64 Cores for $5,000: Threadripper 9980X Is a Workstation Chip With a Gaming Hangover
The Threadripper 9980X costs $5,000, carries a 350W TDP, and exists to run simulations, compilers, and financial models that pay for their own hardware. Judged that way, Zen 5 is a real but deeply uneven upgrade: small gains where memory bandwidth rules, enormous ones where AVX-512 finally gets a full-width datapath. The gaming results are a footnote that mostly says don't.
The Lineup, Priced Like Workstation Hardware
Sixty-four cores, 128 threads, 256MB of L3, a 5.4 GHz peak boost, the sTR5 socket, and 350W: this is the high-end desktop flagship of the Threadripper 9000 family, launched at the end of July 2025. Its siblings are the same silicon with cores fused off. The 9970X runs 32 cores and 64 threads for $2,500, the 9960X runs 24 cores and 48 threads for $1,500, and both carry 128MB of L3 with the same boost ceiling. Prices land almost exactly where the 7000 series did, which is AMD's way of saying the segment has no competitive pressure worth responding to.
That number is doing a lot of work. Nobody buys this for value. They buy it because 64 cores with a deep PCIe lane count behind them is the cheapest way to feed multiple accelerators or a pile of NVMe. The Pro WX parts above it add management, security, and platform-stability commitments for enterprises. The non-Pro chips are for labs, studios, and anyone whose time gets billed by the hour.
Zen 5's AVX Bet Pays Out in Narrow Bursts
The architectural change that matters is the floating-point datapath. Zen 4 executed a 512-bit AVX-512 operation as two 256-bit halves; Zen 5 does it in a single pass, with L2-to-core bandwidth raised to keep pace. AMD's own framing is careful about this, calling it a theoretical doubling that code has to be shaped to exploit, and that caution turns out to be correct.
Published runs show exactly that profile. Most tests land in single-digit generational gains, with a handful of AVX-heavy workloads jumping far past everything else. The standout is a SpecWorkstation convolution result that sits wildly outside the rest of the distribution, and it survived third-party review and repeats across several generations of chips. Whether that figure means anything for real signal-processing work is a genuinely open question. A benchmark can be perfectly repeatable and still be a microbenchmark. Read the big percentage as a property of the test first and the processor second.
Gaming Is Playable; Consistent Is Another Matter
Gaming on Threadripper was never about speed. It is about confirming nothing is broken. By that standard the 9980X passes: the games run, framerates are unremarkable rather than catastrophic, and the chip is nowhere near a modern gaming part in either averages or low percentiles.
The problem is variance. High core counts run into scheduler behavior, clocks fall when 128 threads compete for a 350W budget, and run-to-run spread widens well past what desktop parts show. One title produced a spread across repeated passes wide enough that any single average should be read with suspicion, and the lows were worse than the averages implied. That is not a defect AMD patches out; it is the cost of the configuration. If the machine renders overnight and plays after, fine. If it was bought to play, every dollar is in the wrong place.
Power, Heat, and Why 64 Cores Are Not Automatically Better
GamersNexus measured around 371W at the EPS12V cables under an all-core Blender load, up from roughly 354W on the previous flagship: stable, consistent, and a serious cooling obligation. The saving grace is die area. The same heat spread across that much silicon produces lower temperatures per hot spot, with CCD deltas in the single digits and package temperatures in the high fifties on a 360mm cooler at full fan speed. Big surface area is a thermal feature, not a footnote.
All-core clocks settle near 3950 MHz in rendering with brief spikes to 4200 MHz as tiles complete, while single-thread work reaches about 5425 MHz. That gap explains everything about the results. Workloads that scale with threads love this chip. Workloads that live on one fast core do not. Adobe's photo and video tools are the clearest case: too many threads for the software to schedule, so the 9980X lands behind cheaper desktop parts, and in some editing work a 32-core part is the better purchase.
What I Think
The 9980X is a good chip in a market that barely exists, and AMD knows it, which is why the price did not move. If your work is simulation, compilation, or batch rendering that can saturate 128 threads, the Zen 5 bump is worth having. It is still a modest upgrade in most places rather than a generational rewrite, and anyone already on a 7980X should sit this one out.
What bothers me is the shape of the marketing. The one enormous outlier will get quoted for years while the compression results that barely moved get forgotten, and that is how workstation buyers end up paying for a core count their software cannot use. Buy the benchmark that matches your job, not the one on the slide. And if gaming is the actual goal, a $300 desktop chip will beat this thing, which is exactly as it should be.