The AMD 9070 / 9070XT Reviews and Discussion Thread

[Bondrewd]

The majority of sales aren't halo part sales. They just need to offer very good performance at the most popular price points

Oh no no, noooooo.
That only works if you have a Name of Power.
To get one, you gotta win.
Ryzen wasn't a thing until 3950X won most things, and 5950X won it all!
Server is the same story.

Puts it awfully close to their less OCed 9070XT.

It's not a big chop, and the fundamental limit is bandwidth anyway.
Looking at PCBs, these were definitely designed as GDDR7 ones. Guess they went G6 for costs.

 

[DegustatoR]

and the fundamental limit is bandwidth anyway

Doesn't look like it to me considering that 5070 beats even the XT on that.

 

[DegustatoR]

AMD Radeon RX 9070 (XT) im Test: FSR 4 Super Resolution: Analysen und Benchmarks

AMD Radeon RX 9070 (XT) im Test: FSR 4 Super Resolution: Analysen und Benchmarks / FidelityFX Super Resolution musste liefern

www.computerbase.de

So yeah Nvidia's ML h/w seem to be better at doing AI upscaling as well.
Not by much but you can certainly see the effects in these benches from CB.de - in all but R&C benchmarks there 5070 Ti gains more performance when enabling DLSS than 9070 XT does from FSR4.
On a positive side though FSR4 runs on the 9070 XT at about similar performance as FSR3 so there's no downsides to using it on the GPU.

 

[eastmen]

Oh no no, noooooo.
That only works if you have a Name of Power.
To get one, you gotta win.
Ryzen wasn't a thing until 3950X won most things, and 5950X won it all!
Server is the same story.

It's not a big chop, and the fundamental limit is bandwidth anyway.
Looking at PCBs, these were definitely designed as GDDR7 ones. Guess they went G6 for costs.

I gotta disagree , the ryzen 1700 was a huge performer for amd and the start of the market shift.

 

[Bondrewd]

I gotta disagree , the ryzen 1700 was a huge performer for amd and the start of the market shift.

Oh no, 6/8c CFLs neutered its market impact pretty well.
But you couldn't argue against 3950X and 5950X even moreso.

 

[Sega_Model_4]

Avg Perf Differential	3840x2160	2560x1440	1920x1080
RTX 5070 Ti	105.33%	102.93%	101.83%
RX 9070 XT	100.00%	100.00%	100.00%
RX 7900 XTX	99.70%	97.07%	93.04%
RTX 3090 Ti	93.95%	90.31%	88.72%
RTX 4070 Ti Super	93.34%	93.07%	93.17%
RX 9070	89.54%	90.14%	91.18%
RX 7900 XT	86.03%	86.86%	87.31%
RTX 4070 Ti	85.89%	86.93%	88.79%
RTX 3090	83.54%	81.39%	81.37%
RTX 3080 Ti	82.96%	81.10%	80.65%
RTX 5070	81.17%	82.92%	84.20%
RTX 3080	80.13%	77.14%	75.50%
RTX 4070 Super	77.56%	79.72%	81.87%

 

[RobertR1]

The 5070 is so bad lol

 

[arandomguy]

So yeah Nvidia's ML h/w seem to be better at doing AI upscaling as well.
Not by much but you can certainly see the effects in these benches from CB.de - in all but R&C benchmarks there 5070 Ti gains more performance when enabling DLSS than 9070 XT does from FSR4.
On a positive side though FSR4 runs on the 9070 XT at about similar performance as FSR3 so there's no downsides to using it on the GPU.

Is the hardware better or is the software better (or just different) or a combination?

I'd be interested in looking at the VRAM cost for the upscaling and frame gen. If I were AMD and knew my product stack would have less VRAM contention essentially an optimization choice (for their product stack) could be to use a relatively larger model. Or at least I might not be prioritizing keeping the VRAM and cost down as much as possible versus other considerations.

Has it been mentioned if FSR4 is a CNN model? Or a transformer model? Or something else?

 

[DegustatoR]

Is the hardware better or is the software better (or just different) or a combination?

No real way of telling as the s/w is different as well as the h/w.
Still the gains with DLSS are higher and we have no indication that FSR4 does more than DLSS3/4.

Has it been mentioned if FSR4 is a CNN model?

It is a CNN model I believe.

 

[trinibwoy]

But you couldn't argue against 3950X and 5950X even moreso.

AMD doesn’t need a big GPU to win. They just need to be so far ahead in PPA and PPW that no amount of die harvesting from the competition would matter. As it stands today it’s trivial for Nvidia to negate RDNA4 as they hold the advantage on both metrics.

Now of course if AMD had a huge advantage in PPA or PPW then it would make sense to make a big chip too so they win either way.

The main things they need though are consistency and leadership. They can’t do it all in one generation and they can’t win by constantly riding Nvidia's coattails on features.

 

[Bondrewd]

AMD doesn’t need a big GPU to win

yeah they do, Ryzen and EPYC were built in their modern states by cracking the comp skulls in public.

The main things they need though are consistency and leadership.

That's implicit everywhere.
Good execution is nice but the needle ain't moving anywhere unless you get really ambitious.
Which is the problem pretty much only client gfx folks at AMD have.

 

[techuse]

These cards are overpriced. I can’t imagine many people choosing these unless Nvidia prices remain far over MSRP.

 

[DegustatoR]

These cards are overpriced. I can’t imagine many people choosing these unless Nvidia prices remain far over MSRP.

9070 XT has a good MSRP I'd say.
9070 though should probably be a 50 or more less.

 

[Bondrewd]

9070 though should probably be a 50 or more less.

it's made so that you buy the XT.

 

[trinibwoy]

9070 XT has a good MSRP I'd say.

If there's any truth to HUB's info that MSRP is about as real as the 5070's.

 

[techuse]

9070 XT has a good MSRP I'd say.
9070 though should probably be a 50 or more less.

What scenarios can you envision other than a super strict price ceiling, which I don't think is going to be very common at this price point, where someone would determine the Nvidia card is not worth the extra money? Assuming MSRP prices of course.

 

[DegustatoR]

What scenarios can you envision other than a super strict price ceiling, which I don't think is going to be very common at this price point, where someone would determine the Nvidia card is not worth the extra money? Assuming MSRP prices of course.

The most obvious comparison would be 5070 to 9070 XT where for +$50 you get somewhat similar average RT performance (heavy PT generally isn't suitable for 5070 so such comparisons are of dubious value) and quite a bit of advantage over 5070 in non-RT and lite RT titles. I don't think that aiming at Ti comparisons here make much sense as this one is just overall better - dunno if +$150 better but XT's position perf wise between the non-Ti and Ti seem to be closer to the former.

It is just a tad less solid upgrade option for anyone with cards at or below 3080/6800XT level than the Ti and this "tad" is compensated by the price quite nicely I'd say.

The 9070 vs 5070 is a bit harder to parse. A lot here will depend on the actual retail prices on these I'd say.

 

[itsmydamnation]

from a uarch perspective ( not an SKU ) every time someone compares to a 5070ti remember the 9070XT has 40% less bandwidth. the fact its anywhere near it is crazy good from a uarch perspective. Makes me sad there is no 96CU+ 384bit+ card from AMD

 

[trinibwoy]

Linus weighs in. I must be getting old cause it’s hard to follow this new style of videos where people are screaming information at you at top speed.

To his credit he tests video encoding quality and some other non-gaming stuff that others didn’t cover.

 

[trinibwoy]

from a uarch perspective ( not an SKU ) every time someone compares to a 5070ti remember the 9070XT has 40% less bandwidth. the fact its anywhere near it is crazy good from a uarch perspective. Makes me sad there is no 96CU+ 384bit+ card from AMD

The 9070 XT has 60% more fillrate and 10% more flops than the 5070 Ti. To do a proper comparison you probably want to compare full N48 to full GB203 at same power.