard? AMD can only do serial business, and their business throughput is way too low. Clearly their current job is delivering all those console chips first, after that Sony and Microsoft can finally come out with the launch dates for their consoles, then AMD can finally start to worry about how and when to launch Zen 3 and RDNA2. /s

The limit isn’t AMD, it’s TSMC. Consider how many chips they need to push out for the initial run (10s of millions). Do you have any idea how long it takes to do that?

There is evidence to support this as well. Popular chips like the Ryzen 3600 have drifted in and out of stock, for example. If you are going to question me you should try and learn a about how the industry works.

 

[Hitman928]

The limit isn’t AMD, it’s TSMC. Consider how many chips they need to push out for the initial run (10s of millions). Do you have any idea how long it takes to do that?

There is evidence to support this as well. Popular chips like the Ryzen 3600 have drifted in and out of stock, for example. If you are going to question me you should try and learn a about how the industry works.

According to AMD, Q2 and Q3 is when the vast majority of console chips will be produced with ramp up in early Q2 and ramp down in late Q3. In Q4, there is a large back-off until again they start to ramp up in Q2 of 2021. This follows the seasonality of the console market in which the holiday season dominates for console sales so AMD has most output for consoles in Q2-Q3 so that the actual consoles can be ready in Q4.

So if AMD is launching Zen3 and Navi in early to mid Q4 as has been rumored, then they should have plenty of time to build inventory for launch followed by a quick ramp up in volume for the main holiday season, at least for the consumer market which has a much quicker time to market.

 

[Kenmitch]

There is evidence to support this as well. Popular chips like the Ryzen 3600 have drifted in and out of stock, for example. If you are going to question me you should try and learn a about how the industry works.

I'll clarify it for them just to help you out.

If AMD products are sold out it's because of a supply issue.

If Nvidia or Intel products are sold out it's because of the pent up demand.

I doubt I could make it any clearer. /s

 

[lobz]

I'm leaning towards she'd make a human popsicle out of them and put them on display at the main entrance....Just don't visualize it.

I don't have to. I have a couple of those at my front door.

 

[blckgrffn]

I don't have to. I have a couple of those at my front door.

And no one asks you about where you managed to get such lifelike replicas of AMD marketing associates?!?

Follow up question - did you leave them the color of carbonite or did you dress them up with some color?

 

[sandorski]

Let's not get the hype train going too fast. After all there isn't even a conductor....Raja left for Intel and so far there hasn't been any volunteers.

One Youtube RumourMonger seemed to suggest that Raja was let go after the continued disappointing results and all of the overhype RTG was involved in. Makes sense, wouldn't be surprised if it's true or not, for that matter.

 

[sandorski]

3080 levels of Performance keeps coming up and that seems like it would be a good place to be from AMDs perspective. It is another Step Up in competitiveness and also for Pricing.

One thing seems to be certain, RDNA2 is going to consume less Power, perhaps even significantly less than their 3000 series counterparts. That may alter the Market response, at the very least it should show how much that Marketing line actually influences people.

 

[lobz]

And no one asks you about where you managed to get such lifelike replicas of AMD marketing associates?!?

Follow up question - did you leave them the color of carbonite or did you dress them up with some color?

Nah, they know better than to ask things they don't really wanna know the answer to
Carbonite.