Google Tensor G6 vs Tensor G5: Benchmarks Tested and Compared

Every year the Tensor chip is the part of a new Pixel I am most curious about and most nervous about. The Tensor G6 is a bigger shift than the number suggests. Google has reportedly moved to TSMC's 2nm process, jumped straight from an old Cortex-X4 prime core to Arm's new C1-Ultra, and finally dropped the Samsung modem for a MediaTek one. On paper, that reads like the generational leap Pixel fans have been asking for since the first Tensor.

So I ran the two chips through the same suite, the Tensor G6 in the Pixel 11 Pro XL against last year's Tensor G5, to see how much of that shows up in the numbers. Some of it does. Some of it went the other way, and I am not going to pretend it didn't.

Specs at a Glance

Here is how the two chips compare before we get to the benchmarks:

AnTuTu Performance: A Strange Split

AnTuTu rolls CPU, GPU, memory, and UX into one composite score, so it is a decent first look at where a chip stands. The G6 landed at 1,392,068, up from the G5's 1,227,867. So it is ahead overall, but the margin is smaller than the spec sheet had me expecting.

What this means:

• The GPU score is the surprise. Despite the older architecture, the G6 pulls ahead of the G5 here, likely helped by more compute units, a higher clock, and better efficiency this generation. It is still nowhere near a Snapdragon flagship, but for once the Pixel's graphics number moved in the right direction.
• The UX and memory scores climbed nicely, which points to a chip that handles everyday navigation more smoothly than the G5 did.
• The CPU score is the head-scratcher. AnTuTu actually put the G6 slightly behind the G5 in the CPU portion, which is not what a C1-Ultra core is supposed to do. That lines up with what I saw in Geekbench, so let's get to that.

Temperatures during the run are where the G6 started to look genuinely different. It reported a battery temperature of 29.1°C and an estimated CPU temperature of 28.6°C, with only about a 4°C rise across the full benchmark. That is a remarkably flat thermal curve for a Tensor, and an early sign the new silicon runs cooler than anything Google has shipped before (at least for now).

Geekbench 6: The Number That Gave Me Pause

Geekbench zooms in on the CPU, testing single-threaded and multi-threaded performance. This is where the G6 did not behave.

What this means:

Single-core is a win. The G6's 2411 edges past the G5's 2268, which is the C1-Ultra doing what a new prime core should. Multi-core is the softer spot at 4765 against the G5's 5930, so the older chip still pulls ahead when every core is loaded. The gap is not huge, but it is there, and on paper a jump to the C1-Ultra plus four C1-Pro performance cores should have closed it.

The likely culprit is the layout. The G6 runs seven cores to the G5's eight, and dropping a core costs you most in exactly the multi-threaded workloads Geekbench multi-core leans on. Single-core climbing while multi-core trails fits that trade cleanly; Google has gone for a stronger prime core and fewer of them. It is worth adding that this is still early retail silicon on Android 17, and Tensor scores have a habit of creeping up as the software settles, so the multi-core number may yet close some of that gap.

3DMark: Stress and Stability

3DMark leans on the GPU and, in the stress runs, shows how well a chip holds performance across twenty back-to-back loops.

What this means:

The two tests tell slightly different stories. In Wild Life Extreme, the G6's best loop of 2975 sits a touch under the G5, while in the newer Steel Nomad Light, the G6 pulls ahead at 1158 against 1004. So depending on the workload, the graphics performance is either level with last year or a little better, which matches the small GPU bump from AnTuTu.

The stability figures are the more interesting part. Wild Life Extreme held 67.1% stability, dropping from a best loop of 2975 to a lowest of 1997 as the run went on. Steel Nomad Light was steadier at 72.8%, sliding from 1158 to 843. Neither is a flat line, and the chip does shed a meaningful chunk of its peak once the loops pile up, but the newer silicon looks to be keeping the throttling more controlled than the heavy dips I have come to expect from a Tensor under sustained graphics load.

Storage and Memory Speeds

Storage speed shapes how responsive a device feels day to day, from installing apps to loading heavy games and moving files around.

What this means:

The sequential numbers are close enough to call a draw, with the G5 reading a little quicker and the write speeds sitting almost on top of each other. The real difference is in random access, where the Pixel 11 Pro XL is miles ahead at 1208 and 1166 MB/s against the Pixel 10 Pro's 680 and 528. Random read and write is what you feel when apps are launching, and several are open at once, so this is the storage result that actually matters, and it goes the G6's way by a wide margin.

CPU Throttling: Sustained Performance

Peak numbers are one thing, but how a chip holds up over a long session is what actually decides sustained performance, and this is where Tensor has stumbled before. I ran the CPU Throttle Test to see how much the G6 sheds once it heats up.

The G6 held 51% of its peak output, throttling down from a maximum of 86,799 ips to a floor of around 43,945 ips over the run. The CPU clock told the same story, sliding from 3,274MHz at full tilt to 888MHz once the chip settled into its thermal limit. The graph shows a steep drop across the first 40 seconds or so before it levels off into a stable band and holds there for the rest of the test.

For context, the G5 throttled to 47% in the same test last year. So the G6 is a little better, hanging on to a bit more of its peak, which fits the cooler-running behaviour I saw everywhere else. That said, this is still roughly half the peak performance gone once the chip warms up, and it reads alongside the 3DMark stability figures of 67.1% in Wild Life Extreme and 72.8% in Steel Nomad Light. The move to a newer node keeps the G6 cool and controlled under load, but for sustained frame rates in a long session it still gives up a large chunk of its ceiling. It is an improvement on the G5, just not the fix that closes the gap to a Snapdragon.

Verdict

Lining up all the numbers, the G6 is a more complicated chip than the spec sheet led me to expect. The wins are real. Overall AnTuTu is up, the GPU score finally moved in the right direction after years of the Pixel falling behind on graphics, the UX and memory numbers climbed, and the thermals are the standout, a flat curve and controlled throttling that no earlier Tensor could manage. The cooler-running silicon and the MediaTek modem address the two things that have dogged Tensor for years, heat and connectivity, and that alone makes the G6 the most sensible Tensor Google has shipped.

That said, the CPU results hold it back from being the clean generational leap it looked like on paper. Both Geekbench scores came in under the G5, and the multi-core gap is hard to ignore on a chip built around Arm's newest cores. I lean towards this being early-silicon and seven-core behaviour that improves with software updates, and I will retest, but I am not going to call the G6 faster than the G5 on the CPU when my numbers say otherwise right now.

So where does that leave the Tensor G6? It runs cooler, holds performance better, feeds the TPU harder, and nudges the GPU forward, which is a meaningful step for the chip that has always been the Pixel's weak point. What it still does not do is trade blows with Qualcomm and MediaTek on raw CPU power. Google has fixed a lot of what was wrong with Tensor. Raw processing is the one box that stays unticked.