AMD Radeon RX 9070 XT, Ryzen 9950X3D, Z2 Extreme SOC for Ally / Legion, & More

AMD announced the existence of its Radeon RX 9070 and RX 9070 XT today, noting Q1 2025 availability.

Specs are extremely limited right now. The company did picture several partner model cards in its announcement slide, including at least one that looks like a Yeston model, an ASRock model, XFX, and all the others listed. The cards pictured above are generally 3-slot coolers that are 2-3 slots thick.

AMD says the GPUs will use its RDNA 4 architecture and will utilize a 4nm process from TSMC. AMD mostly described its specs with adjectives, which is unfortunately not particularly useful. Words like “optimized compute units,” “supercharged AI compute,” and “improved raytracing per CU” don’t tell us a whole lot. We’ll have to wait for that information.

Likewise, AMD announced that FidelityFX Super Resolution, or FSR4, will be released and has been built for RDNA 4. It intends to re-launch its Anti-Lag software solution that was intended to compete with Reflex, now in its Anti-Lag 2 iteration. We were not pre-briefed with any further information than this at the time of briefing.

The company is clearly self-aware, as it also presented a slide about its naming choices for the RX 9070 series. Go figure. The slide above shows that AMD intends to line-up the 9070 series, including both XT and non-XT models, with the RX 7900 XT down to the middle of the RX 7800 XT (read our review), whatever that means. The worst 9070 is apparently half of one RX 7800 XT -- or maybe that means 1.5 7800 XTs? All we’re missing is a note telling us that the image is not to scale.

Anyway, against NVIDIA, this roughly positions the 9070 series as comparable, according to this image, to the 4070 Ti (watch our review) and 4070 Super (read our review). We won’t get out our scrying stones for this hastily thrown-together image since it’s hard to judge without real numbers, but that’s at least how AMD seems to be positioning it.

The company claims the change is to line-up with its Ryzen 9000 CPUs and says it will reserve 8000 naming for its mobile CPUs.

AMD CPUs

On the CPU side, AMD’s two new desktop CPUs are predictable: The 9950X3D and the 9900X3D, which use the Zen 5 architecture that shipped at the end of last year and follow-up the wildly successful 9800X3D. The 9800X3D has been constantly out of stock due to the demand.

The 9950X3D is a 16C/32T part that advertises a maximum boost frequency of “up to 5.7GHz,” noting a 144MB total cache size. At the time of writing this, AMD has not provided pricing details or a specific release date beyond sometime within the next few months. 

For comparison, the 9800X3D (read our review) has a total 104MB cache and $480 MSRP. The advertised boost of the 9800X3D is 5.2GHz, so the 9950X3D has a greater cache size and may benefit from higher boosting. The 144MB cache on the 9950X3D comes from the additional CCD in the configuration and could have some specific benefits that we’ll explore in our eventual review. 

The 9900X3D is a 12C/24T component that advertises up to a 5.5 GHz boost max. It has a 140MB total cache. The 9950X3D runs a 170W TDP, with the 9900X3D at 120W. The 9800X3D is also 120W. Because of this, the 9800X3D may benefit from additional power available during fully loaded workloads. There could be some shuffling of the CPU stack in specific benchmarks due to the power budget differences.

AMD’s updated chipset drivers should more intelligently park CCDs and, in theory, should make it easier to upgrade in-socket without needing to blow away the whole OS prior to moving from a single-CCD part to a dual-CCD part. We will still be using isolated SSDs for our reviews, but this should be a benefit for end users who may later seek to upgrade in-socket.

AMD published some first-party claims. As usual, we’ll have our own numbers soon -- as will basically all other reviewers -- and so you should wait for those prior to making decisions. To set the stage for what we’re verifying against, AMD is claiming the following:

AMD says this is “the world’s best gaming processor,” though note that they are comparing it to the 7950X3D (watch our review). The 9950X3D is shown as being on average 8% better across 40 games that AMD tested, with a range of no change to a 58% uplift over baseline compared to the 7950X3D.

AMD also claims that it outperforms the 285K (read our review) by 20% on average across 40 games. We definitely believe this, based on numbers we ran for the 9800X3D and 285K already.

Its first-party numbers also point to performance improvement in Blender, with lesser improvements in Photoshop and Premiere in PugetBench testing against AMD’s own prior processor.

Historically, these CPUs do not necessarily provide significant improvements over the single-CCD X3D CPUs of the same generation. You might see rough equivalence or slight changes in specific games. The primary advantage would be for someone who does a lot of gaming but also wants the additional cores for production workloads. 

The other historical challenge has been behavior with core parking, something we’ve now detailed extensively. While core parking is still a “thing,” AMD says its new chipset drivers should resolve a lot of the past issues.

AMD Z2 SOC

AMD’s Z2 SOC follows-up the Z1 and Z1 Extreme mobile solutions that were found on some handheld devices. AMD has also offered mobile chips like the 7840U and 8840U that have been in handheld devices and are comparable.

The Z2 is light on information: AMD again defers to descriptors like ‘breathtaking” and “exhilarating speed,” which we assume is the next speed setting for a Back to the Future movie. 

It also gets into some business-y stuff, like the addressable market and increase in competition in this market.

As for actual news, the Z2 family comes in 3 variations currently known: The Z2 Extreme, Z2 Go, and the…Z2 non-Extreme, non-Go.

The Z2 Extreme and Z2 are both 8C/16T parts with the same cache and a boost frequency separated only by 100 MHz. The actual change comes in the form of the integrated graphics. This is also where the Z1 and Z1 Extreme deviated most heavily. The configurable TDP allows up to 5W more driven to the Z2 Extreme, which will cost battery life but help power the GPU. The Z2 has a higher boost clock despite a lower cTDP, likely due to overall package budget allocation with the GPU change (and also density).

The Z2 Go is new. This is a 4C/8T part that only boosts to 4.3 GHz maximum advertised, only has 10 MB of cache, keeps the 15-30W cTDP, and keeps the 12 CUs. This CPU is far weaker than the others listed here, especially with that clock drop, so we’re curious to see what types of devices make the best use of it. We’ll also be curious to see battery life and if it can stick closer to that 15W number while still providing meaningful performance.

These are all listed for Q1 2025 availability, so we’ll be busy on our team running handheld benchmarks once again here soon. We ran several reviews last generation and will need to do a total refresh.

AMD HX3D

We’d like to dedicate this next section to our late friend Gordon Mah Ung, who once joined us to complain about AMD’s mobile CPU naming scheme.

AMD’s new mobile CPU is the FireRange AMD Ryzen 9 9955HX3D that’s releasing in 1H 2025, and it’s joining the 9955HX and 9850HX mobile parts.

The 9955HX3D is advertised as what AMD claims is the best gaming and content creation part for mobile. We don’t really test mobile, but we certainly use high-end laptops for our travel and might try these out.

The 9955HX3D is a 16C/32T part that boosts up to 5.4GHz. It has 144 MB of cache and a TDP of 54W. The 9955HX is the same, but with less cache. The 9850HX is a 12C/24T part with lower boost, no X3D cache by mercy of its easier name, and the same TDP.

AMD also spent some time on its new “AI” brand name mobile processors, but we’ll leave that to someone else to cover as that’s not really our area of focus.