It looks like the 'real' /affordable RDNA3 + next gen NV desktop launch won't launch until September. Thoughts?

[eddyr]

I see that or a highly clocked N32 as the 7800XT options and both will probably perform around 6950XT levels. N32 would give AMD a bit more pricing freedom because the die is 33% smaller than N31 but if they are running similar power envelopes then board costs are probably a tie so overall BOM probably won't be that different. If they released this around £550-600 it would be a better option than the 4070 IMO.

I'm leaning more towards the latter with 7800xt being Navi32.
If there is anything to the rumour that Navi31 fell short of its performance target due a hardware bug, then if these do not affect Navi 32, at 60CU, that combined with a bump in clocks would surely leave no room in the stack for a further cut down N31 with 70CU. They'd be within spitting distance! Setting aside the bug rumour, smaller dies tend to pushed to run higher clock speeds anyway, the gap feels to narrow.

 

[timorous]

I'm leaning more towards the latter with 7800xt being Navi32.
If there is anything to the rumour that Navi31 fell short of its performance target due a hardware bug, then if these do not affect Navi 32, at 60CU, that combined with a bump in clocks would surely leave no room in the stack for a further cut down N31 with 70CU. They'd be within spitting distance! Setting aside the bug rumour, smaller dies tend to pushed to run higher clock speeds anyway, the gap feels to narrow.

That is how I am leaning but if N32 is a while away AMD may want to offer something more immediately so that cut N31 would be an option. Then N32 would be for 7800 non XT, 7700XT and maybe a 7700 non XT.

The 7700 question is do they go 3 MCD and 12GB or do they keep 4 MCDs, maybe user slower RAM but keep it at 16GB?

 

[eddyr]

That is how I am leaning but if N32 is a while away AMD may want to offer something more immediately so that cut N31 would be an option. Then N32 would be for 7800 non XT, 7700XT and maybe a 7700 non XT.

The 7700 question is do they go 3 MCD and 12GB or do they keep 4 MCDs, maybe user slower RAM but keep it at 16GB?

I suppose it depends if a cut down N32 7700 competes against a AD106 chip, then I can see them quite happily going 3 MCD 12GB vs an 8GB AD106.
They still get to advertise 50% more ram than the competition, save on an MCD, pair of vram chips per card and in doing so use a smidge less power than they would otherwise.

 

[CAT-THE-FIFTH]

I suppose it depends if a cut down N31 7700 competes against a AD106 chip, then I can see them quite happily going 3 MCD 12GB vs an 8GB AD106.
They still get to advertise 50% more ram than the competition, save on an MCD, pair of vram chips per card and in doing so use a smidge less power than they would otherwise.

You mean AD104? Navi33 is the AD106 competitor and seems to be doing reasonably well against in laptops IIRC.

 

[g67575]

Lol, people still clinging on to Navi32 rumours.

Yes eventually, but no sign of life at the moment.

The rumour mill got it wrong once again.

If you think about it, N32 is more likely to be the successor to the 6700X and the 6700 (both are Navi22), you should get around 20 extra CUs.

 

[Xelene]

AMD will want to shift the RDNA2 cards from the channel before dropping 78xx and lower to desktop, presumably they still have some contracted capacity at TSMC for 7nm.

 

[eddyr]

You mean AD104? Navi33 is the AD106 competitor and seems to be doing reasonably well against in laptops IIRC.

I've seen that the cut down Navi33 7600s is around the laptop 4060, but that is using AD107. https://www.notebookcheck.net/NVIDIA-GeForce-RTX-4060-Laptop-GPU-Benchmarks-and-Specs.675692.0.html
Not seen any reviews of the laptop 4070 which should use AD106, a quick google didn't throw up any benchmarks.

In desktop I can see the cut down N32 being up against AD106, at least if AMD want a favourable comparison/reception.

Chip	AD102	AD103	AD104	AD104 - 4070 Desktop	AD106	AD107 - 4060 laptop
Die size	608 mm2	378.6 mm2	295 mm2		190 mm2	146 mm2
Transistors	76.3B	45.9B	35.8B		Unknown	Unknown
Transistor density	125.5 MTr/mm2	121.1 MTr/mm2	121.4 MTr/mm2		Unknown	Unknown
Graphics processing clusters (GPC)	12	7	5	4	3	2
Streaming multiprocessors (SM)	144	80	60	46	36	24
CUDA cores	18432	10240	7680	5888	4608	3072
Texture mapping units	576	320	240	184	144	96
Render output units	192	112	80	64	64	32
Tensor cores	576	320	240	184	144	96
RT cores	144	80	60	46	36	24
L1 cache	18 MB	10 MB	7.5 MB		4.5 MB	3 MB
128 KB per SM
L2 cache	96 MB	64 MB	48 MB	36 MB	32 MB

 

[CAT-THE-FIFTH]

I've seen that the cut down Navi33 7600s is around the laptop 4060, but that is using AD107. https://www.notebookcheck.net/NVIDIA-GeForce-RTX-4060-Laptop-GPU-Benchmarks-and-Specs.675692.0.html
Not seen any reviews of the laptop 4070 which should use AD106, a quick google didn't throw up any benchmarks.

I can see the cut down N32 being up against AD106, at least if AMD want a favourable reception.

Chip	AD102	AD103	AD104	AD104 - 4070 Desktop	AD106	AD107 - 4060 laptop
Die size	608 mm2	378.6 mm2	295 mm2		190 mm2	146 mm2
Transistors	76.3B	45.9B	35.8B		Unknown	Unknown
Transistor density	125.5 MTr/mm2	121.1 MTr/mm2	121.4 MTr/mm2		Unknown	Unknown
Graphics processing clusters (GPC)	12	7	5	4	3	2
Streaming multiprocessors (SM)	144	80	60	46	36	24
CUDA cores	18432	10240	7680	5888	4608	3072
Texture mapping units	576	320	240	184	144	96
Render output units	192	112	80	64	64	32
Tensor cores	576	320	240	184	144	96
RT cores	144	80	60	46	36	24
L1 cache	18 MB	10 MB	7.5 MB		4.5 MB	3 MB
128 KB per SM
L2 cache	96 MB	64 MB	48 MB	36 MB	32 MB

The AD104 is the RTX4070/RTX4070TI series so I can see a cut down Navi 31 competing well with the RTX4070TI,but if Navi 32 clocks high enough it could get close. Navi 32 should in theory compete fine with the RTX4070.

I forgot the laptop RTX4060 is an AD107 dGPU,but the desktop RTX4060 will use a cut down AD106. So if a cut down Navi 33 can compete with a full AD107,then a full Navi 33 should be in theory compete with a desktop RTX4060 or get close enough. The RTX4060TI,OTH,might need a cut down Navi 32 IMHO. But if AMD are cunning,they should make it have at least 10GB VRAM or even 12GB.

 

[muon]

AMD says it could have competed with the RTX4090 if it wanted to, but they do not want to make a 600watt GPU and costs over $1000, AMD says it wants to keep making GPUs priced no higher than $999 as this is the mainstream sweet spot

It's nice of AMD to confirm that $999usd is mainstream pricing for GPUs

Reddit - Dive into anything

www.reddit.com

Their flagship card is already 384bit with 24GB of memory, 529mm^2 and not that power efficient.

As if they could have built a much of a larger chip.

The 4080 is 379mm^2 and the 4090 (and future 4090 ti) are 608mm^2.

 

[g67575]

AMD will want to shift the RDNA2 cards from the channel before dropping 78xx and lower to desktop, presumably they still have some contracted capacity at TSMC for 7nm.

The Navi21 cards are all on sale now, priced competitively against the RTX 4070, all with 4GB more VRAM.

They'll be gone this quarter I reckon.

 

[Jono8]

The Navi21 cards are all on sale now, priced competitively against the RTX 4070, all with 4GB more VRAM.

They'll be gone this quarter I reckon.

Not really.

Apart from the extra 4gb ram, there is little point in choosing the 69xx series over the 4070 at current prices.

 

[humbug]

Their flagship card is already 384bit with 24GB of memory, 529mm^2 and not that power efficient.

As if they could have built a much of a larger chip.

The 4080 is 379mm^2 and the 4090 (and future 4090 ti) are 608mm^2.

The logic die is 301mm, its actually smaller than the 6700XT at 335mm.

The 529mm you're quoting includes the 6 memory dies which are 6nm.

 

[g67575]

Not really.

Apart from the extra 4gb ram, there is little point in choosing the 69xx series over the 4070 at current prices.

You can get the top Navi 21 card for around £640. Or, you could yesterday, it's sold out now.

Both the RX 6900 XT and the XTX are (a bit) faster than the RTX 4070.

The RX 6800 is going for <£500.

Some websites aren't even selling any Navi21 graphics cards are more.

The RX 6900 XT is pretty hard to find anywhere.

You think AMD will still be selling Navi21 cards in July then?

I think the RX 6800 XT will sell reasonably well too, because of it's higher performance in demanding games like The Last of Us (which the 6900 series do well in also):

 

[muon]

The logic die is 301mm, its actually smaller than the 6700XT at 335mm.

The 529mm you're quoting includes the 6 memory dies which are 6nm.

So that's my point. It would take 80mm^2 to come to the same size as the 4090 and consume more power at the same time. Is the 4090 just shaders? No it isn't. I don't get your point.

 

[g67575]

Navi 33 is trash tier (mostly laptop GPUs), surprised people care about it.

 

[eddyr]

The AD104 is the RTX4070/RTX4070TI series so I can see a cut down Navi 31 competing well with the RTX4070TI,but if Navi 32 clocks high enough it could get close. Navi 32 should in theory compete fine with the RTX4070.

I forgot the laptop RTX4060 is an AD107 dGPU,but the desktop RTX4060 will use a cut down AD106. So if a cut down Navi 33 can compete with a full AD107,then a full Navi 33 should be in theory compete with a desktop RTX4060 or get close enough. The RTX4060TI,OTH,might need a cut down Navi 32 IMHO. But if AMD are cunning,they should make it have at least 10GB VRAM or even 12GB.

Yeah, if they are cunning. AMD really need to make sure they edge out NV at price/perf points, especially in the more price sensitive segment of the market, really put forward a compelling proposition.
If they offer more vram and are ahead in rasterization than the NV equivalent, then that can offset their weaker RT performance bringing it in line with the NV counterpart and negate that (valid) criticism. RT is here now, they could get away with it last gen, but not going forward. Having that RT deficit sediment in peoples minds for another 2+ years will do damage.

More ram, faster in rasterization, at least equal in RT. I think that works. They can keep their "premium/not a budget brand" mantra, just offer people a compelling option.

 

[g67575]

70 CU Navi 31 is approx. equal to an RTX 3080 TI according to techpowerup. I suppose they might be able to eke a bit more out of it by clocking it higher - but the boost clock speed of the W7800 is already 2499 Mhz (~ 100 Mhz higher than the RX 7900 XT), which they presumably calculated by looking at the FP32 processing power.

The other thing to note, is that the number of ROPs and amount of L3 cache is lower than the RX 7900 XT.

I think it will need to be a cheaper than the RTX 4070 TI, because it is likely to be a bit slower.

EDIT - Just realised that the FP32 processing power of RDNA3 (RDNA2 as well I think?) cards is calculated based on the boost clocks, not the game clocks. So, the actual FP32 processing power is lower in games, than the specs suggest.

For the RX 7900 XT - In games, you can expect an FP32 processing power of (2025 MHz x 5376) x 2 = 21.7 TFlops x 2 = 43.5 TFlops at stock. Not 51.48 TFlops as the specs suggest.

For the 70 CU Navi 21 GPU - In games, you can expect an FP32 processing power of (2025 MHz x 4480) x 2 = 18.1 TFlops x 2 = 36.2 TFlops at stock. We don't know the exact game clock yet, but it's likely to be similar to the RX 7900 XT.

The same applies to the pixel and texture rates also.

The RTX 3080 TI has a FP32 processing power of 34.1 TFlops, according to the spec.

 

[Cyber-Mav]

Navi 33 is trash tier (mostly laptop GPUs), surprised people care about it.

If amd can match performance of 4070 it wont be at the same power consumption level, amd gonna need 100w more to get the job done. Hopefully amd play it smart and undercut the 4070 price by 100 quid or more

 

[g67575]

Is it possible to fix the voltage on RDNA3 GPUs?

Techpowerup's RX 7900 XT review suggests that the actual clock rate in games is largely determined by the voltage the card is using at the time:

From here: https://www.techpowerup.com/review/amd-radeon-rx-7900-xt/38.html

So, fixing the voltage at 0.95v or 1.0v should allow for higher and more consistent clocks in games.

 

[CAT-THE-FIFTH]

If amd can match performance of 4070 it wont be at the same power consumption level, amd gonna need 100w more to get the job done. Hopefully amd play it smart and undercut the 4070 price by 100 quid or more

According to TPU(who do actual measurements of power consumption at the connectors),the RX7900XTX is around 50W~60W more compared with the RTX4080(AD103). That is also with a 384 bit memory controller with 24GB of VRAM too. The RX7900XT is around 10% faster than the RTX4070TI(AD104) and consumes about 40W~50W more with a 320 bit memory controller with 20GB of VRAM. That extra power consumption could be the extra VRAM.

Also it depends on what AMD uses to compete with the RTX4070. If it is Navi 31,it might be cut down a lot like the RX6800 was,and use a 256 bit memory bus. That should make it more efficient(the RX6800 was very efficient). It might be a pushed Navi 32? But I would find it hard to believe,that it would take 100W more than an RTX4070,because it would make the RX6800XT and RX7800XT have the same power consumption and performance!

Agreed,that it should be under £500. To catch the RTX4070 it would need to be around 40% faster than an RX6700XT:

https://tpucdn.com/review/nvidia-geforce-rtx-4070-founders-edition/images/relative-performance-2560-1440.png

The RX6700XT can be found for well under £400 now.