AMD's $350 RX 9060 XT Is the Volume Play; the Workstation Parts Are Plumbing
AMD used Computex to fill in the bottom of its RDNA 4 stack and the top of its workstation stack at the same time: a $350 RX 9060 XT with 16GB, a 96-core Threadripper flagship, and a 32GB AI card for people running models locally. Only one of those products sells in volume, and the interesting question is whether the cheap one is the honest one.
One Die, Two Memory Configurations
The 9060 XT arrives as an 8GB card at $300 and a 16GB card at $350, both on June 5, built on a Navi 44 die measuring 199mm² against the 357mm² of the 9070-class silicon. Shared specifications: 32 compute units, 32 ray accelerators, 64 of AMD's AI accelerators, a 3.13 GHz boost clock, PCIe 5.0 x16, DisplayPort 2.1a, HDMI 2.1b, and a 150W to 182W board power window, with the bottom of that range belonging to the 8GB model. A single 8-pin feeds it.
That is a deliberately modest card, and the specification sheet is the part that matters rather than any framerate, because AMD shipped far less first-party testing data than it usually does. The sheet is also where the real competitive argument sits: this is aimed straight at NVIDIA's 5060 Ti, at matching memory capacities, using GDDR6 where the GeForce parts use GDDR7.
The 16GB Model Is the Only One Worth Defending
Fifty dollars is a strange amount of money to decide a product's usefulness, and here it decides everything. Modern titles at 1440p already punish 8GB cards, and VRAM is the one specification you cannot tune around, because the textures either fit or they do not. Double the memory on the same die for a small step in price is not a premium tier. It is the correct configuration, and the 8GB model exists mainly to advertise a $299 price point.
AMD's first-party comparison targets the 8GB 5060 Ti, which is fair on price parity and also the least informative matchup available. The x16 link is the quieter advantage: against NVIDIA's x8 interface it makes no difference on a current board, but it matters when the card goes into an older platform where halving the lanes turns into a real limitation. Budget buyers are precisely the people who reuse old boards.
Threadripper 9000 Has Specs Now, Still No Prices
The Shimada Peak family lands as three non-Pro Zen 5 parts. The 9980X carries 64 cores and 128 threads with a 3.2 GHz base and 256MB of L3. The 9970X carries 32 cores and 64 threads at a 4.0 GHz base with 128MB. The 9960X carries 24 cores and 48 threads at a 4.2 GHz base with the same 128MB. All three boost to 5.4 GHz, all three draw 350W, and all three use sTR5 with PCIe 5.0. The Pro WX line goes wider, six chips running from a 12-core 9945WX up to the 96-core 9995WX, with enterprise management features and, historically, WRX90 support alongside TRX50.
Availability is listed as July 2025 and pricing is listed as nothing at all. That is what makes this a specification dump instead of a buying decision. The architecture changes are genuine, more memory support and wider AVX-512 among them, but the generational delta on these parts is a question for benchmarks, and AMD is not supplying any.
The R9700 Is Aimed at a Different Buyer Entirely
The Radeon AI Pro R9700 is the genuinely new object in the announcement: 128 AI accelerators, 32GB of GDDR6, 1531 TOPS claimed, a 300W TDP, PCIe 5.0, RDNA 4. Against the W7700 it replaces, memory doubles, FP16 throughput climbs from 56.54 to 96 TFLOPS, and 32 accelerators get added, which puts it closer to a W7800 than to its own predecessor.
The positioning is entirely about model capacity. Thirty-two gigabytes lets you load models that will not fit in sixteen, and less aggressive quantization means better answers. The comparison slide uses an RTX 5080, which is an odd choice when the 5090 also carries 32GB and would be the like-for-like matchup. The multi-GPU pitch is the more interesting one: four cards on a PCIe 5.0 platform, presented as 128GB of pooled memory for models no single card can hold. No price, July availability.
What I Think
The 9060 XT 16GB is the only product here I would tell someone to care about. It is the card that gets bought by people building $900 systems, and the 16GB version at $350 is the one that will not be obsolete before the warranty expires. The 8GB model exists to hit a price tag, and AMD should have skipped it.
The workstation announcements are a different genre: impressive specifications, no prices, no performance data, and a July date. That is fine as a roadmap and useless as a purchase decision. When the AI card compares itself to a 16GB gaming part instead of the 32GB part one tier up, AMD is telling you which comparison flatters it. Wait for independent numbers and real prices before treating any of this as a recommendation.