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

 

[leoneazzurro]

Ther was a comparison (now deleted) about the "Big Navi" pagkage being comparable to the Radeon VII one, and the Radeon VII was 331 mm^2 but it had also four stacks of HBM on the package itself. If that's the case, then Big navi is big indeed.

 

[Karnak]

Ther was a comparison (now deleted) about the "Big Navi" pagkage being comparable to the Radeon VII one, and the Radeon VII was 331 mm^2 but it had also four stacks of HBM on the package itself. If that's the case, then Big navi is big indeed.

You mean this one?

 

[leoneazzurro]

yep

 

[Karnak]

Was watching Buildzoid's video about the reference GPU and there was an interesting thing he mentioned: Looking at the back, there are two screws left and right to the GPU. He thinks that this points towards HBM because with GDDR they most likely just won't fit there. On the 5700(XT) e.g. there are no such screws - but they are present on the Radeon VII and Vega10. Both using HBM...

Navi10 with GDDR6:

RadeonVII with HBM:

RDNA2:

 

[leoneazzurro]

These screws are simply required for a better thermal contact in the case the die/package Is bigger than usual (as in the case of HBM but it may be the same in the case of a big almost dquare shaped die

 

[Glo.]

If it is around the Radeon VII size its Navi 22. The 60 CU GPU with 256 bit bus.

So everything fits, so far.

 

[leoneazzurro]

Package is around Radeon VII size, but Radeon VII has also four HBM dies on that package. So two possibilities:

1) Huge die with GDDR6
2) Vega 20 sized die with HBM

Moreover, the cooler is clearly targetted at 275+W cards

 

[Karnak]

If it is around the Radeon VII size its Navi 22. The 60 CU GPU with 256 bit bus.

So everything fits, so far.

I highly doubt that N22 will featue 2x8pin. That's (most likely) definitely N21.

 

[Glo.]

I highly doubt that N22 will featue 2x8pin. That's (most likely) definitely N21.

True.

 

[Helis4life]

Looking at the fortnite renders of the rx6000 cooler, the retention bracket is larger than the one on vega7. If big navi was using anything other than hbm than it would not have a larger retention bracket, the vega7 interposer was approx 840mm2


Not sure who put together the comparison image above, but using a pixel count on the rx6000 image and the pcie connector as a reference, the rx6000 has a diametrically opposite mounting hole width of approx 100mm, vega7 has 95.

 

[Olikan]

Well, if N21 uses HBM, no way it's a 505mm2 die... there would be what? ~110mm2 unused?

 

[Konan]

@leoneazzurro If there is this massive 128mb L2 cache that is also rumored to be there how much space will that take up? And would it be blocks of 2MB slices or 4MB?
I.e. something like 30-35% of 500mm2 (if using 2MB blocks for comparison)

 

[raghu78]

Package is around Radeon VII size, but Radeon VII has also four HBM dies on that package. So two possibilities:

1) Huge die with GDDR6
2) Vega 20 sized die with HBM

Moreover, the cooler is clearly targetted at 275+W cards

Vega 20 was a 331 sq mm die with 4 HBM2 stacks. Navi 21 is likely to be a 500 sq mm die with 2 HBM2E stacks.

Well, if N21 uses HBM, no way it's a 505mm2 die... there would be what? ~110mm2 unused?

Not really. It depends on the die area spent for Radeon PRO features like SR-IOV and MXGPU, xGMI (for multi GPU), FP16 etc.

 

[A///]

Do you have a price list?

Was there any point in NVIDIA announcing RT for Fortnite when the target audience can't afford a card to run it?

Which is also ironic considering most of Jay's audience are not adults like him but in the same field as Linus' fans. It's mostly teenagers without jobs. Yes, there are adults, but the humor on Jay's channel attracts younger people. I'm sure GamersNexus attracts an older crowd, particularly one seeking a free sleeping aid.

 

[MrTeal]

AMD is mixing slotted and philips screws on the same bracket. I'm out.

 

[A///]

I haven’t actually seen the stuff about VRMs and memory. Most of the YouTube channels talking about upcoming hardware are complete garbage. They are making a video every day, it seems, even through there hasn’t been much new information in months. They get a lot of views even though they don’t have any new information. I wonder how much money they are making from talking for 20 minutes without any real content.

Correct. I don't know why they do that. It must be to get the Youtube algorithm to keep recommending them or get more views on that new video which may translate to more views on other channels. Don't forget, they all have got a Patron, too.

As to to the actual “rumors”, there is no way they can put GDDR6 on a separate board from the gpu if that is what they were implying. The VRM on a separate board doesn’t make sense either, at least not in the desktop market. There was some rumors about some new power management system in his latest video. The only thing I can think of is that they are making a mezzanine type card that includes gpu and some voltage regulating hardware. That would more likely be a CDNA card, so it isn’t necessarily complete garbage, it just might be for a different product. The GPUs that AMD will be making for their super computer wins could be a mezzanine type boards/packages with VRMs on the gpu board.

One of the more prominent channels said that, and then a few others ran with it or retracted it once someone with knowledge on the topic probably informed them in private. The first time I heard it in a video I couldn't control my laughter. A mezz board makes sense in the datacenter. AMD hired Dan McNamara earlier this year to head their server division. He was head of Intel's FPGA division, and AMD has been on a hiring streak of FPGA engineers if their job listings are anything to go by. We shall see. I'd love to comment on what AMD may have in store for the Epyc line but I'd be no better than those rumor mills, and they'd likely try and pass parts of the post off as fact.

If their top end is using a 256-bit bus, then it seems like they need something to compete with Nvidia. That is why I am taking the cache rumor a bit more seriously. Infinity fabric connected chips do not necessarily need to be on the same board, but if they are using such cache die, then they would much more likely be on the same package as the gpu using IFOP style connections. That is much lower power than IFIS off package links. Such a cache die might be reusable for Epyc, so it makes some sense. An Epyc with 128 MB L4 cache chips would be very interesting.

As Andrei pointed out, the resuability of such a move is cost shattering because it's an inherintly cheap method of production that gets results. Playing off of Dr. Su's words that RDNA will mature like fine such like Ryzen, I'm inclined to place as much faith in Andrei's theory as you have. An integrated cache on chip is a lot more expensive and adds complexity that may hinder overall yields. I could see the chips being used in the higher spectrum of the Threadrippers, too. I'm curious if AMD will divide TR and TR Pro again for Zen 3. If AMD can provide a higher performance punch with Zen3 chiplets, and as the TR board platform is new, they could provide more performance and price it whatever they like. The masses would eat it up.

 

[A///]

Who else is going to be working on a variable rate shading bug, and has access to an engineering board? It's not an AIB

Claims and a board, without the ability to verify the claims or if the board is it. Ok.

 

[Tup3x]

Looks like it will be as hungry card as Ampere.

 

[Gideon]

Looks like it will be as hungry card as Ampere.

The cooler actually looks very similar to the one on Vega VII, which had a 295 W TDP. So yeah, it'll burn "many a watt" but might not quite be at TU102 levels

vs

The one thing I do not like about the similarity is that it was a very noisy cooler cofirmed by both:

Anandtech

The noise levels of the card look surprising at first blush. Ultimately, what's happening here is the consequence of a very aggressive fan curve, one that invests all potential acoustic improvements of an open-air triple fan card for cooling capability. Going this route makes the fan noise comparable to RX Vega 64's blower.

and Techpowerup

Gaming noise levels are just way too high; the fans ramp up very quickly and become a nuisance mere seconds after putting some gaming load on the card. If you compare this to NVIDIA's offerings, it'll be a day-and-night difference! If you value low noise, the Radeon VII is definitely not for you.

A 8-9 dbA difference compared to a 2080 and ~6 dBA compared to a 2080 Ti. As the decibel scale is logarithmicm that's nearly 10x noisier.

Overall I really hope they have improved upon the cooler and don't need as agressive ramp-up.

 

[Saylick]

AMD is mixing slotted and philips screws on the same bracket. I'm out.

Hey, as long as it ain't wood screws...

 

[A///]

RX6000 looks like it's got larger overall fans with the tips going beyond just under the housing covering as opposed to the VEGA VII and its shorter fans with wider fan blade curve. Smaller fans, too. There's a gap between the fan blade tips and the housing . The VEGA VII was also based on the now dead GCN architecture which was never designed for gaming, but for compute. IIRC AMD pushed the power envelope on the VII to try and compete with NVidia. If we're judging just by design than just how different are these than most AIB cards from team green? Dual and tri axial fan designs are common from the likes of Asus and EVGA.

 

[Saylick]

I do think it's rather interesting that each fan has that perimeter ring of plastic tying all of the blades together. I remember Nvidia did this for their blower card for the GTX 580 and it helped the fan blades stay together or something... not sure if it helped with acoustics. Anyways, I'm no aerodynamicist but there's got to be better axial flow of air when you have that perimeter band of plastic since it is pretty much a ducted fan, right? Better axial flow means lower fan speeds needed for a given volume of air moved (CFM), therefore reducing noise.

 

[leoneazzurro]

The cooler actually looks very similar to the one on Vega VII, which had a 295 W TDP. So yeah, it'll burn "many a watt" but might not quite be at TU102 levels

vs

The one thing I do not like about the similarity is that it was a very noisy cooler cofirmed by both:

Anandtech

and Techpowerup

A 8-9 dbA difference compared to a 2080 and ~6 dBA compared to a 2080 Ti. As the decibel scale is logarithmicm that's nearly 10x noisier.

Overall I really hope they have improved upon the cooler and don't need as agressive ramp-up.

Fans are larger, though, with bigger blades. Also, we don't know about rotational speed. Even if the design is similar, noise cannot be determined by looking at an older card.

 

[coercitiv]

I do think it's rather interesting that each fan has that perimeter ring of plastic tying all of the blades together. I remember Nvidia did this for their blower card for the GTX 580 and it helped the fan blades stay together or something...

Ideally you want that "duct" to be as tight as possible around the blades. As you probably know, air speed reaches the highest velocity around the outer edge of the blades, so minimizing surface area for air to escape around the edge of the blades brings double yields: higher fan performance and lower noise (since there's a lot of air turbulence created there in the absence of the duct). It all depends on how tight that duct is though.

 

[Gideon]

Fans are larger, though, with bigger blades. Also, we don't know about rotational speed. Even if the design is similar, noise cannot be determined by looking at an older card.

I agree. Even if it's exactly the same the way fans are spinned up and how fast they'll changes the results significantly.

But regarding the fans being larger, I'm not quite sold on that yet. They certainly look bigger because of the styling, but if you disregard the black edges and instead compared to the outputs panel, they seem to align almost exactly to the VII design:

Top:






Roughly the same when disregarding perspective.


Bottom:







Then again the Fortnite 3D-model has a wrong number of blades on the fans ( 8 vs 9 in the render) so the model might not be all that exact.

EDIT:
The fans "holes" seem to be of very similar size. The fan's themselves with their "duct" seem to be a tiny bit bigger (filling the entire hole) but only slightly so.