AMD Radeon 9070 XT Rumored to Outpace RTX 5070 Ti by Almost 15%

[Rabit]

if price right 9070xt 32 Gb will be mine personal pick for local LLM
Also we have similar situation in past RX 480 4Gb and 8GB not much difference when released by after some time we seen it when game star using above 4GB.

 

[OneSorcerer]

if price right 9070xt 32 Gb will be mine personal pick for local LLM
Also we have similar situation in past RX 480 4Gb and 8GB not much difference when released by after some time we seen it when game star using above 4GB.

Do you really think 16gb vs 32GB is comparable to 4 vs 8GB? Like, you expect a 16GB jump in memory usage in < 5 years?

 

[Sound_Card]

Do you really think 16gb vs 32GB is comparable to 4 vs 8GB? Like, you expect a 16GB jump in memory usage in < 5 years?

He's talking about running a local LLM.

My last guess was conservative, a more hopeful outlook:

9070xt ~ 7900xt raster / 4080 ray - $550 and up
9070 ~ 7900gre raster / 4070 super ray - $450 and up
9060xt ~ 7800xt raster / 4070 ray - $400

 

[Rabit]

Currently consoles have 16GB and we see games use even 16GB Vram, if next gen consoles in 2028 will have 32GB is sure that new games will star utilise it.
This cycle always repeat new gen consoles arrive VRAM requirement increase.
Today 32GB will be nice for LLM or other work application.

 

[mkppo]

This is ridiculous and stupid. rdna3 register files are 196k, while RDNA2 has 128kB, and rtx 30/40/50 has 64kB. The one that really needs improvement here is NV) That's why we see 0% IPC SM improvements for RTX40->50
As for L2, NV went to such a waste of crystal area for one (may be not only) function - SER, which reloads all register files via L2. But they were wrong about the scale of this feature. One game in two years is nothing.

Your claims about no changes are not true. The recommendations for configuring the last level cache and memory are also incompetent.

A calm(er) discussion would be appreciated and I do appreciate your response except the, erm, aggressive part. The reason I said I expect changes to L2/registers isn't necessarily to do with capacity. I was thinking more along flexibility - two operands cannot read from the same register bank and the destination registers also can't be both even or odd. As of now, that massive increase in FP32 throughput isn't really realized in games and such because dual-issue capability is limited and reliant on hand-compiling the code.

As for L2, Ada only does a two tier cache subsystem so their L2 also works to prevent memory bandwidth bottlenecks as compute capacity gets scaled up. RDNA2/3's does 4 way though, and L2 plays a smaller role and L3/Infinity cache rather deals with preventing bandwidth bottlenecks. I can see them rebalancing L2 size along with infinity cache, it's a given because it'll depend on memory configuration and compute throughput among other things. I didn't claim anything about no changes though, nor did i recommend anything specific about L3/IC or memory so i'm not sure what you meant?

 

[medi01]

´´´´´´


9070xt ~ 7900xt for $600 or less
9070 ~ 7900gre for $500
9060xt ~ 7800xt for $400

Sounds right.

Also, HUBs comment, to go lower price upfront, instead of opening with higher MSRP and dropping it later, makes a lot of sense
Reviews rate cards based on MSRP.

Right now RX 7900 XT costs 635€ and RX 7900 GRE 546€. Similar improvement to blackwell

Why are you comparing MSRP to MSRP for NV cards and discounted last gen price to MSRP for AMD cards?

Also, 680 Euro is the cheapest I see here, with most being 700:

https://geizhals.de/?fs=7900xt&hloc=at&hloc=de