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

[amrnuke]

I can - I doubt it will happen from AMD, at least on standard N7.

I doubt we will see anything beyond 64 CU / 32 WGP at N7+ either.

I'm still wondering when we will see chiplets usng RDNA.

The only thing holding back chiplet-based GPUs, as I understand it, is an external bus that's as fast and high bandwidth as an on-die bus.

Of course, that's a BIG "only thing."

 

[soresu]

The only thing holding back chiplet-based GPUs, as I understand it, is an external bus that's as fast and high bandwidth as an on-die bus.

Of course, that's a BIG "only thing."

If the chiplets are close enough together, I can't see why the bus should need to be dramatically different, GPU's can get quite big after all, so the signalling distance shouldn't be much longer than a single die.

Especially with the high density/width of connection you get with an interposer as used for HBM already.

 

[uzzi38]

The only thing holding back chiplet-based GPUs, as I understand it, is an external bus that's as fast and high bandwidth as an on-die bus.

Of course, that's a BIG "only thing."

Latency is a big issue for gaming GPUs as well as bandwidth. For enterprise use-cases, bandwidth is the biggest problem, but for gaming, you also need to tack latency onto the list of problems.

 

[amrnuke]

If the chiplets are close enough together, I can't see why the bus should need to be dramatically different, GPU's can get quite big after all, so the signalling distance shouldn't be much longer than a single die.

Especially with the high density/width of connection you get with an interposer as used for HBM already.

I do wonder if (similar to on the R9) an active interposer can basically provide something equivalent to (or near-enough-to-not-matter) direct-connect between chiplets. It didn't work, obviously, for the R9, but AMD weren't doing chiplets at the time.

 

[Hitman928]

Nvidia being late to the N7 party is because they made a bad bet. Afaik, the original plan was 2020 for everything. That might not end up being the case now for consumers specifically. There is legitimately a chance now that by the time they release Turing+1 on 7nm, AMD will be looking at releasing RDNA3 on 5nm GPUs, which is... unexpected.

Also, here's a fun fact for you guys. Navi 1X is gfx101X, as you'd expect. More or less anyway. Navi 2X is gfx103X. At some point or another, RDNA2 underwent a significant change is my theory as to why, though that's naught more than a guess.

AMD are earlier to big dies than first expected. No comment on whether there is one by N7 (depends on Navi 12, which makes no sense to me), but N7+? Oh there most certainly is.

From what I understand RDNA1 is only part of the architecture. Basically AMD is trying to be really , really consistent in delivering new products within a certain time frame for both GPUs and CPUs and the full RDNA architecture wasn't ready in time for the set launch window so they took what was ready and gave us the current RX5000 models.

For Nvidia, I had read that they were planning on using the Samsung 7 nm process with its EUV utilization and by the time they realized how much later Samsung 7 nm was going to be compared to TSMC 7 nm it was too late to port the design and be able to secure the needed wafer orders at TSMC considering how in demand the TSMC process ended up being. Could be wrong on both of these items, it's just the rumors/reports that I remember reading.

 

[Asterox]

I guess they might drop the price once the stocks of RX 500s are gone, because it's pretty bad right now for the 5500xt

That is logical AMD plan, because there is very solid stock of RX 580 or RX 590 GPU-s.

For people who estimated that GPU power consumption is not very important, here is very good comparison.

 

[uzzi38]

From what I understand RDNA1 is only part of the architecture. Basically AMD is trying to be really , really consistent in delivering new products within a certain time frame for both GPUs and CPUs and the full RDNA architecture wasn't ready in time for the set launch window so they took what was ready and gave us the current RX5000 models.

Yearly cadence from RDNA onwards. Or maybe I should refer to it as relentless execution?

In any case, you might be right, but Navi 2X is more like Zen 2 than Zen+, if I'm to relate it to something.

For Nvidia, I had read that they were planning on using the Samsung 7 nm process with its EUV utilization and by the time they realized how much later Samsung 7 nm was going to be compared to TSMC 7 nm it was too late to port the design and be able to secure the needed wafer orders at TSMC considering how in demand the TSMC process ended up being. Could be wrong on both of these items, it's just the rumors/reports that I remember reading.

And therein lies the failed bet. Samsung's 7nm EUV appears to be a complete dud.

 

[DeathReborn]

Yearly cadence from RDNA onwards. Or maybe I should refer to it as relentless execution?

In any case, you might be right, but Navi 2X is more like Zen 2 than Zen+, if I'm to relate it to something.



And therein lies the failed bet. Samsung's 7nm EUV appears to be a complete dud.

Samsung 7nm EUV has been producing chips for months, the Exynos 9825 in the Note 10 etc is made using it. You can't blame the process for a botched SoC design though. Nvidia will be releasing 7nm EUV chips in 2020.

 

[uzzi38]

Yeah, and the 9825 is basically the same as the 9820 which is on 8nm - an improved 10nm process.

For the record, the SD865 was also supposed to be on SS's 7nmEUV. Look at how that turned out.

 

[lifeblood]

Now what?

For discrete cards we have the 5500 and 5700 series out. What can we expect to see next year?

The 5300 exists as a mobile part but not as a discrete card. The mobile 5300 is just a 5500 with a lower clock speed, it’s not even a cut down part. I expect AMD will release a 5300 discrete card at some point to take the place of the RX 550 & 560, but I’m not sure if it will be cut down or just slower like the mobile parts. It would seem kinda awkward if the discrete card had less CU’s and such compared to the mobile parts, but that doesn’t mean AMD wouldn’t do it since they appear to be doing it with the 5600. I have no idea when they will appear as AMD appears to be chip constrained and bottom level cards have minimum profit margin. My guess is much later rather than sooner.

Based on rumors, the 5600 appears to be the real mid-range replacement. I say that given the huge price gap between the 5500 & 5700. I expect the 5600 to come in around the $250 mark. What’s curious is the 6GB & 8GB memory config. I expect it will be a 256bit bus but the 6GB version just won’t be fully populated giving it an effective 192bit bus. I also expect to see it in Q1 2020, produced on the current 7nm process.

I think that will be the end of the RDNA v1 cards. I suspect the 20 series (Navi 21, 22, 23) are all RDNA v2 cards. As such I have no idea when or even if they actually show up. Maybe they’ll be the improved 7nm+ process.

 

[Glo.]

Now what?

For discrete cards we have the 5500 and 5700 series out. What can we expect to see next year?

The 5300 exists as a mobile part but not as a discrete card. The mobile 5300 is just a 5500 with a lower clock speed, it’s not even a cut down part. I expect AMD will release a 5300 discrete card at some point to take the place of the RX 550 & 560, but I’m not sure if it will be cut down or just slower like the mobile parts. It would seem kinda awkward if the discrete card had less CU’s and such compared to the mobile parts, but that doesn’t mean AMD wouldn’t do it since they appear to be doing it with the 5600. I have no idea when they will appear as AMD appears to be chip constrained and bottom level cards have minimum profit margin. My guess is much later rather than sooner.

Based on rumors, the 5600 appears to be the real mid-range replacement. I say that given the huge price gap between the 5500 & 5700. I expect the 5600 to come in around the $250 mark. What’s curious is the 6GB & 8GB memory config. I expect it will be a 256bit bus but the 6GB version just won’t be fully populated giving it an effective 192bit bus. I also expect to see it in Q1 2020, produced on the current 7nm process.

I think that will be the end of the RDNA v1 cards. I suspect the 20 series (Navi 21, 22, 23) are all RDNA v2 cards. As such I have no idea when or even if they actually show up. Maybe they’ll be the improved 7nm+ process.

5600 series SKU has 192 bit bus.

 

[amrnuke]

Now what?

For discrete cards we have the 5500 and 5700 series out. What can we expect to see next year?

The 5300 exists as a mobile part but not as a discrete card. The mobile 5300 is just a 5500 with a lower clock speed, it’s not even a cut down part. I expect AMD will release a 5300 discrete card at some point to take the place of the RX 550 & 560, but I’m not sure if it will be cut down or just slower like the mobile parts. It would seem kinda awkward if the discrete card had less CU’s and such compared to the mobile parts, but that doesn’t mean AMD wouldn’t do it since they appear to be doing it with the 5600. I have no idea when they will appear as AMD appears to be chip constrained and bottom level cards have minimum profit margin. My guess is much later rather than sooner.

Based on rumors, the 5600 appears to be the real mid-range replacement. I say that given the huge price gap between the 5500 & 5700. I expect the 5600 to come in around the $250 mark. What’s curious is the 6GB & 8GB memory config. I expect it will be a 256bit bus but the 6GB version just won’t be fully populated giving it an effective 192bit bus. I also expect to see it in Q1 2020, produced on the current 7nm process.

I think that will be the end of the RDNA v1 cards. I suspect the 20 series (Navi 21, 22, 23) are all RDNA v2 cards. As such I have no idea when or even if they actually show up. Maybe they’ll be the improved 7nm+ process.

I fully expect just a 5600 XT, but would be happy to be wrong. I expect it will be 6GB 192bit bus. I would expect a release price, of around $229 to combat the 1660S.

The fan in me wants to see a 5600 to combat the 1660S at $10 cheaper than the 1660S, $219 at release. I would also like a 5600XT that sits somewhere around the Vega56/1070Ti performance range at ~$259-269. At that price I think there's a possibility they could afford to do a 5600XT with 256bit bus and 8GB RAM. I guess it's financially possible. However, I don't think it's likely.

 

[lifeblood]

5600 series SKU has 192 bit bus.

I’ve read that rumor but we’re also seeing talk of both 6GB and 8GB versions. They can have 8 & 6 GB version if the GPU has 256 bit bus, but not with a 192 bit bus (At least no way I know of).

 

[DeathReborn]

I’ve read that rumor but we’re also seeing talk of both 6GB and 8GB versions. They can have 8 & 6 GB version if the GPU has 256 bit bus, but not with a 192 bit bus (At least no way I know of).

They can use Mixed Density VRAM with some chips being 8Gbit & other 16Gbit, Nvidia did it before with the 550 Ti 1GB 192bit.

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

I do however think it'll be Navi 10 cut down with 192bit for 5600 6GB & 256bit for 5600XT 8GB.

 

[uzzi38]

I think it's 192 bit bus for both, but not like it matters for stock config.

A 192 bit bus would be able to supply both a 28 or a 32 CU GPU, the two most likely configurations for the 5600 series.

 

[amrnuke]

Making some assumptions (calculated values in red). Using Sapphire 5500 XT Pulse 4GB as baseline, and the reference 5700 and 5700XT.

I assume scaling based on what appears to be:

• 5.39% gain per CU added at 1080p
• 5.75% gain per CU added at 1440p
• 5W extra power draw per CU added at peak gaming (though largely this is skewed by the big difference between 5700 and 5700XT, and I think realistically the 5600 and XT will have power draws around 150 and 165W respectively)

Card	CU	% of 5500XT 1080p	% of 5500XT 1400p	W Peak Gaming
5500XT (4GB)	24	100	100	131
5600 (?GB)	28	122	123	151
5600XT (?GB)	32	143	146	171
5700 (8GB)	36	165	170	180
5700XT (8GB)	40	186	191	227

That's a neat and tidy lineup.

At 1080p:
1660 115%
5600 122%
1660S 128%
Vega56 141%
5600XT 143%
5700 165%

At 1440p:
1660 113%
5600 123%
1660S 128%
5600XT 146%
Vega56 147%
5700 170%

Peak Gaming W:
1660 124W
1660S 134W
5600 151W
5600XT 171W
5700 180W
Vega56 237W

If Navi scales as it appears it might, then I think we could probably expect a 28CU Navi chip to sit just below a 1660S, and a 32CU Navi chip to blow a 1660S clean out of the water matching 1070Ti performance. Sandwiching the 1660S with prices of $209 for the 5600 and $259 for the 5600XT would make for some nice competition.

 

[lifeblood]

I think it's 192 bit bus for both, but not like it matters for stock config.

A 192 bit bus would be able to supply both a 28 or a 32 CU GPU, the two most likely configurations for the 5600 series.

Can a 192 bit bus support 8GB of memory and use it efficiently?

My understanding is a 192 bit bus can support 6 memory chips with 32 bits per chip. If we assume 1GB per chip then it maxes out a 6GB of vram. To equip it with 12GB which is 12 chips then you'll have two chips sharing the 32 bit bus which would tend to cut in half the available bandwidth for each chip. To have an irregular number of chips, 8 in this case, the two additional chips would have to share the bus with two other chips. That means 4 chips will have full use of the bus while 4 would be sharing their bus. That would result in uneven memory performance which would be a nightmare for the GPU's memory manager.

I do however think it'll be Navi 10 cut down with 192bit for 5600 6GB & 256bit for 5600XT 8GB.

This sounds much more probable. A cut down Navi 10 GPU with it's 256bit bus for the 5600 XT and "genuine" Navi 12 GPU's for the vanilla 5600. That solves the whole 6 & 8 GB memory conundrum while allowing the use of Navi 10's chips that can't pass QC for 5700's.

 

[uzzi38]

In the first place I'm not even sure about the 8GB thing. There's only one company that's filed any 8GB skus afaik.

 

[lifeblood]

In the first place I'm not even sure about the 8GB thing. There's only one company that's filed any 8GB skus afaik.

Which would also nicely solve the memory conundrum, 8GB is a mistake.

 

[lifeblood]

In the first place I'm not even sure about the 8GB thing. There's only one company that's filed any 8GB skus afaik.

Toms is reporting that Gigabyte just registered a bunch of names for cards and all the 5600's were 6GB models. It's looking more like the 8GB thing was fake news.

 

[lifeblood]

Toms Hardware is now reporting on a rumor that the 5500 XT GPU was produced by Samsung, not TSMC.

I don't put a lot of confidence in that claim but if true may explain why 5500 was so damn late. AMD may have tried to get TSMC to make it but when it became clear it would take too long they went to Samsung. Of course AMD made that deal with Samsung on mobile GPU's and the 5500 may well have been part of that deal from the beginning.

 

[uzzi38]

Its a fake rumour. It was first spread by Fudzilla, who also claimed Rome is shipping with unsafe voltages.

 

[Glo.]

I’ve read that rumor but we’re also seeing talk of both 6GB and 8GB versions. They can have 8 & 6 GB version if the GPU has 256 bit bus, but not with a 192 bit bus (At least no way I know of).

Its not a Rumor. Only ASUS posted 8 GB models to Eurasian commision, for approval.

Everybody else is posting 6 GB/192 Bit GDDR6 GPUs, in both: RX 5600 and RX 5600 XT SKUs.

 

[lifeblood]

Its a fake rumour. It was first spread by Fudzilla, who also claimed Rome is shipping with unsafe voltages.

AMD just issued a statement saying the 5500 series is manufactured at TSMC. I don't know which is worse, wccftech or fudzilla.

 

[Mopetar]

At least fudzilla is appropriately named for a person to come to the conclusion that the rumors should be taken with a pinch of salt.

If we could just get wccftech to rebrand as wtftech they’d be on even footing in that regard.