Support NeoGAF

[Shogmaster]

RDNA 4 was a huge improvement over RDNA 3 in Q1 2025, giving us FSR 4 image scaling and big improvements in Ray Tracing.

Now RDNA 5 will be coming out in 2027 (Q2? Q3?), with even bigger improvements, using TSMC N3P. So let's look at the leaked line up with speculated clocks, RAM, and performance vs RDNA 4 / RTX. (leaks says 1.1x raster and 2X RT of RDNA 4):

RX 10050XT? (AT4) -
24 CU @ 3.5 GHz (?)
12~16GB 128bit LPDDR5X 10000 (?) 160 GB/s
~95% raster, ~175% RT of 9060XT
~100% raster, ?RT of RTX 3070
$300~350

RX 10060XT? (AT3) -
48 CU @ 3.4 GHz (?)
16~24GB 256bit LPDDR5X 10000 (?) 320 GB/s
~190% raster, ~340% RT of 9060XT
~105% raster, ?RT of 4080 Super
$450~550

RX 10070XT? (AT2) -
68 CU @ 3.2 GHz (?)
16~24GB 196bit GDDR7 864 GB/s
~125% raster, ~227% RT of 9070XT
~ 104% raster, ?RT of RTX 5080
$700~800

RX 10090XT? (AT0) -
96 CU @ 3 GHz (?)
24~36GB 384bit GDDR7 (?) 1.7TB/s
~165% raster, ~300% RT of 9070XT
~90% raster, ?RT of RTX 5090
$1000~1200

I am soooo fucking excited for RDNA 5....

 

[Durin]

I'm rooting for them to be competition for Nvidia, and I'd like them to pump out some better chips for handhelds.

 

[tusharngf]

They've let everyone down every single time. I am not expecting much from AMD.

 

[Shogmaster]

I'm rooting for them to be competition for Nvidia, and I'd like them to pump out some better chips for handhelds.

Binned and down clocked Madusa Point Premium with AT4 will be your handheld SoC. It will destroy Arc G3.

 

[Gameplay Gods Bless]

RDNA 4 was a huge improvement over RDNA 3 in Q1 2025, giving us FSR 4 image scaling and big improvements in Ray Tracing.

Now RDNA 5 will be coming out in 2027 (Q2? Q3?), with even bigger improvements, using TSMC N3P. So let's look at the leaked line up with speculated clocks, RAM, and performance vs RDNA 4 / RTX. (leaks says 1.1x raster and 2X RT of RDNA 4):

RX 10050XT? (AT4) -
24 CU @ 3.5 GHz (?)
12~16GB 128bit LPDDR5X 10000 (?) 160 GB/s
~95% raster, ~175% RT of 9060XT
~100% raster, ?RT of RTX 3070
$300~350

RX 10060XT? (AT3) -
48 CU @ 3.4 GHz (?)
16~24GB 256bit LPDDR5X 10000 (?) 320 GB/s
~190% raster, ~340% RT of 9060XT
~105% raster, ?RT of 4080 Super
$450~550

RX 10070XT? (AT2) -
68 CU @ 3.2 GHz (?)
16~24GB 196bit GDDR7 864 GB/s
~125% raster, ~227% RT of 9070XT
~ 104% raster, ?RT of RTX 5080
$700~800

RX 10090XT? (AT0) -
96 CU @ 3 GHz (?)
24~36GB 384bit GDDR7 (?) 1.7TB/s
~165% raster, ~300% RT of 9070XT
~90% raster, ?RT of RTX 5090
$1000~1200

I am soooo fucking excited for RDNA 5....

1.1 raster is dog shit thats 10% improvement, the 5080 brought 13% raster gains over the 4080 and that was on the same exact node as the 4080. People are still clowning on the 5080s gains to this day. I expect at least 20% gains over RDNA4.

 

[Tobimacoss]

Is this a new leaked lineup?

I thought AT1 was 96 CUs. Wtf happened to AT0 with 192 CUs?

Binned and down clocked Madusa Point Premium with AT4 will be your handheld SoC. It will destroy Arc G3.

I don't think Xbox Handhelds will be using binned chips. Magnus uses 8 Zen 6C cores, 3 Zen 6 cores unlike Medusa Point Premium/mini.

Medusa Point Premium is just for Windows Mini PCs and non Xbox Handhelds, like given to Lenovo etc.

My theory is only Magnus chiplet devices will be branded Xbox. ROG Ally/X being the only exception because they had no choice at the time.

 

[Loomy]

Time for a rebrand of the line

 

[Lagspike_exe]

Sources for this?

 

[simpatico]

10700 having 16GB would be a tragedy.

 

[Mr.Phoenix]

They've let everyone down every single time. I am not expecting much from AMD.

They will be alright... the only year AMD GPUs show any real innovation is whatever year consoles are releasing... then they spend the next 7 years iterating on whatever architecture showed up in the consoles.

 

[willothedog]

leaks, like the 9080XT with 32GB which was 30% faster than the 70...

 

[Shogmaster]

1.1 raster is dog shit thats 10% improvement, the 5080 brought 13% raster gains over the 4080 and that was on the same exact node as the 4080. People are still clowning on the 5080s gains to this day. I expect at least 20% gains over RDNA4.

AMD was already doing well against NVidia in raster so they concentrated on RT. No big wup.

[EDIT]
4080 to 5080 had CUDA core increase. 9070XT to 10070XT will have CU uptick and will give you 25% increase, which bitch slaps 13% increase that 5080 had.

 

[OLED_Gamer]

I'm in for RDNA5. Would be my fourth AMD GPU in a row (previous 6700xt, 7900xt, current 9070xt).

Looking at either the highest end or 2nd from the top so long as the gains are solid.

 

[Shogmaster]

Sources for this?

Kepler and MLiD

 

[Shogmaster]

Is this a new leaked lineup?

I thought AT1 was 96 CUs. Wtf happened to AT0 with 192 CUs?

I think 192 CU is only for server/AI purposes.

I don't think Xbox Handhelds will be using binned chips. Magnus uses 8 Zen 6C cores, 3 Zen 6 cores unlike Medusa Point Premium/mini.

Medusa Point Premium is just for Windows Mini PCs and non Xbox Handhelds, like given to Lenovo etc.

My theory is only Magnus chiplet devices will be branded Xbox. ROG Ally/X being the only exception because they had no choice at the time.

All AMD Z series handheld SoC have been binned XXX Point SoCs. Z2 Extreme is binned HX 370. Literally.

 

[twilo99]

They've let everyone down every single time. I am not expecting much from AMD.

I've been using AMD cards for almost a decade now .. haven't let me down yet

 

[Tobimacoss]

I think 192 CU is only for server/AI purposes.


All AMD Z series handheld SoC have been binned XXX Point SoCs. Z2 Extreme is binned HX 370. Literally.

192 CU AT0 is for Cloud yes, multipurpose for Azure Enterprise, AI, Xbox Cloud Gaming. They would likely divide that into 2-4 instances, 2 x 96 CUs or 4 x 48 CUs. However, reports were AMD was going to make a Flagship Discrete GPU out of that die with 154 CUs. To compete with the RTX 6090.

Z series lineup was cancelled according to KeplerL2, so no Z3. There was another report corroborating him.

The APUs that MS would be supplying to OEMs are going to be Magnus powered, exact same specifications. OEMs can buy non Magnus APUs and do whatever.

Magnus powered portfolio of devices are going to be closed bootloader, at the very least the Console form factors.

 

[Shogmaster]

192 CU AT0 is for Cloud yes, multipurpose for Azure Enterprise, AI, Xbox Cloud Gaming. They would likely divide that into 2-4 instances, 2 x 96 or 4 x 48. However, reports were AMD was going to make a Flagship Discrete GPU out of that die with 154 CUs. To compete with the RTX 6090.

Z series lineup was cancelled according to KeplerL2, so no Z3. There was another report corroborating him.

The APUs that MS would be supplying to OEMs are going to be Magnus powered, exact same specifications. OEMs can buy non Magnus APUs and do whatever.

Magnus powered portfolio of devices are going to be closed bootloader, at the very least the Console form factors.

Yeah, AT1 and AT0 have conflicting info from different sources, which is fun AMIRITE

 

[kasrkinsquad]

1.1 raster is dog shit thats 10% improvement, the 5080 brought 13% raster gains over the 4080 and that was on the same exact node as the 4080. People are still clowning on the 5080s gains to this day. I expect at least 20% gains over RDNA4.

Consumers are pretty satisfied with existing tech so AMD and co have no interest in major upgrades. They are just gonna milk us. I have a 7900xtx and I'm sitting this out. 96CU (same as my 7900xtx) with better ray tracing and AI is cool.

For example I'm 99.99% sure the Z3 Extreme will be 20CUs. It's just better to keep what you have and wait for a major upgrade.

 

[YCoCg]

 

[Boss Mog]

I wanna say those prices are terrible, but with the current state of things they're actually quite good. That naming has to change though. 10070 should at least change to 10700 if you want to keep 5 digits or you could replace the 10 by the roman numeral X and have it only be 4 digits X070.

 

[Victhemic]

RDNA 4 was a huge improvement over RDNA 3 in Q1 2025, giving us FSR 4 image scaling and big improvements in Ray Tracing.

Now RDNA 5 will be coming out in 2027 (Q2? Q3?), with even bigger improvements, using TSMC N3P. So let's look at the leaked line up with speculated clocks, RAM, and performance vs RDNA 4 / RTX. (leaks says 1.1x raster and 2X RT of RDNA 4):

RX 10050XT? (AT4) -
24 CU @ 3.5 GHz (?)
12~16GB 128bit LPDDR5X 10000 (?) 160 GB/s
~95% raster, ~175% RT of 9060XT
~100% raster, ?RT of RTX 3070
$300~350

RX 10060XT? (AT3) -
48 CU @ 3.4 GHz (?)
16~24GB 256bit LPDDR5X 10000 (?) 320 GB/s
~190% raster, ~340% RT of 9060XT
~105% raster, ?RT of 4080 Super
$450~550

RX 10070XT? (AT2) -
68 CU @ 3.2 GHz (?)
16~24GB 196bit GDDR7 864 GB/s
~125% raster, ~227% RT of 9070XT
~ 104% raster, ?RT of RTX 5080
$700~800

RX 10090XT? (AT0) -
96 CU @ 3 GHz (?)
24~36GB 384bit GDDR7 (?) 1.7TB/s
~165% raster, ~300% RT of 9070XT
~90% raster, ?RT of RTX 5090
$1000~1200

I am soooo fucking excited for RDNA 5....

What's the estimate on TF performance for each? YES, TF performance DOES mean A LOT.

 

[Gameplay Gods Bless]

They've let everyone down every single time. I am not expecting much from AMD.

The only ones who brag about AMD GPUs are console gamers.

10700 having 16GB would be a tragedy.

RAM-ageddon is slowing down memory gains, we'll see what happens

 

[Hohenheim]

If their RDNA5 top card (10090XT ?) is close to 5090, i'll probably get that as replacement for my 5080 in my living room "Steam Machine " PC.
Unless the 6080 is miles better and have as much VRAM.

 

[HerjansEagleFeeder]

I just hope we're getting proper models with at least 24GB VRAM. I need that for my AI agents. Gaming performance is good as is, I'm not upgrading because of that. Nvidia can go f* itself

 

[PeteBull]

My pov on amd rdna5 lineup- it doesnt have to be at 5090 lvl, but i would want it to at least match 4090 from 2022 aka by 2027 5yo card, with 32gigs of vram but most importantly much better value/much lower price, definitely below 1500$ streetprice(cant believe msrp anymore), preferably 1200$ 4090 equivalent with much improved vs rdna4 rt and ai upscaling capabilities.

TLDR I want actual value, streetprices dont need to match their fake msrp but currently highend(anything above 5070ti) is extremly overpriced and bad value, so thats where i see amd has nvidia's weakness ready to exploit, ofc if they fuck it up and charge 1999$ for their flagship i wont be surprised either

 

[Shogmaster]

What's the estimate on TF performance for each? YES, TF performance DOES mean A LOT.

LOL. If they are still doing dual issue FP32 for RDNA 5, then just do the math:

AT0: 96 x 3 x 128 x 2 = 73.7 TFLOPs.

But hold up, 9070XT is 48.7 TF already, so that doesn't reflect the gigantic 300% RT gains at all!

WHAT WILL WE DO! Will they go TRIPLE ISSUE FP32!? QUAD?!?@????

FLOPs are DUM brah.

 

[Shogmaster]

I wanna say those prices are terrible, but with the current state of things they're actually quite good. That naming has to change though. 10070 should at least change to 10700 if you want to keep 5 digits or you could replace the 10 by the roman numeral X and have it only be 4 digits X070.

RX X070XT?

Can I get an X? 'BING! BING! BING!'
I would like to solve the puzzle.

[EDIT]
I've got a better naming scheme. C for 100:

RX 1C70 XT

Then the gen after can be RX 2C70XT. So on and so forth.

 

[Shogmaster]

If their RDNA5 top card (10090XT ?) is close to 5090, i'll probably get that as replacement for my 5080 in my living room "Steam Machine " PC.
Unless the 6080 is miles better and have as much VRAM.

60 series is 2028, so if thry dont have huge gains, then they ded.

My bet is they do have the gains, launch with same MSRP as 50 series, but will never come close to.selling at MSRP on the street. LOLOL

 

[hardwood007]

Consumers are pretty satisfied with existing tech so AMD and co have no interest in major upgrades. They are just gonna milk us. I have a 7900xtx and I'm sitting this out. 96CU (same as my 7900xtx) with better ray tracing and AI is cool.

For example I'm 99.99% sure the Z3 Extreme will be 20CUs. It's just better to keep what you have and wait for a major upgrade.

Based on what Kepler said, there won't be a Z3 Extreme. AMD is losing interest and not investing in the handheld space

 

[Shogmaster]

Based on what Kepler said, there won't be a Z3 Extreme. AMD is losing interest and not investing in the handheld space

I could see that, but at that point, ODMs will simply use stock Madusa Premium and use it on their handhelds.

It would certainly be cheaper than Current Strix Halo handhelds, and maybe even cheaper than Arc G3E handhelds.

 

[Victhemic]

LOL. If they are still doing dual issue FP32 for RDNA 5, then just do the math:

AT0: 96 x 3 x 128 x 2 = 73.7 TFLOPs.

But hold up, 9070XT is 48.7 TF already, so that doesn't reflect the gigantic 300% RT gains at all!

WHAT WILL WE DO! Will they go TRIPLE ISSUE FP32!? QUAD?!?@????

FLOPs are DUM brah.

They aren't. You are.

 

[Shogmaster]

They aren't. You are.

OK then. $1200 RX 10090XT only has ~40% more Floppity FLOPs than $600 9070XT.

You happy?

 

[Shogmaster]

I just hope we're getting proper models with at least 24GB VRAM. I need that for my AI agents. Gaming performance is good as is, I'm not upgrading because of that. Nvidia can go f* itself

Madusa Halo should have upto 196GB of RAM so if you are happy with better than RTX 4080 Super gaming performance, just get an AI Max+ 588 ITX mobo and build a SFF box? 588 will have 12 core Zen 6 + 2 core Zen 6 LP to entice you.

 

[Cumulus]

You guys really don't learn....

Hyping yourselves just to get slapped in the face by reality in the end.
Every AMD GPU gen the same drama.

5090ish performance for a grand, lol....

 

[Shogmaster]

You guys really don't learn....

Hyping yourselves just to get slapped in the face by reality in the end.
Every AMD GPU gen the same drama....

RDNA 4 slapped the right way so I dunno wat yer talking about elf boi.

 

[Gameplay Gods Bless]

Consumers are pretty satisfied with existing tech so AMD and co have no interest in major upgrades. They are just gonna milk us. I have a 7900xtx and I'm sitting this out. 96CU (same as my 7900xtx) with better ray tracing and AI is cool.

For example I'm 99.99% sure the Z3 Extreme will be 20CUs. It's just better to keep what you have and wait for a major upgrade.

It depends on the type of person, I think console gamers seem satisfied with the PS5 but the 50 series getting attacked as hard as it was was due to people not being satisfied with current performance. I have a 5080 and I would have loved if the 5080 had more VRAM, memory bandwidth and compute. If the 5080 was 50% faster than a 4090 I would have been a happy camper, instead not even the 5090 reached that performance threshold. Next gen from Nvidia will bring bigger gains (6080 over 5080) but now we'll have to deal with memory prices. I wish the 6080 had 32GB or more of vram but at this point I would be happy with 24.

 

[Shogmaster]

It depends on the type of person, I think console gamers seem satisfied with the PS5 but the 50 series getting attacked as hard as it was was due to people not being satisfied with current performance. I have a 5080 and I would have loved if the 5080 had more VRAM, memory bandwidth and compute. If the 5080 was 50% faster than a 4090 I would have been a happy camper, instead not even the 5090 reached that performance threshold. Next gen from Nvidia will bring bigger gains (6080 over 5080) but now we'll have to deal with memory prices. I wish the 6080 had 32GB or more of vram but at this point I would be happy with 24.

Are you fuming about the 5090SE news rn?

 

[hardwood007]

I could see that, but at that point, ODMs will simply use stock Madusa Premium and use it on their handhelds.

It would certainly be cheaper than Current Strix Halo handhelds, and maybe even cheaper than Arc G3E handhelds.

Issue with Medusa Premium or Halo or whatever it will be called, is that there's too many cpu cores drawing power and bandwidth away from the gpu. The chip isn't geared for gaming. It's the same issue with strip halo, though AMD has released the 388 recently to partially address this. That chip is still too power inefficient for gaming.

ARC G3E is a chip that's tailored for gaming, where the cpu doesn't draw much power and bandwidth. As to your point about Medusa being cheaper than G3E, I doubt that

 

[hardwood007]

At this point, I am hoping AMD will release a variant of the chip that's is going into the ps6 handheld, perhaps one or two years after the ps6 handheld is released.

That chip has low power cpu and a 192 memory bus, which should help with memory bandwidth issues all these mobile chips have.

 

[Sanepar]

10070 XT should do more than 25% UP in Raster than 9070 XT. I would say at least 30-40%

 

[Shogmaster]

Issue with Medusa Premium or Halo or whatever it will be called, is that there's too many cpu cores drawing power and bandwidth away from the gpu. The chip isn't geared for gaming. It's the same issue with strip halo, though AMD has released the 388 recently to partially address this. That chip is still too power inefficient for gaming.

ARC G3E is a chip that's tailored for gaming, where the cpu doesn't draw much power and bandwidth. As to your point about Medusa being cheaper than G3E, I doubt that

Madusa Premium (Kepler) a.k.a Madusa Halo Mini (MLiD) is only using 4+8c+2LP Zen 6 config. Its not doing full Zen 5 like Strix Halo. It has otion to add 12x Zen 6 chiplet, but that will only be for mini PCs.

So its meant yo be mobile from get go. I'm betting full Zen 6 cores are gated off from Compact and LP cores and only kicks in when needed.

G3E is just a binned Panther Lake 358H. It wasn't tailor made for gaming. Its was only tailor marketed for gaming. LOL

As for price, its a tiny weeny die package, but that damn 18A is expensive! By die size Panther Lake/G3E should be like half the price, but 18A yields must be shit.

By contrast, TSMC 3nm is pretty damn mature. Yeild are good, and thus even with bigger dies, Madusa Premium in Q2 2027 should be cheaper than what Arc G3 is now.

 

[Shogmaster]

10070 XT should do more than 25% UP in Raster than 9070 XT. I would say at least 30-40%

My calculation says 25%

9070XT: 64 x 3 = 192
1C70XT: 68 x 3.2 x 1.1 = 239.4 / 192 = 1.25

Pretty simple math, not accounting VRAM bandwidth.

 

[Shogmaster]

At this point, I am hoping AMD will release a variant of the chip that's is going into the ps6 handheld, perhaps one or two years after the ps6 handheld is released.

That chip has low power cpu and a 192 memory bus, which should help with memory bandwidth issues all these mobile chips have.

Extremely unlikely since:

1. Orion is a monolithic SoC, thus AMD would have to create a new RDNA 5 chiplet from scratch

2. Orion probably missing features ftom other RDNA 5, like PS5 missing features from RDNA 2

3. Orion will be using TSMC N3E (efficiency), and won't clock that high vs RDNA 5 chiplets that will use N3P (performance)

 

[HerjansEagleFeeder]

Madusa Halo should have upto 196GB of RAM so if you are happy with better than RTX 4080 Super gaming performance, just get an AI Max+ 588 ITX mobo and build a SFF box? 588 will have 12 core Zen 6 + 2 core Zen 6 LP to entice you.

No, I need a proper desktop gaming rig, not some gimped bullshit AI bait product with unified memory crap. 24+GB will do just fine for the time being, QWEN and other LLM can perform quite well within those constraints already.

 

[Hideous Snake]

I will get excited when I see them being a proper competition for NVIDIA

 

[Shogmaster]

No, I need a proper desktop gaming rig, not some gimped bullshit AI bait product with unified memory crap. 24+GB will do just fine for the time being, QWEN and other LLM can perform quite well within those constraints already.

1. LLM values RAM amount more. 24GB will force you to use heavily quantized models. Hallucinations all day lol

2. No practical difference between AT3 in discrete card and the APU version as far as performance is concerned, unless some card ODM decides to juice the shit out of the card for OC.

What are we doing anyways? Meme image generation? Vibe coding? Making those shitty AI voice over YouTube videos? I suppose that's light years better than starting your videos with "What's up YouTube It's yer Boi blah blah blah" lol

 

[rm082e]

I jumped to the 9070 XT after 13 years of Nvidia cards. I've been quite happy with it. Nvidia makes great cards, but I'm sick of supporting them when they treat gamers like shit.

Based on the leaked specs and expected performance, I will definitely go for their biggest card next time if I can get one.

 

[Rocco Schiavone]

Definitely getting the 10090XT If they really release one and a Ryzen9800X3D(successor) so I most likely will have still better performance than a PS7 Pro.

 

[riko]

LPDDR on dgpus? Seems like a terrible idea.