AMD's Ryzen 9800X3D Beats 5800X3D by 58%, Not on Clock

AMD's Ryzen 7 9800X3D scores 58 percent higher than the Ryzen 7 5800X3D on Geekbench 6's multi-core test, and only a small share of that gap comes from clock speed. Daniel Lemire, a computer science professor at Université du Québec's TÉLUQ campus in Montreal, published the comparison on his blog on Sept. 18, running all three chips through the same benchmark and then working backward through AMD's own architecture disclosures to explain the gap.

The three chips are AMD's flagship 3D V-Cache gaming parts, released two years apart: the Ryzen 7 5800X3D (2022, built on the Zen 3 core), the Ryzen 7 7800X3D (2023, Zen 4) and the Ryzen 7 9800X3D (2024, Zen 5). All three are eight-core, 3D-stacked-cache chips meant as drop-in upgrades on the same style of socket. On Geekbench 6, single-core scores rose from 2,016 to 2,969, a 47 percent gain; multi-core scores rose from 11,832 to 18,751.

Clock speed explains 15 percent, not 58

Max boost clock rose from 4.5 GHz on the 5800X3D to 5.2 GHz on the 9800X3D, a 15 percent increase, according to Lemire's figures, which match the base and boost speeds AMD lists for both chips. Base clock climbed further, from 3.4 GHz to 4.7 GHz. If clock speed were the whole story, performance would have scaled with it. It did not: multi-core throughput nearly quadrupled the clock gain in percentage terms.

Lemire's answer is transistor count. The 5800X3D carries roughly 11 billion transistors; the 9800X3D carries roughly 16 billion, a 50 percent increase, and Lemire found most of the added budget went into the core rather than the cache. Dispatch width, the number of instructions the chip can issue in a single cycle, rose from six to eight. Integer ALUs, the execution units that do arithmetic, went from four to six. The reorder buffer, which holds instructions the chip has issued but not yet finished, grew from 256 to 448 entries. L2 cache per core doubled, from 512 KB to 1 MB, and L1 data cache grew from 32 KB to 48 KB. The chip's SIMD units, which handle vector math for compression, encoding and some AI workloads, doubled in width too, from four 256-bit units to four 512-bit units, with matching increases in per-cycle load and store bandwidth.

The widening landed in one generation, not two

Wikipedia's entry on Zen 5, AMD's architecture inside the 9800X3D, narrows the timing further: the dispatch-width jump from six to eight macro-ops, and the ALU jump from four to six, both happened specifically between Zen 4 (the 7800X3D) and Zen 5, not gradually across all three chips. The Zen 3-to-Zen 4 step, by contrast, was a smaller, more conventional refinement. Two years of headline gains compress a story that is really about one generation doing most of the structural work.

Metric	5800X3D (2022)	7800X3D (2023)	9800X3D (2024)
Geekbench 6 multi-core	11,832	15,508	18,751
Max boost clock	4.5 GHz	4.5 GHz	5.2 GHz
Dispatch width	6/cycle	6/cycle	8/cycle
Reorder buffer	256 entries	320 entries	448 entries
Rated TDP	105W	120W	120W

1. 5800X3D (2022)12K score
2. 7800X3D (2023)16K score
3. 9800X3D (2024)19K score

Source: Daniel Lemire, lemire.me, accessed 2026-09-23

What the comparison leaves out

The number that matters here is not the 58 percent score gain but the 14 percent rise in rated power draw needed to get it. AMD's own specification sheets, cited by TechReviewer, list the 5800X3D at a 105-watt TDP and both the 7800X3D and 9800X3D at 120 watts. A chip that gains 58 percent more throughput for 14 percent more rated power is getting meaningfully more efficient, not just faster, but neither Lemire's post nor AMD's marketing frames the result that way.

Geekbench 6 is also a general-purpose benchmark, not a gaming workload, and these are gaming chips first. The 3D V-Cache that defines the X3D line trades clock headroom for cache size specifically to help game engines that are sensitive to memory latency; a synthetic multi-core score can undercount exactly the workload AMD built the chip for. What would change this read is a like-for-like game-frame-rate comparison across the same three chips, run on identical memory and cooling, which neither Lemire's post nor AMD has published. AMD has not said whether its next X3D chip will repeat the pattern of concentrating structural change in a single generation rather than spreading it evenly, and until a fourth data point exists, Lemire's two-generation sample is the only rule anyone can test it against.

The pattern is not unique to gaming chips. System76's Thelio Mira AI workstation ships its base configuration on a Ryzen 7 9700X and, like AMD's X3D launches, arrived with no benchmark figures attached, only specifications. Intel's Crescent Island inference GPU did the same this month, publishing a tokens-per-watt pitch with no tokens-per-watt number. Across the industry, the spec sheet keeps getting more detailed while the benchmark that would let a buyer compare it to the previous generation keeps getting left out, which is why an independent comparison like Lemire's, run by someone with no product to sell, is doing work vendor disclosures no longer do.