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

[g67575]

There's just no reason for them to do that. They can't produce enough for that anyway, judging by the stock levels on many websites.

Prices should fall a bit when they get more in stock.

Navi32 is the bulk chip, these should be reasonably competitive with Nvidia on pricing.

 

[g67575]

Prices are mostly controlled by fairly simple algorithms these days anyway, on most (maybe all) major retail websites these days.

If people stop buying the latest cards for a while, the stock levels should improve, and hopefully prices too.

 

[g67575]

No, not a chance.

You say all that, but the performance numbers suggest otherwise.

The only thing that might be a disadvantage for AMD, is the lack of a large L2 cache (instead there's infinity cache).

Some people insist that Navi32 has a 192bit memory interface for some reason, it's very likely to be wrong.

The other thing that looks wrong about the speed is the GDDR6 memory speed, it's likely to be the same as the high end RDNA3 GPUs.

 

[g67575]

I think you must be ignoring the FP32 processing power, which is 38.40 TFlops. This is likely to be very accurate, because we know the max number of Compute Units on Navi32 (60), and a clock rate of 2500Mhz, which be quite a modest estimate.

What AMD have done with RDNA3, appears to be similar to what Nvidia did with Ampere, in terms of floating point performance.

This increased floating point performance has been seen on the RDNA3 mobile chips (announced recently).

 

[g67575]

Well, I was basing off the ones on Angstronomics, which were made available ages ago.

​

 

[g67575]

Here you go:

AMD's RDNA 3 Graphics

3rd time's the charm?

www.angstronomics.com

They are for Navi32, not 7800XT specifically.

The article came out months before the release of the Navi31 chips, and it was accurate.

Seems to have been correct in predicting the Compute Units, memory bus and infinity cache of Navi33 chips also:

https://cdn.arstechnica.net/wp-content/uploads/2023/01/rdna3-lapotp-05-2.jpeg

 

[g67575]

Well, the RDNA2 based RX 6600 XT was slower (on average) than the RTX 3060, according to Techspot:

EDIT - Oops, it depends on the resolution actually, which is faster.

Spec wise, it's basically identical to the RDNA3 based RX 7600M XT:

AMD Radeon RX 6600 XT Specs

AMD Navi 23, 2589 MHz, 2048 Cores, 128 TMUs, 64 ROPs, 8192 MB GDDR6, 2000 MHz, 128 bit

www.techpowerup.com

AMD Radeon RX 7600M XT Specs

AMD Navi 33, 2469 MHz, 2048 Cores, 128 TMUs, 64 ROPs, 8192 MB GDDR6, 2250 MHz, 128 bit

www.techpowerup.com

And this is what AMD claims about the 7600M XT's performance:

More games compared here:

https://www.notebookcheck.net/fileadmin/Notebooks/News/_nc3/7600M_XT_2.jpg

Seems like it's around 20% faster than a RTX 3060 8GB, depending on the game.

So, I'd assume at least a 20% performance increase for RDNA3 Compute Units. The floating point (FP32) processing power is approx. double.

That's before taking clock speeds into account.

The Boost clock on the RX 6800 is just 2100 Mhz, so 2500-2600Mhz for Navi32 wouldn't be a surprise.

​

 

[g67575]

So, it looks like I wasn't far off saying a 20% performance improvement for RDNA3 Compute Units:

The magic number is 17.4% for the CUs.

AMD also claims a 20% improvement for the overall silicon design:

https://www.kitguru.net/wp-content/uploads/2022/11/YJUIZmrw-5.jpg

I think this includes cache system improvements also:

https://www.hardwareluxx.de/images/cdn02/uploads/2022/Dec/swanky_web_d3/amd-rdna3-breakout-025_1920px.jpeg

AMD says that there are less infinity cache hits with RDNA3.

I think we will get another 10 or 20% boost in performance from higher clock speeds.

I think the earlier claims of a 50% performance per watt improvement, must have assumed large increases in clock frequency, like rumours suggest we could see with an RDNA3 refresh (and probably conflated things by including compute unit increases in this figure also).

So, it looks to me, like the performance of a 60 Compute unit RDNA3 GPU will be as fast as either an RTX 3080, or RTX 3090 (closer to the RTX 3090), depending on the max boost clock frequencies.

 

[g67575]

We already know they are not 17.4% faster per compute unit.

Perhaps AMD meant the performance improvement is actually half that amount per CU, because Compute Units are paired?

 

[g67575]

Interesting:

Custom AMD Radeon RX 7900 XTX can hit 3.5 GHz clock with Blender, 2.8 GHz in games - VideoCardz.com

AMD Navi 31 OC clocks: up to 2.8 GHz in games, 3.5 GHz in Blender ComputerBase have a detailed overview of AMD RDNA3 and NVIDIA Ada architecture frequency analysis. This generation allows both architectures to reach the 3 GHz barrier with ease, but such clocks will not be seen at all times...

videocardz.com

This suggests that high clocks in games might be possible on some Navi32/Navi33 cards.

 

[g67575]

Is the launch of more AD104 cards imminent? Like in feb. or March?

Or are we talking late Q2?

 

[g67575]

Cheers mods

 

[g67575]

@timorous OK, didn't know what a shader engine was but this helps:

At least we know the total bandwidth is probably a bit higher than Nvidia's AD104 chips.

3 Shader Engines with 20 Compute Units each seems plausible for the top N32 die.

 

[g67575]

I guess one way of estimating the performance of a Navi32 die with 60 CUs, would be to base it on the performance of the top Navi33 chip (7600M XT, with 32 CUs), which is significantly faster than a desktop RTX 3060 (8GB) according to AMD.

I'm gonna assume that it's actually about the same as the RTX 3060 8GB in performance.

I think what we will get with a 60 CU RDNA3 GPU (nearly 2x the CUs of Navi 33), will be a graphics card that is roughly twice as fast as the RTX 3060 8GB, so about the level of a RX 6900 XT:

NVIDIA GeForce RTX 3060 8 GB Specs

NVIDIA GA106, 1777 MHz, 3584 Cores, 112 TMUs, 48 ROPs, 8192 MB GDDR6, 1875 MHz, 128 bit

www.techpowerup.com

In a lot of ways, Navi32 seems to double up on the spec of Navi33 (256 bit memory bus vs 128 bit, 64MB infinity cache, vs 32MB).

My impression is that the 'game clock' is limited to around 2300 Mhz on RDNA 3 (Navi 33 is 2300 Mhz, Navi 31 is 2300 Mhz also), so not sure if we will see much improvement there - Might need high end cooling to go higher?
7600m XT spec:

https://www.amd.com/en/products/graphics/amd-radeon-rx-7600m-xt

The RX 7600m XT is supposed to be released already, so maybe we'll see some real performance figures soon.

 

[g67575]

I'm thinking about buying an RX6800 soon, it's roughly 2x as fast as the card I've got at 1440p and costs ~£490 new.

I will wait a little more, to see if the RTX 4070 is going to come out, I'm beginning to wonder if Nvidia has delayed their launch deliberately.

 

[g67575]

I'd buy a RX 6800, but it can't quite handle Cyberpunk at 1440p, at 60 FPS:

https://www.kitguru.net/wp-content/uploads/2022/12/Cyber2.png

It looks like it can handle Warhammer III though:

https://www.kitguru.net/wp-content/uploads/2022/12/TW2.png

I'm leaning towards either a RX 6800 XT or RTX 3080, but both are still outside what I'm willing to pay.

 

[g67575]

I had a used RTX 3080 before and it performed well until one of the fans started making very loud grinding noises. I think it was probably just a problem with the fans EVGA used.

 

[g67575]

I think the next card in Nvidia's series is gonna be the same price as the RTX 3080. Many people were glad to pay £650 for the RTX 3080, so they will charge that again. There's been some inflation, so I doubt it will be less than that.

I don't think it will be an affordable card on launch, unless there is a Founders Edition version.

One problem Nvidia has, is that lots of people will be looking at buying used RTX 3080s (because there won't be a big performance difference between the next AD104 card and the RTX 3080) again when they launch less expensive RTX 4000 series cards - which of course they won't profit from at all. So I think they will need a reference model to help keep prices closer to MSRP.

Prices are slowly coming down on used RTX 3080s, some cards are selling for ~£500 in auctions...

 

[g67575]

This guy just guesses:

https://twitter.com/kopite7kimi/status/1555053945151975425?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1555053945151975425%7Ctwgr%5E52d845a9e58d0ad522a30e8128321b32f7d50f06%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Fpublish.twitter.com%2F%3Fquery%3Dhttps3A2F2Ftwitter.com2Fkopite7kimi2Fstatus2F1555053945151975425widget%3DTweet

He reckoned that the RTX 4070 would be a 300w card in August 2022.

The RTX 4070 TI has a tdp of 285w.

 

[g67575]

Didn't the 'leakers' miss the fact that Ampere had 2x the shader units per core?

Which was the biggest design change from the previous gen.

Basically proving that they just guess.