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

 

[GodisanAtheist]

AMD Radeon RX 6700 XT BIOS Analysis Reveals Extreme GPU Clock Limits

AMD is expected to debut its Radeon RX 6700 series based on the "Navi 22" silicon following its RX 6900 XT launch, to compete with NVIDIA's GeForce RTX 3060/Ti. Several rumored specifications of the RX 6700 series surfaced in an older report from last week, which referenced a similar compute...

www.techpowerup.com

- So are we still talking about RDNA2 stuff? Looks like RX6700 is going to be clocked to the hilt (certainly higher than the RX6800XT, and maybe more with an OC). 2080S level performance looks like a very achievable target for a 40CU die.
Remember, we all agreed to never apologize to Glo.

 

[.vodka]

AMD Radeon RX 6700 XT BIOS Analysis Reveals Extreme GPU Clock Limits

AMD is expected to debut its Radeon RX 6700 series based on the "Navi 22" silicon following its RX 6900 XT launch, to compete with NVIDIA's GeForce RTX 3060/Ti. Several rumored specifications of the RX 6700 series surfaced in an older report from last week, which referenced a similar compute...

www.techpowerup.com

- So are we still talking about RDNA2 stuff? Looks like RX6700 is going to be clocked to the hilt (certainly higher than the RX6800XT, and maybe more with an OC). 2080S level performance looks like a very achievable target for a 40CU die.
Remember, we all agreed to never apologize to Glo.

2080S?

With all the perf/clk improvements RDNA2 has (+21%), plus that ridiculous frequency (~2850MHz gfx clk) the 6700XT should probably be at ~2080Ti/3070 level at 1080p, but probably not quite make it at 1440p. Still closer to them than the 2080S I guess.

It all comes down to how much IC AMD is going to use on Navi22. AMD themselves provided the cache hit rates for 128, 96, 64, 48, 32MB etc of cache in one of their slides. 96MB should make for a quite solid 1440p card, 64MB would be a bit limited, maybe, for that resolution considering there's a 192 bit bus behind that cache.

Considering there's a few less memory controllers apart from having half the resources of Navi21 (~510mm²), I'd guess they didn't skimp on the ~30mm² it'd require to implement the extra 32MB to do 96MB. Navi10 is ~250mm², Navi22 should be around that die size after all the cutting down.

It'd be ridiculous to see a little 250mm² die mostly match up against a 754mm² x02 monster and the typical ~400mm² X04 die, lol

If they go for 96MB cache, it should be an excellent high refresh rate 1080p card that can comfortably do not quite as high refresh rate 1440p gaming.

64MB, at 1440p it'd probably behave like Navi21 does at 4k, it loses some of its punch, but it can still do its thing.

 

[GodisanAtheist]

If the 6700XT encroaches on 3070 territory then the 6800's days are numbered. Would make me wonder what the hell AMD is doing wrong that an 80 CU die would only be +30-40% a 40CU die.

Maybe that's the plan, 6800 only exists to move every single last N21 die (despite supposedly solid yields) and once those are depleted move the 6700xt into that slot as a "bargain" $450 dollar 300mm2 die that has nearly the same performance.

 

[.vodka]

Going by current 5700XT performance as seen in Techspot's 6800 review, that card is ~10% slower than the 2080S on average at 1080p, and that gap increases to ~14% on average at 1440p. Now, 2080Ti/3070 is 22% faster on average at 1080p, and 33% faster on average at 1440p. The gap widens quite a bit.

2080S performance (+15%) is a given between the perf/clk improvements and the extra ~600-800MHz over the 5700XT.

~+25-30% should be a realistic speedup over Navi10 with all these upgrades, so it should be more than a match at 1080p for the 2080Ti/3070, but things will probably not be as good at 1440p. 96 or 64MB IC, we'll see where it ends up at. 4k is out of the question, that resolution requires the full 128MB solution + the extra memory bus width the 6800/6800XT have.

But yeah, I agree, seeing how big Navi21 is, it probably doesn't scale as well to such high CU counts seeing what Navi22 should be capable of, taking the 5700XT as the starting point. Scaling is miles better than GCN's, that's for sure.

Having said that, the 6800 is in a weird place... It sure is a better choice for higher resolution uses than the 6700XT, though.

 

[Hitman928]

2080S?

With all the perf/clk improvements RDNA2 has (+21%), plus that ridiculous frequency (~2850MHz gfx clk) the 6700XT should probably be at ~2080Ti/3070 level at 1080p, but probably not quite make it at 1440p. Still closer to them than the 2080S I guess.

It all comes down to how much IC AMD is going to use on Navi22. AMD themselves provided the cache hit rates for 128, 96, 64, 48, 32MB etc of cache in one of their slides. 96MB should make for a quite solid 1440p card, 64MB would be a bit limited, maybe, for that resolution considering there's a 192 bit bus behind that cache.

Considering there's a few less memory controllers apart from having half the resources of Navi21 (~510mm²), I'd guess they didn't skimp on the ~30mm² it'd require to implement the extra 32MB to do 96MB. Navi10 is ~250mm², Navi22 should be around that die size after all the cutting down.

It'd be ridiculous to see a little 250mm² die mostly match up against a 754mm² x02 monster and the typical ~400mm² X04 die, lol

If they go for 96MB cache, it should be an excellent high refresh rate 1080p card that can comfortably do not quite as high refresh rate 1440p gaming.

64MB, at 1440p it'd probably behave like Navi21 does at 4k, it loses some of its punch, but it can still do its thing.

~2850 MHz is the max engine clock, not the game clock. N21 has a max game clock of 2800 MHz. N22 will likely clock higher stock than N21 but not that much higher.

 

[kurosaki]

Gosh just checked local shop, no model has any availabilty date and the few that have a price are priced like there is no tomorrow.

RX 6800 with MSRP of $579 currency adjust is listed around $800
RX 6800 XT with MSRP of $649 currency adjust is listed around $980 (some models over $1000)

I'm aware with sales tax etc prices are a bit higher than MSRP but for the Ryzen 3900x I got last year that was $50 and most of it due to lack of supply and not $250...

Eg. retailers are gouging as well

EDIT: And with RTX3080 / 3070 the difference to MSRP is smaller making the 3080 actually cheaper than a 6800xt here...

Same in Sweden..

 

[Glo.]

AMD Radeon RX 6700 XT BIOS Analysis Reveals Extreme GPU Clock Limits

AMD is expected to debut its Radeon RX 6700 series based on the "Navi 22" silicon following its RX 6900 XT launch, to compete with NVIDIA's GeForce RTX 3060/Ti. Several rumored specifications of the RX 6700 series surfaced in an older report from last week, which referenced a similar compute...

www.techpowerup.com

- So are we still talking about RDNA2 stuff? Looks like RX6700 is going to be clocked to the hilt (certainly higher than the RX6800XT, and maybe more with an OC). 2080S level performance looks like a very achievable target for a 40CU die.
Remember, we all agreed to never apologize to Glo.

The performance target was 2080S +10%. It was never meant to compete with RTX 2080 Ti.

Also, was supposed to be much more efficient. AMD, it appears however, decided to clock the hell out of the GPU, and target it at 3070, while also making the GPU much less efficient.

And you guys know me well enough to know, that I would rather 3060 Ti competitor, but much more efficient, than 3070 but less efficient .

 

[Qwertilot]

Especially when they've already cut down the 6800 so much that its more or less a 3070 competitor as well. A bit weird all round.

Better to let the AIB's do any mad over clocking in premium SKU's.

 

[uzzi38]

~2850 MHz is the max engine clock, not the game clock. N21 has a max game clock of 2800 MHz. N22 will likely clock higher stock than N21 but not that much higher.

No. The maximum settable clock that equates to the 6800XT's 2.8GHz max clock is the 2.95GHz clock. The 2.854GHz clock is the clock target set for an AIB card - just like how the 6800XT and 6900XT have a clock target of 2.33GHz set in their BIOSes.

Yes, they have a clock target higher than the rated boost on the spec sheet. Don't ask.

 

[moinmoin]

Yes, they have a clock target higher than the rated boost on the spec sheet. Don't ask.

That's how AMD advertises boost clocks since Zen 3, it's the lower end of the boost range. Considering the mess AMD created with Zen 2 where the advertised boost clock is the upper end of the boost range and barely anybody was able to see it consistently this is a good change.

 

[Mopetar]

The performance target was 2080S +10%. It was never meant to compete with RTX 2080 Ti.

Also, was supposed to be much more efficient. AMD, it appears however, decided to clock the hell out of the GPU, and target it at 3070, while also making the GPU much less efficient.

And you guys know me well enough to know, that I would rather 3060 Ti competitor, but much more efficient, than 3070 but less efficient .

Seems like binning should let you have both. Why not make an efficient version, call it the 6700 that's more in line with the 3060 Ti / 2080 SUPER, and then make one that pushes the clocks to those absurd levels and call it a 6700 XT that drifts closer to 2080 Ti performance, at least in sub-4K resolutions.

Given how many different cards came out of Navi 10 it's not unreasonable for AMD to have two bins for a full die and just sell any of the chips with disabled hardware as 6600 cards. One of the reasons that chips end up having hardware disabled is that even though it's functional it can't hit performance targets. Look at Zen 3 where 95% of chiplets are going to come back 8-core capable, but a lot are being sold as 6-core parts for that reason.

Just do something like the following:

6700XT - 40 CU @ 2.6 GHz - 12 GB GDDR6
6700   - 40 CU @ 2.1 GHz - 12 GB GDDR6
6600XT - 36 CU @ 2.3 GHz - 10 GB GDDR6
6600   - 36 CU @ 2.0 GHz - 10 GB GDDR6

I'm assuming that they won't have a 32 CU bin since that's a full Navi 23 die, but the rumors surrounding that have suggested it could even be an HBM2 part meant for professional notebook chips, so who knows if Navi 22 might stretch down into the 6500-series cards as well.

The AIB cards can push the top card all the way up to 2.8 GHz with better cooling solutions.

 

[uzzi38]

Seems like binning should let you have both. Why not make an efficient version, call it the 6700 that's more in line with the 3060 Ti / 2080 SUPER, and then make one that pushes the clocks to those absurd levels and call it a 6700 XT that drifts closer to 2080 Ti performance, at least in sub-4K resolutions.

Given how many different cards came out of Navi 10 it's not unreasonable for AMD to have two bins for a full die and just sell any of the chips with disabled hardware as 6600 cards. One of the reasons that chips end up having hardware disabled is that even though it's functional it can't hit performance targets. Look at Zen 3 where 95% of chiplets are going to come back 8-core capable, but a lot are being sold as 6-core parts for that reason.

Just do something like the following:

6700XT - 40 CU @ 2.6 GHz - 12 GB GDDR6
6700   - 40 CU @ 2.1 GHz - 12 GB GDDR6
6600XT - 36 CU @ 2.3 GHz - 10 GB GDDR6
6600   - 36 CU @ 2.0 GHz - 10 GB GDDR6

I'm assuming that they won't have a 32 CU bin since that's a full Navi 23 die, but the rumors surrounding that have suggested it could even be an HBM2 part meant for professional notebook chips, so who knows if Navi 22 might stretch down into the 6500-series cards as well.

Navi23 is GDDR6

 

[Stuka87]

Seems like binning should let you have both. Why not make an efficient version, call it the 6700 that's more in line with the 3060 Ti / 2080 SUPER, and then make one that pushes the clocks to those absurd levels and call it a 6700 XT that drifts closer to 2080 Ti performance, at least in sub-4K resolutions.

Given how many different cards came out of Navi 10 it's not unreasonable for AMD to have two bins for a full die and just sell any of the chips with disabled hardware as 6600 cards. One of the reasons that chips end up having hardware disabled is that even though it's functional it can't hit performance targets. Look at Zen 3 where 95% of chiplets are going to come back 8-core capable, but a lot are being sold as 6-core parts for that reason.

Just do something like the following:

6700XT - 40 CU @ 2.6 GHz - 12 GB GDDR6
6700   - 40 CU @ 2.1 GHz - 12 GB GDDR6
6600XT - 36 CU @ 2.3 GHz - 10 GB GDDR6
6600   - 36 CU @ 2.0 GHz - 10 GB GDDR6

I'm assuming that they won't have a 32 CU bin since that's a full Navi 23 die, but the rumors surrounding that have suggested it could even be an HBM2 part meant for professional notebook chips, so who knows if Navi 22 might stretch down into the 6500-series cards as well.

The AIB cards can push the top card all the way up to 2.8 GHz with better cooling solutions.

The 6700 (non-XT) will be 36CU. I don't see any reason for them to offer two full die chips. Plus we have already seen what they did here with the 5600/5700 cards. The 5700XT was 40CU, the 5700 was 36. The 5600XT was also 36, but with less memory and a smaller bus.

 

[Hitman928]

No. The maximum settable clock that equates to the 6800XT's 2.8GHz max clock is the 2.95GHz clock. The 2.854GHz clock is the clock target set for an AIB card - just like how the 6800XT and 6900XT have a clock target of 2.33GHz set in their BIOSes.

Yes, they have a clock target higher than the rated boost on the spec sheet. Don't ask.

This is the 6800XT clock target Igor shared from one of the AIB cards:

None of the AIB cards clock that high without overclocking that I am aware of, the highest I've seen is ~2360 MHz from the Nitro+ model. My point was, just because its listed in the BIOS as the max (even non-overdrive) frequency, doesn't mean you're going to get that at stock.

AMD Radeon RX 6800XT board partner card allegedly features a 2577 MHz boost clock - VideoCardz.com

AMD Navi 21 with boost clock up to 2577 MHz? According to Igor Wallossek from Igor’sLAB, a ‘very special board partner’ graphics card based on Navi 21 GPU will feature a boost clock of 2577 MHz. Now, before we all get too excited, Igor did not mention which SKU this is (XTX, XT or XL). […]

videocardz.com

 

[Timorous]

This is the 6800XT clock target Igor shared from one of the AIB cards:

None of the AIB cards clock that high without overclocking that I am aware of, the highest I've seen is ~2360 MHz from the Nitro+ model. My point was, just because its listed in the BIOS as the max (even non-overdrive) frequency, doesn't mean you're going to get that at stock.

AMD Radeon RX 6800XT board partner card allegedly features a 2577 MHz boost clock - VideoCardz.com

AMD Navi 21 with boost clock up to 2577 MHz? According to Igor Wallossek from Igor’sLAB, a ‘very special board partner’ graphics card based on Navi 21 GPU will feature a boost clock of 2577 MHz. Now, before we all get too excited, Igor did not mention which SKU this is (XTX, XT or XL). […]

videocardz.com

The Strix LC averages 2416Mhz across the TPU suite.

When OCd it can average 2743 in unigine heaven.

Edit: It is critical to point out though that what we have are not top N21 bins. Those will be in the 6900XT so if we are going to use N21 to guess at potential N22 clocks we need to look at the top bins because the 6700XT will be using top N22 bins so its clock/voltage curve should be pretty similar to top N21 bins just like top N14 is similar to top N10

 

[Mopetar]

The 6700 (non-XT) will be 36CU. I don't see any reason for them to offer two full die chips. Plus we have already seen what they did here with the 5600/5700 cards. The 5700XT was 40CU, the 5700 was 36. The 5600XT was also 36, but with less memory and a smaller bus.

If Glo is correct (and he seems to have had good information the whole time) about the 6700XT being pushed to hit higher clock speeds and land closer to the 3070 than the 3060 Ti, then it makes sense. The higher the clock target for your 40 CU part, the more full dies that won't be able to meet it and must have some otherwise functional CUs (or WGPs now) disabled. If you have two parts that both use the full die you can make one of them an absolutely ridiculous part and not have to bin everything else that fails to live up to that. Then you can have an efficiency version that still uses the full die.

Because we also know that there's a Navi 23 part that's 32 CU, it makes a lot less sense to bin for more lower CU chips that can clock higher since you have that covered with the best performing full-die Navi 23 parts.

 

[Stuka87]

If Glo is correct (and he seems to have had good information the whole time) about the 6700XT being pushed to hit higher clock speeds and land closer to the 3070 than the 3060 Ti, then it makes sense. The higher the clock target for your 40 CU part, the more full dies that won't be able to meet it and must have some otherwise functional CUs (or WGPs now) disabled. If you have two parts that both use the full die you can make one of them an absolutely ridiculous part and not have to bin everything else that fails to live up to that. Then you can have an efficiency version that still uses the full die.

Because we also know that there's a Navi 23 part that's 32 CU, it makes a lot less sense to bin for more lower CU chips that can clock higher since you have that covered with the best performing full-die Navi 23 parts.

Yeah, thats a valid point. Maybe they are very confident that the 40CU will have very good yields at X clock speed, and don't have a reason for a cut down version.

 

[Glo.]

Because we also know that there's a Navi 23 part that's 32 CU, it makes a lot less sense to bin for more lower CU chips that can clock higher since you have that covered with the best performing full-die Navi 23 parts.

Now, here's the thing.

If we look at the lineup of what AMD did, compared to what was rumored, the decision at some point has been made to clock the hell out of those GPUs.

And lose efficiency.

I don't expect that desktop N23 will follow different path, and it may be closer to 150W of TDP, rather than the rumored initially 125W(that was power target for this die).

Yesterday I was confident that it won't be more than 130W SKU(the full, 32 CU/128 bit GDDR6 design) for the whole board. Today - Im leaning more towards AMD pushing the boundaries of performance, rather than low power draw.

It might actually end up being 150W Total Board Power GPU.

At least it will be cheap...

 

[TESKATLIPOKA]

When will they actually show N22 and N23?

 

[uzzi38]

When will they actually show N22 and N23?

At the very least N22 should first show up at CES.

N23 idk. Depends on how AMD wanted to allocate the few wafers they have - but N23 can't be too much further behind, it's the main chip they're pushing for mobile.

 

[Glo.]

When will they actually show N22 and N23?

Last time I heard was January for N22, N23 - I don't know. But it should follow soon after.

 

[Timorous]

Now, here's the thing.

If we look at the lineup of what AMD did, compared to what was rumored, the decision at some point has been made to clock the hell out of those GPUs.

And lose efficiency.

I don't expect that desktop N23 will follow different path, and it may be closer to 150W of TDP, rather than the rumored initially 125W(that was power target for this die).

Yesterday I was confident that it won't be more than 130W SKU(the full, 32 CU/128 bit GDDR6 design) for the whole board. Today - Im leaning more towards AMD pushing the boundaries of performance, rather than low power draw.

It might actually end up being 150W Total Board Power GPU.

At least it will be cheap...

If the 6700XT is highly strung I can see it using more power than the 6800.

 

[Glo.]

If the 6700XT is highly strung I can see it using more power than the 6800.

No, it won't use more power than 6800.

 

[Mopetar]

I don't doubt that they'll have some version of Navi 23 that's also had the clock speeds pushed as far as possible, but I don't think that necessarily matters for the point of this discussion.

Just look at a site like Silicon Lottery to see what I'm getting at since they have posted statistics about how as you push clock speeds higher, fewer and fewer chips can reach those levels. Let's just use their results for the 10900K (comet lake) to illustrate this:

Clock Speed	Percent of CPUs Capable
4.8 GHz	100%
4.9 GHz	99%
5.0 GHz	68%
5.1 GHz	21%
5.2 GHz	1%

Some dies are going to have defects in the shaders, etc. that mean they get binned to a 36CU part regardless of performance, but we know that TSMC has really good 7nm yields and that most (80% plus) of the dies AMD will be getting back will be good dies. Depending on the economics how many cards you can expect to sell at a particular price you may still bin some otherwise functional dies just because you've saturated the demand for the highest performance and priced part.

So let's just imagine a hypothetical scenario with Navi 22. I'll just use the same spread as with 10900K since I have nothing else to go off of or a better starting point. It's probably not accurate, but this is just for illustration.

Clock Speed	Percent of GPUs Capable
1.8 GHz	100%
2.1 GHz	99%
2.4 GHz	68%
2.7 GHz	21%
2.85 GHz	1%

Suppose you wanted to target 2.7 GHz for your 6700XT. Only 21% of your full dies make that cut, so if you want the 6700 to be a 36 CU part you're actually binning 80% of your dies even though they could otherwise still hit reasonable clock speeds that where your previous intention in the first place. If you had stayed at something like 2.2 GHz, you might only need to bin 10% of those full die chips.

But you can have both. The top 20% of chips, go into a 6700XT at 2.7 GHz. The next 70% go into a 6700 at 2.2 GHz, and the bottom 10% get the worst performing CUs disabled to allow them to clock higher as a 6600-level part. They can obviously choose other points along that curve to get the kinds of numbers that they need based on sales expectations, etc.

The point about bringing up Navi 23 was that since they'll have another die that will top out at 32 CU, they probably don't have as much incentive to bin 80% of their full Navi 22 dies because they need cards with 32 or 36 CU to target the $300 - $375 part of the market.

This does make the pricing a little more interesting. If AMD does have a Navi 22 that equals the 3070 is 1080p and 1440p due to high clock speeds what do you think they charge for it? I can see an argument for $500, but something like $470 seems more likely. They also need to leave room for whatever they'll have at $400 to square off against the 3060 Ti.

 

[TESKATLIPOKA]

Last time I heard was January for N22, N23 - I don't know. But it should follow soon after.

So 2 months before review are out. Can't say I am happy about that.