Engineering samples of "Medusa," AMD's Zen 6-based notebook APU (a processor that integrates the CPU and GPU on one chip) due in 2027, appeared in succession in the Geekbench 7 database on November 4–5, 2026. All were Lenovo prototype machines, and three configurations have been spotted: 10-core, 8-core, and 6-core. On November 7, Wccftech reported that the 10-core model scored 3100 single-core and 18313 multi-core, presenting it as a major step up from the current Ryzen AI series. But a closer look at the 11 results for the same 10-core model shows that 3100 is merely the highest score, and six of the results fall below 2000 single-core. In addition, the iGPU's "4CU" listing may be tied to Geekbench's particular way of counting, so judging these numbers as Zen 6's true capability calls for careful verification.
Results for 10-, 8-, and 6-core chips found in four Lenovo prototypes
Searching Geekbench 7 CPU results for "AuthenticAMD Family 26 Model 129" turns up 11 results for a 10-core, 20-thread chip as of November 8, 2026. All were uploaded on November 4–5, under two model names: "LENOVO 22ETPF6DTA" and "LENOVO 22EUFVT023." Their ES part numbers also differ: "100-000001713-32_Y" and "100-000002192-30_Y," respectively. The top score among them is the result with 3100 single-core and 18313 multi-core (v7/cpu/804566), which is the figure Wccftech picked up.
For the 8-core and 6-core models, a separate identifier, "Family 26 Model 224," shows 13 results in the same period. The 8-core model accounts for eight of them, under "LENOVO 22EUFVT027" (100-000002152-50_Y), and the 6-core model accounts for five, under "LENOVO 22EQFVT006" (100-000001658-50_Y). The 8-core figures Wccftech reported, 2913 single-core and 15075 multi-core, are based on the result with the highest single-core score (804572). However, the highest multi-core score was 15151, recorded in a different result (804742), so both peaks did not come from the same run. The 6-core model's 2651 single-core and 10670 multi-core (804574) likewise come from the result with the highest single-core score.
| CPU configuration | Model name (all Lenovo) | Results counted in the original article | Run with highest single-core score | Multi-core in the same run |
|---|---|---|---|---|
| 10 cores / 20 threads | 22ETPF6DTA, 22EUFVT023 | 11 | 3100 | 18313 |
| 8 cores | 22EUFVT027 | 8 | 2913 | 15075 |
| 6 cores / 12 threads | 22EQFVT006 | 5 | 2651 | 10670 |
There is still no definitive information on the core configuration of each model. Wccftech describes the 8-core model as "2 Zen 6 + 6 Zen 6c" and the 6-core model as "1 Zen 6 + 5 Zen 6c." Meanwhile, AnandTech's forum has speculated about other layouts including low-power cores, such as "2+4+2" and "1+3+2," none of which AMD has confirmed. All that can be read from the Geekbench results at present is that the 10-core model is shown as consisting of two clusters, of 4 and 6 cores.
The GPU result page for the 10-core model (v7/gpu/361996) also retains some specification information. This result was posted from "LENOVO 22ETPF6DTA" on November 4 and is a separate run from the CPU result that scored 3100 single-core. It was run on Geekbench 7.1.0 and Windows 11 Pro, with 16MB of L3 cache, 1MB of L2 cache per core, and a base frequency of 2.00GHz. Memory is listed as four-channel DDR5 at 5188MT/s, with 55.58GB recognized. However, the maximum frequency field reads 1996MHz, so this result alone cannot tell us the actual boost clock, and the L3 cache capacity should be treated as information shown on the result page.
| Item | Geekbench listing |
|---|---|
| Benchmark | Geekbench 7.1.0 for Windows AVX2 |
| OS / power plan | Windows 11 Pro / Balanced |
| ES part number | 100-000001713-32_Y |
| Clusters | 4 cores + 6 cores |
| L2 / L3 cache | 1MB×10 / 16MB×1 |
| Base / max frequency | 2.00GHz / 1996MHz |
| Memory | 55.58GB, DDR5, 5188MT/s, Channels: 4 |
At its Financial Analyst Day on November 11, 2025, AMD laid out a roadmap with "Gorgon," a refresh of Zen 5, arriving in 2026 and the Zen 6-based "Medusa" in 2027. Wccftech explained that these results are from OEMs evaluating AMD processors as they develop 2027 models. In other words, the scores seen here are measurements from a prototype stage roughly a year before the product's launch.
The 10-core model's 11 results split into a 3000-range group and a 1500–1900 group
Comparing the 11 Geekbench 7 scores for the 10-core ES chip shows they fall into two groups with very different single-core performance. Five results reached the 3000 range, with scores of 3030, 3033, 3059, 3080, and 3100 and multi-core scores of 16969–18313. The other six single-core scores were 1560, 1595, 1870, 1900, 1920, and 1931, with multi-core scores of just 11575–13416. No results exist between 1931 and 3030, so the performance gap between the two groups is clear.
Notably, the 3000-range scores were not recorded on only one particular prototype. "22ETPF6DTA" recorded 3030 and 3100, and "22EUFVT023" recorded 3033, 3059, and 3080, so similar performance has been seen on both models with different ES part numbers. The low-score group also includes both models: "22ETPF6DTA" recorded 1560, 1900, and 1920, and "22EUFVT023" recorded 1595, 1870, and 1931. The 3100 is therefore not a fluke peak from a single prototype but the top value of a high-performing group repeatedly confirmed on two different models. Even so, because the same models also recorded far lower scores, the variability in the results cannot be ignored.
This distribution can be checked in the public Geekbench Browser database. In Geekbench Browser, user-submitted benchmark results are published as individual pages, and searching by CPU identifier lists the matching model names, upload dates, and scores. The numbers that spread as leaks are just examples picked from among such results. Comparing results under the same identifier makes it possible to verify where a reported score sits in the overall picture and whether different models achieve similar performance.
The 8-core and 6-core models also show large gaps between results. Of the eight results for the 8-core model, five had single-core scores of 2875–2913, while the other three managed only 1415–1833. On the 6-core model, two of five results recorded 2607 and 2651, and the 2651 Wccftech cited is the highest of them. In short, the 10-core 3100, 8-core 2913, and 6-core 2651 that Wccftech picked up were each the highest single-core score confirmed for that configuration.
The cause of the low-scoring results cannot be determined from the publicly available measurement information. All are in-development ES chips and do not directly indicate the performance of production models. Even so, the fact that single-core scores in the 3000 range have been repeatedly confirmed across different models and ES part numbers deserves attention when estimating Zen 6's single-thread performance. It is not appropriate, however, to use the peak 3100 as a figure representing all 11 results.
Wccftech's performance comparison mixes measurement conditions and Geekbench versions
According to Wccftech's comparison, the 10-core Medusa's single-core score of 3100 beats the Ryzen AI MAX+ 395's 2422 by about 28% and the Ryzen AI 9 HX 370's 2354 by about 32%. The 8-core model's 2913 is about 23.7% above the HX 370, and the 6-core model's 2651 is about 25.1% above the Ryzen AI 5 340's 2119. In multi-core, the 10-core model's 18313 is about 21.5% above the HX 370's 15064, and the 8-core model's 15075 is said to be roughly on par with the HX 370.
However, this comparison does not use unified measurement conditions. Wccftech explains that it used comparison values listed in the Geekbench database for the comparison chips. The Medusa side, meanwhile, uses peak values selected from ES measurement results, so database comparison values are being compared directly against the prototype's best result. As noted above, only five of the 10-core model's 11 results scored in the 3000 range in single-core. Hassan Mujtaba, senior editor at Wccftech, suggested that if the ES stage already delivers this level of performance, production models could be faster still, but he gave no concrete basis for that prediction.
Another point requiring caution is the difference in Geekbench versions. Developer Primate Labs released Geekbench 7 on July 23, 2026. Geekbench 6 result pages state explicitly that scores can only be compared between Geekbench 6 results. Because Geekbench 6 and Geekbench 7 differ in measurement methods and scoring baselines, their numbers cannot simply be compared to derive performance differences.
Wccftech's same article also includes a separate comparison table titled "Geekbench 6," listing "AMD Medusa (10-Core ES)" at 3407 single-core and 16826 multi-core. Looking this result up in the Geekbench 6 database shows an upload date of August 27 and a machine name that is not a Lenovo notebook but AMD's evaluation board "AMD Plum-MDS1." Its ES part number is "100-000001713-33_N," which also differs from the November Lenovo prototypes. In addition, the comparison chips' scores are values Wccftech measured itself with modified power settings. The 3407 figure is accurate as an August 27 result, but the comparison table does not state the measurement date or machine used, so it must be read separately from the November Lenovo prototypes.
The main measurement results confirmed so far for the Medusa 10-core model are summarized chronologically below.
| Date (2026) | Result | Machine | Version |
|---|---|---|---|
| July 17 | 3329 single-core, 16555 multi-core | AMD Plum-MDS1 | Geekbench 6 |
| July 23 | Primate Labs releases Geekbench 7 | — | — |
| August 27 | 3407 single-core, 16826 multi-core | AMD Plum-MDS1 | Geekbench 6 |
| November 4–5 | 11 results confirmed; best is 3100 single-core, 18313 multi-core | LENOVO 22ETPF6DTA, 22EUFVT023 | Geekbench 7 |
| November 4–5 | 2997–3060 single-core, 10266–15101 multi-core | SU AEN55 and others | Geekbench 6 |
| November 7 | Wccftech publishes two comparison tables | — | — |
The Medusa 10-core model had already appeared in Geekbench 6 as of July, and the August 27 score of 3407 was likewise recorded by an AMD evaluation board during development. The November Lenovo prototypes, by contrast, were measured in Geekbench 7, so both the machines and the benchmark version differ. Simply lining these up does not show whether Zen 6's performance has improved or declined over time.
Is the iGPU really 4CU? Geekbench lists the Radeon 890M as "8"
In the 10-core model's GPU result (v7/gpu/361996), the integrated GPU's device name is shown as "gfx1170" and its Compute Units (CUs, the GPU's processing units) count as "4." The maximum frequency is 3200MHz, and the OpenCL score was 22249. AMD's ray-tracing library "HIP-RT 3.2.0" has added gfx1170, gfx1171, and gfx115E for Medusa. Of these, GFX117X is classified as RDNA 4m and GFX115X as RDNA 3.5.
Wccftech adopted Geekbench's "4CU" listing as is and presented the RDNA 4m 4CU GPU as having beaten the 8CU Radeon 860M (20642). Its comparison table also includes the 16CU Radeon 890M (31475), the 12CU Radeon 880M (26814), and the 4CU Radeon 840M (12019). If Medusa's GPU really has 4CUs and the measurement conditions match those of the comparison chips, it would have beaten the Radeon 860M with half the CU count, which could be read as a large improvement in per-CU performance.
However, there are confirmed cases in which Geekbench's Compute Units display does not match AMD's stated CU count. In a Geekbench 6.4.0 OpenCL result, the Radeon 890M (gfx1150), which has 16 CUs by specification, was listed with Compute Units "8." That run had an OpenCL score of 30946 and a maximum frequency of 2900MHz. Because there is an actual example of Geekbench showing half of AMD's stated CU count, the "4" for gfx1170 could also represent 8CU-equivalent.
This difference is thought to be related to the internal structure of RDNA-series GPUs. According to AMD's technical documentation, the RDNA architecture groups two CUs into one WGP (Workgroup Processor). It is possible that the Compute Units reported by Geekbench's OpenCL correspond to this WGP count. However, a single Radeon 890M measurement cannot establish Geekbench's display specification as a whole. Note that the number of compute units reported in OpenCL device information does not necessarily match the CU count a manufacturer publishes as a product specification.
Given this display difference, two interpretations remain for the gfx1170 GPU configuration. If it is 4CU as displayed, then, as Wccftech pointed out, it posted a high score with half the CU count of the Radeon 860M. If the same counting method as the Radeon 890M applies, it would actually be 8CU-equivalent, making the comparison with the Radeon 860M one at the same CU count. AMD has not announced gfx1170's official CU count, and it is unknown whether the same display method is used for RDNA 4m. Note also that the Radeon 890M measurement was run on Geekbench 6.4.0 and the gfx1170 measurement on Geekbench 7.1.0, so this inference assumes the CU display method stays the same across versions.
The GPU score comparison itself also has a benchmark-version problem. The Radeon 890M score of 31475 that Wccftech used does not state its measurement version and is close to the 30946 result confirmed in Geekbench 6. It cannot be ruled out that gfx1170's 22249, measured in Geekbench 7, is being mixed with Geekbench 6 scores. For that reason, an exact performance improvement over the Radeon 860M or 840M cannot be calculated from 22249, and it is difficult to draw a conclusion about per-CU performance at this stage.
What to confirm before Medusa arrives in 2027: performance and official GPU specs
AMD's product roadmap calls for Gorgon, a Zen 5 refresh, in 2026 and the Zen 6-based Medusa in 2027. Gorgon is used as "Gorgon Point" in the Ryzen AI 400 series, and the 12-core, 24-thread Ryzen AI 9 HX 470 was announced at CES 2026. Medusa is positioned as the next generation after that, and these Geekbench results suggest OEMs are advancing product development for 2027. AMD has not, however, announced official product specifications or performance targets.
If you are considering a notebook with the Ryzen AI 300 series' HX 370 or the Ryzen AI 400 series' HX 470 right now, do these results give you reason to wait for Medusa? On CPU single-core performance, scores in the 3000 range have been repeatedly confirmed on two different models, which is a result that raises expectations for Zen 6's gains. But the 28% and 32% gaps over current products were calculated from the ES chip's peak value and comparison values obtained under different conditions. These percentages therefore cannot be treated as performance improvements for production models and used as-is in a purchasing decision. For the iGPU as well, until it is known whether the actual CU count is 4 or 8-equivalent, it is impossible to assess how much it improves over the current Radeon 860M or 890M.
The first key point for future verification is whether AMD discloses gfx1170's official CU count. That would show whether Wccftech's comparison with the Radeon 860M reflects half the CU count or the same CU count. Another important data point will be Geekbench 7 results from production models arriving in 2027. If enough Geekbench 7 data is also collected for the Ryzen AI 300 and 400 series, performance can be compared while avoiding version-difference problems. Whether the low single-core scores in the 1500–1900 range seen on these ES chips disappear in production models will also be worth watching.
Accurately judging Medusa's performance requires waiting for AMD to publish official specifications and for measurements from production machines released by OEMs such as Lenovo. The single-core scores in the 3000 range, repeatedly confirmed on two different models, are a strong clue to Zen 6's potential. On the other hand, performance gains over current products and the iGPU's per-CU performance should not be given a definitive assessment until measurement conditions and specifications are clear.