Discussion Radeon 6500XT and 6400

[GodisanAtheist]

Just getting a thread up for this. Navi 24 will ride... Q1 2022... ish.

AMD Readies Radeon RX 6500 XT and RX 6400 Graphics Cards

AMD is preparing to wrap up its Radeon RX 6000 series desktop discrete graphics card family with two new SKUs, the RX 6500 XT and the RX 6400. The two debut the company's smallest piece of silicon based on the RDNA2 graphics architecture, codenamed "Navi 24," to the desktop space. The RX 6500 XT...

www.techpowerup.com

I fugure the 6500XT lands ~5500XT territory for ~$200.

 

[LightningZ71]

Look at the amount of die area taken up by the 16GB L3 cache in the CCX. To be useful, that's about the amount of space that adding an IfC would take on Rembrandt. That's a significant amount of space and would seriously dent their die per wafer throughput in a market that's literally devouring ANYTHING that's thrown at it. For Rembrandt, there was just no competitive need. What they have is significantly better than Xe in Tiger Lake and likely also in Alder Lake. Raptor Lake isn't looking to make a significant change there either. It's going to be 2023 before Intel has anything that's going to be threatening AMD in that regard. AMD's biggest performance deficit is in their CPU performance, and that's certainly where their next APU is expected to make the biggest changes. While there will likely be some tweaking of the iGPU for it's move to N5, It should still be largely similar.

We might see something significant for 2024 on an improved/tweaked N5 process based around RDNA3 and an IfC.

 

[Shivansps]

-Its weird cause I thought that was sort of the whole point of RDNA2... Running fast under restricted bandwidth situations.

My first thought was "this thing is going to take IGPs by storm".

Yet... No IC for IGPs with RDNA2 (wow that was a mouthful)? I understand having a tight transistor and power budget there, but letting IGP and Cores share a nice fat IC L3 pool seems like a no brainer.

Maybe the worry was the IGP would cannibalize their low end or add in card sales?

It looks to me they havent figured that one out yet, it may come with RDNA3/Zen4 APUs.

But, RMB IGP still have a larger than usual L2. It has to be doing something.

 

[GodisanAtheist]

It looks to me they havent figured that one out yet, it may come with RDNA3/Zen4 APUs.

But, RMB IGP still have a larger than usual L2. It has to be doing something.

-Zen 4 is rumored (confirmed?) to have every processor ship with an IGP on the IO die.

By getting away from monolithic APUs where any additional die space has a disproportionate impact on yields due to an already relatively large die size, we may see AMD include a splash of IC on the IO die since its already reasonably compact to begin with.

Then again, Zen4 will be DDR5 only which should naturally push up bandwidth available to the IGP, so AMD might just lean on that exclusively for a performance bump.

Maybe Zen5? Anyhow, this is way off topic at this point.

 

[waffleironhead]

People won't be stupid enough to pay $280 for a $100 card. Right now, most of them think $80 above MSRP is bearable. Their minds will give them a hard brake if the difference is $180.

The $250 rx550 people have been buying previous to the 6500xt coming out says hello.

 

[waffleironhead]

Looks like we're back to AMD's derped out naming schemes. The 680M (WHY AMD WHY CAN'T YOU JUST STICK WITH A GOD DAMN NAMING CONVENTION?!) is now faster than the 6500XT, in at least a couple scenarios.

Que OEMs pairing the 6900HX with a 6500XT and the weaker discrete card bringing the overall laptop performance down...

Oh AMD, you were executing too flawlessly for too long. Got too close to the sun I guess.

Would you prefer they keep their old naming scheme, where various IGP were all called vega8 regardless of their clockspeed? I mean the Vega8 in the 5700g is leaps ahead of the vega8 in the 2200g.

 

[LightningZ71]

It looks to me they havent figured that one out yet, it may come with RDNA3/Zen4 APUs.

But, RMB IGP still have a larger than usual L2. It has to be doing something.

Are you sure that Rembrandt's iGPU L2 cache is larger than the 6500xt? Looking at the TPU database, they have them both listed as the same. Granted, due to the 680m having fewer CUs, it has more L2 per CU, so, that may be your point.

 

[AtenRa]

It's shameful for 6500XT as a dGPU to even be compared to an iGPU. And now AMD will release the even more embarrassing 6500 soon. As Anand used to say, there is no bad product, only bad prices. 6500XT is VERY badly priced. I hope AMD kills it soon and doesn't produce more of this crap.

Why it is shameful to compare those two ??

680M = 12 Compute Units or 768 Cores
6500XT = 16 Compute Units or 1024 cores

 

[AtenRa]

People won't be stupid enough to pay $280 for a $100 card. Right now, most of them think $80 above MSRP is bearable. Their minds will give them a hard brake if the difference is $180.

They paid 300+ for the GTX1650 which is a $150 MSRP card

 

[igor_kavinski]

Why it is shameful to compare those two ??

680M = 12 Compute Units or 768 Cores
6500XT = 16 Compute Units or 1024 cores

15W TDP vs. 107W. It should be a no contest. Yet the 680M comes out on top in RT. Someone at AMD needs to be demoted, if not fired.

 

[igor_kavinski]

6500 XT was billed by AMD as the "For Gamers, NOT for miners" card. They succeeded in the latter but not in the former. They insulted gamers. I'm waiting to see if Nvidia pays more respect to gamers with their 3050 4GB card.

 

[igor_kavinski]

They paid 300+ for the GTX1650 which is a $150 MSRP card

Nvidia cards have always been worth more than AMD cards. Even then, I do admit that I may have underestimated the stupidity of some people.

 

[GodisanAtheist]

Would you prefer they keep their old naming scheme, where various IGP were all called vega8 regardless of their clockspeed? I mean the Vega8 in the 5700g is leaps ahead of the vega8 in the 2200g.

-I'd prefer they kept it in line with the rest of their GPU arch.

12CU RDNA2 should be 6400i or 6300i or something like that.

 

[Shivansps]

Are you sure that Rembrandt's iGPU L2 cache is larger than the 6500xt? Looking at the TPU database, they have them both listed as the same. Granted, due to the 680m having fewer CUs, it has more L2 per CU, so, that may be your point.

Yes, AMD said 2MB L2, RX6500XT has 1MB, the Van Goth on Steam Deck also has 1MB. Its not a lot, but ill take that over nothing.

 

[LightningZ71]

That's probably a bigger deal than we are thinking and likely explains some of the benchmarks that we're seeing.

 

[TESKATLIPOKA]

Double post.

 

[TESKATLIPOKA]

15W TDP vs. 107W. It should be a no contest. Yet the 680M comes out on top in RT. Someone at AMD needs to be demoted, if not fired.

Incorrect.
It was 6900HX and that is a 45W APU. Even the IGP part should consume more than 15W.
BTW, It was already explained why 6500XT loses to Rembrandt in it.
Choose a game without RT or high memory needs, then It will be naturally faster.

 

[igor_kavinski]

AMD Radeon 680M Specs

AMD Rembrandt, 2200 MHz, 768 Cores, 48 TMUs, 32 ROPs

www.techpowerup.com

Says 15W.

RT or not, a dGPU named 6500 "XT" should not be losing to an iGPU.

The model name is incorrect. Should be 6200 non-XT and the upcoming 6500 should be 6100 non-XT.

There's a marketing idiot in AMD's GPU division. His balls need some nice electroshock therapy so he doesn't pull something like this again.

 

[waffleironhead]

-I'd prefer they kept it in line with the rest of their GPU arch.

12CU RDNA2 should be 6400i or 6300i or something like that.

Ok, so now you're ok with them renaming their IPU, you just dont like how they are renaming them. Gotcha.

 

[TESKATLIPOKA]

AMD Radeon 680M Specs

AMD Rembrandt, 2200 MHz, 768 Cores, 48 TMUs, 32 ROPs

www.techpowerup.com

Says 15W.

And that should prove what? 15W R7 6800U also uses 680M as an IGP(official link), and It certainly won't use the whole shared TDP for just the IGP.
There is no official info how much It actually consumes from the shared TDP as far as I know, but with a shared TDP of at least 45W I would expect a lot more W going to the IGP I wouldn't even be surprised If It was 15W(CPU) and 30W(IGP) for example.

RT or not, a dGPU named 6500 "XT" should not be losing to an iGPU.

The model name is incorrect. Should be 6200 non-XT and the upcoming 6500 should be 6100 non-XT.

There's a marketing idiot in AMD's GPU division. His balls need some nice electroshock therapy so he doesn't pull something like this again.

It was already mentioned why 6500XT performs worse(4GB + 4xPCIe), and that only in an RT title. Show me a non-RT test where this IGP beats RX 6500XT?
There is no such test!
Both 6400 and 6500XT are weaker than RX 6600 (XT), so there is no big problem with the name, the IGP has a totally different name.

 

[Shivansps]

It was already mentioned why 6500XT performs worse(4GB + 4xPCIe), and that only in an RT title. Show me a non-RT test where this IGP beats RX 6500XT?
There is no such test!
Both 6400 and 6500XT are weaker than RX 6600 (XT), so there is no big problem with the name, the IGP has a totally different name.

Maybe no now but as things are, there may be eventually, specially vs the desktop version of RMB. The problem the 6500XT has is that there are cases where the effective bandwidth is a lot lower than 144GB/s + wharever the 16MB IC gives, due to texture streaming. And texture streaming is here to stay.

 

[Panino Manino]

Yes, AMD said 2MB L2, RX6500XT has 1MB, the Van Goth on Steam Deck also has 1MB. Its not a lot, but ill take that over nothing.

View attachment 57976

That's it? Just because the iGPU has more L2 cache it is faster on RT because the dGPU has less L2? That simple a reason?

 

[LightningZ71]

While I'm sure that there are some other factors to the speed difference, if the iGPU indeed has twice the L2 as compared to the rx5500 xt, then, it has twice the capacity for ray stream buffers and tree buffers in cache as opposed to going to main memory. Having a sufficient amount of those buffers is important for ray tracing performance. Granted, I'm not sure why they would over-provision the iGPU over the dGPU in this case, as it doesn't make a lot of sense from the outside, however, the larger L2 doesn't just serve the RT units, instead, it works to reduce the demands on the main memory to a certain degree, which is extra important in an iGPU. It may have been cheaper for AMD to stuff a doubled L2 in Rembrandt's iGPU than to include an infinity cache and scale well enough to make it worth the impact to the overall chip size.

Personally, I'm not 100% convinced that there IS a difference between the L2 cache in the iGPU and the 6500XT yet as I keep seeing conflicting specs for both units. It would seem to neatly explain the performance differences, but, its also likely that something else is going on.

 

[Leeea]

At this market class why does anyone care about RT?

RT is an unplayable experience on all of these. nvidia, amd, iGPU, etc. RT is pointless on the low end, completely pointless!

Making arguments based off of unplayable experiences is pointless stupidity!

 

[Panino Manino]

At this market class why does anyone care about RT?

RT is an unplayable experience on all of these. nvidia, amd, iGPU, etc. RT is pointless on the low end, completely pointless!

Making arguments based off of unplayable experiences is pointless stupidity!

But it looks nice when you have youtubers showing Cyperpunk with RT running at playable framerates.
Also, it was a promise that AMD made.

 

[maddie]

But it looks nice when you have youtubers showing Cyperpunk with RT running at playable framerates.
Also, it was a promise that AMD made.

On this site, we should aim to be wiser than the tech plebs. Ray Tracing was always going to be a check box item for low end this generation and probably several more to come.