• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

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

Soldato
OP
Joined
30 Jun 2019
Posts
7,876
It looks like you can fix the minimum clock rate and set the voltage within AMD's software:

bPPXzMv.png

The RX 7900 XT's hotspot temp. seems to be quite high with these settings though (close to 90).
 
Last edited:
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
I hope someone can get their hands on a Radeon PRO W7800 card and do a review, showing performance in games.

Would be interesting to see if it can get close to an RTX 4070 TI.
 
Last edited:
Soldato
Joined
9 Nov 2009
Posts
24,932
Location
Planet Earth
The RX7600S which uses a cut down 6NM die compared to a full TSMC 4N 5NM AD107:

At similar power the RX7600S has similar rasterised performance but OFC worse RT performance. Surprised considering the RTX4060 mobile is on a much better process node.

Apparently the desktop RTX4060 might use the AD107(the RTX4060TI might use the AD106).
 
Last edited:
Associate
Joined
31 Dec 2010
Posts
2,490
Location
Sussex
The RX7600S which uses a cut down 6NM die compared to a full TSMC 4N 5NM AD107:

At similar power the RX7600S has similar rasterised performance but OFC worse RT performance. Surprised considering the RTX4060 mobile is on a much better process node.

Apparently the desktop RTX4060 might use the AD107(the RTX4060TI might use the AD106).
Guess that additional RT performance¹ and those tensor sensors² do take to a lot of space.

Navi 33 seems to be designed to be extra cheap to produce. Nothing wrong with as long as that is reflected in its price.

But for ¹, I actually believe that having most of that done at the shader level like AMD do makes more sense. Longer term, I really expect some fixed function units but not just for rays but also since of the setup as currently RT puts quiet a large load on the CPU.

For ²? Well, don't really think tensors are required for consumer loads but Nvidia have marketed then to death now so we shall see.

Rather than buying into Nvidia's marketing, I wish PCMR would start complaining that - like with Intel iGPUs on CPUs - they don't want to pay for things they don't use like they did with the old "I'd rather have more cores". Of course, that would force Nvidia to use separate chips for games and professional work.
 
Soldato
Joined
9 Nov 2009
Posts
24,932
Location
Planet Earth
Guess that additional RT performance¹ and those tensor sensors² do take to a lot of space.

Navi 33 seems to be designed to be extra cheap to produce. Nothing wrong with as long as that is reflected in its price.

But for ¹, I actually believe that having most of that done at the shader level like AMD do makes more sense. Longer term, I really expect some fixed function units but not just for rays but also since of the setup as currently RT puts quiet a large load on the CPU.

For ²? Well, don't really think tensors are required for consumer loads but Nvidia have marketed then to death now so we shall see.

Rather than buying into Nvidia's marketing, I wish PCMR would start complaining that - like with Intel iGPUs on CPUs - they don't want to pay for things they don't use like they did with the old "I'd rather have more cores". Of course, that would force Nvidia to use separate chips for games and professional work.

If you read some of the bumpf about RDNA2 and the new consoles,it was about how AMD added RT to the existing RDNA1 designs with relatively few extra transistors. This definitely has some advantages for AMD for IGPs used in their APUs,and the ARM based SOCs which Samsung are designing too. The RDNA2 and RDNA3 IGPs look fantastic too:

So it does seem RDNA3 is performing well,and it makes me wonder when Navi 31 performance is being negatively affected by the chiplet design. I wonder how it would have performed as a monolithic design made on the same TSCM 4N 5NM Nvidia is using??
 
Soldato
Joined
6 Feb 2019
Posts
17,874
Guess that additional RT performance¹ and those tensor sensors² do take to a lot of space.

Navi 33 seems to be designed to be extra cheap to produce. Nothing wrong with as long as that is reflected in its price.

But for ¹, I actually believe that having most of that done at the shader level like AMD do makes more sense. Longer term, I really expect some fixed function units but not just for rays but also since of the setup as currently RT puts quiet a large load on the CPU.

For ²? Well, don't really think tensors are required for consumer loads but Nvidia have marketed then to death now so we shall see.

Rather than buying into Nvidia's marketing, I wish PCMR would start complaining that - like with Intel iGPUs on CPUs - they don't want to pay for things they don't use like they did with the old "I'd rather have more cores". Of course, that would force Nvidia to use separate chips for games and professional work.


It would have been cool if instead of sticking RT cores on GPUs Nvidia stuck them on an add-on card like physx back in the day. They'd be able to dedicate the entire GPU to raster and dedicate even more RT cores to the add on card
 
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
I'm thinking of buying a RX 7900 XT at some point, if I can buy a used or new one for ~£650.

I could wait for a RX 7800 tier GPU (Navi31) but I'm fairly sure these will cost £600 or more...
 
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
oof.

They're recommending a RTX 3090 /rx 6950 XT to play this game at 4K 60 FPS:

Dead-Island-2-PC-specs.jpg


For some reason, a Ryzen 7700X is deemed insufficient to play smoothly at 4K Ultra :cry:
 
Last edited:
Associate
Joined
26 Jun 2015
Posts
720
oof.

They're recommending a RTX 3090 /rx 6950 XT to play this game at 4K 60 FPS:

Dead-Island-2-PC-specs.jpg


For some reason, a Ryzen 7700X is deemed insufficient to play smoothly at 4K Ultra :cry:
Those were the GPUs for 4k 60 , seems normal.

Dunno why people think you can do with less
 
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
Techpowerup's rumoured spec is up for the RX 7800 XT, which will apparently be based on the Navi32 GPU die:


It looks okay, seems like it will perform similarly to the top Navi21 GPUs.

And best the RTX 3080 and RTX 4070.

I think it would be a shame if there's no more consumer Navi31 GPUs released...
 
Last edited:
Soldato
Joined
9 Nov 2009
Posts
24,932
Location
Planet Earth
It appears the full Navi 33 is being called the RX7600 non-XT and will be an 8GB VRAM card:

So I wonder if the RX7600XT might be reserved for a model with more VRAM? I hope this is under $300/£300,because this looks more like an overclocked RX6650 8GB with some RDNA3 improvements,so probably is no better than an RX6700 10GB/RX6700XT 12GB but with less VRAM.
 
  • Haha
Reactions: TNA
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
Unfortunately, it looks like the Navi31 yields (RX 7900 XT and XTX) have sucked so far, based on stock levels I've checked on a couple of sites.

Which likely means AMD has invested a lot, but not yet seen much return on their investment for RDNA3.

Especially compared to the RTX 4070, which has 400-500 items in stock for some models (e.g. Palit).

Might be worth waiting for though.

Edit - One site I checked has a couple hundred 7900 XTs in stock for some Sapphire models, so maybe it's not so bad.

Could be that some of the N31 dies are being reserved for workstation GPUs (due to become available soon) to maximize the profits.

I'd put money on the RX 7800 XT having better yields, assuming it's a Navi32 (5nm) based card.
 
Last edited:
Associate
Joined
13 Sep 2010
Posts
2,077
That rumoured n32 7800xt seems to land exactly where I thought it should, just need to see how that translates to pricing.

Still leaves room for another card in between though.... 7800xtx or 7900 (non XT).... something like 64/72 CU and 16GB(256bit)/18GB(288)/20GB(320), depending on how that affects the IO die config.
Regarding yields, I guess it depends whether the defect rate is higher on a fully enabled n32, or a heavily shaved n31
 
Soldato
OP
Joined
30 Jun 2019
Posts
7,876
It would almost certainly be a 70 Compute Unit model based on Navi31 (just like the W7800), but because the yields don't look great, I'm not sure if AMD will bother.

Maybe in 6-9 months?

It doesn't bode well for RDNA4 (likely 100 CUs or more), if AMD + TSMC are struggling to produce GPUs with 70 CUs on 5nm.
 
Last edited:
Associate
Joined
13 Sep 2010
Posts
2,077
Ah, I'm not familiar with the workstation lineup - was just thinking in logically divisible steps from full die.
Yeah, if they're having trouble getting a 70cu sku reliably, that really could spell bad times...
 
Back
Top Bottom