In an embargoed presentation Friday morning, Intel's Chief Performance Strategist Ryan Shrout walked a group of tech journalists through a presentation aimed at taking AMD's Zen 2 (Ryzen 4000 series) laptop CPUs down a peg.
Intel's newest laptop CPU design, Tiger Lake, is a genuinely compelling release—but it comes on the heels of some crushing upsets in that space, leaving Intel looking for an angle to prevent hemorrhaging market share to its rival. Early Tiger Lake systems performed incredibly well—but they were configured for a 28W cTDP, instead of the far more common 15W TDP seen in production laptop systems—and reviewers were barred from testing battery life.This left reviewers like yours truly comparing Intel's i7-1185G7 at 28W cTDP to AMD Ryzen 7 systems at half the power consumption—and although Tiger Lake did come out generally on top, the power discrepancy kept it from being a conclusive or crushing blow to AMD's increasing market share with the OEM vendors who are actually buying laptop CPUs in the first place.
Enter the battery
-
Intel badly needs to increase perception of Tiger Lake as superior to Ryzen 4000—and on-battery performance is a plausible angle to take.
-
Many real-world workloads—what Intel calls RUGs—demonstrate deltas between AMD plugged-in and on-battery performance similar to the one seen in PCMark 10 Modern Apps.
-
An aggrieved Intel points out that the most widely-used benchmarks, like Cinebench R20, run long enough to mask the initial performance delay from Team Red when on-battery.
-
Intel explains the reason for AMD's performance drops on-battery—and when they are and are not seen—by graphing Ryzen 4000's CPU power and voltage states over time.
Intel's original Tiger Lake launch presentations sought to draw attention to on-battery versus off-battery discrepancies in AMD's performance, but those attempts mostly went unheard. Shrout's presentation Friday was an attempt to tell that story again, this time with enough additional information to get people fired up.
We can see this discrepancy between on-battery and off-battery performance easily in the PCMark 10 Applications benchmark and also in many of Intel's RUGs—scripted workloads based around production applications, which the company calls "Realistic Usage Guides." However, the same discrepancy between on- and off-battery performance isn't visible in more commonly used industry benchmarks, such as Cinebench, PassMark, or Geekbench.
Intel's engineering team displays the reason why we don't see the discrepancy in Cinebench in the last image of the gallery above—in Intel's testing, the Ryzen 4000 CPUs didn't ramp up power and voltage to their maximum state until somewhere between eight and 11 seconds after heavy-duty workloads began.
Independent confirmation
-
The x-axis is time elapsed, and the y-axis is compression throughput in MiB/sec. Ryzen 4700u's throughput leaps forward 70 percent between the 11s and 12s marks; Intel's is constant throughout.Jim Salter
-
The phenomenon we saw with performance delay in single-threaded workloads is repeated on this quad-thread workload—a 65 percent performance increase, ramping up between the 11s and 12s marks.Jim Salter
-
The performance leap isn't quite as pronounced in an all-threaded workout—but it's still sharp, with a 48 percent jump between the 11s and 12s marks.Jim Salter
We were able to confirm Intel's findings over the weekend, working with an Acer Swift 3 SF314-42 laptop (with a Ryzen 7 4700u CPU) and an MSI Prestige 14 Evo laptop (with a Core i7-1185G7). In the charts above, we repeatedly compress small chunks of the Linux 5.3 kernel source and graph throughput over time on each CPU.
The 4-core/8-thread i7-1185G7 easily outperforms the 8-core/8-thread Ryzen 7 4700u in both single and quad-thread workloads, even after the Ryzen 7 4700u belatedly jumps to its full performance around the 12-second mark. In the unlimited workload, where the Ryzen 7 is allowed to flex its full octa-core muscle, things are much closer—and the 4700u even ekes out a narrow win in the last four seconds.
There are a few things we need to point out here, though. First and most obviously—Intel is 100 percent correct in its claims that AMD's Zen 2 laptop CPUs delay ramping power and voltage up to their maximum states. This causes a sharp, corresponding, and decreased performance during those first few seconds.
We reached out to AMD representatives for comment on this design decision. Although AMD representatives asked further questions about our observations, we have not yet received a response for the record at press time.
The devil is in the details—so is the heat
-
This MSI Prestige 14 Evo—and its Core i7-1185G7 CPU—is a barn-burner, sure. But the "burn" is a little more literal than we'd prefer, with CPU temps screaming up to 94°C, and CPU power consumption of 34W sustained during Cinebench R23 testing.Jim Salter
-
The Ryzen 7 Pro 4750u in this HP EliteBook laptop stays relatively cool and quiet during its Cinebench R23 run, with max power consumption a whopping 31W lower than the i7-1185G7's.Jim Salter
But Intel is still playing games with its own power consumption. In the above screenshot, we can see the MSI Prestige Evo 14 with Core i7-1185G7 during a Cinebench R23 run. We haven't had this laptop for long enough to fully review it—and particularly, to review its battery life, which we've been very curious about since being forbidden to test that stat in two earlier i7-1185G7 systems.
But we can see that—rather than dial the i7-1185G7's cTDP down to something approximating the typical Ryzen 7 4000 cTDP of 15W, as widely expected—MSI has in this laptop chosen to dial it up even further than what we saw in earlier prototypes. This production i7-1185G7 system has a variable PL1 which hits as high as 36W during the course of a Cinebench R23 run—in addition to its PL2 of 51W, which is unchanged from the prototypes.
During this Cinebench R23 run, the laptop spent its first 10 to 15 seconds running at the full PL2 power limit of 51W, with temperatures up to a blistering 98°C. After that initial, extremely high-performance, power, and heat generating burst, the CPU dropped down to sustain an average power consumption of 34W. By contrast, an 8 core / 16 thread Ryzen 7 Pro 4750U—at cTDP up of 25W—consumed an average of 27.9W, with a high of 29.9W.
While we're veering away from the CPUs themselves and into laptop-design territory, it's perhaps worth noting that system-fan activity was also significantly different between the MSI Prestige 14 Evo—which reached nearly gaming-laptop levels of fan noise almost immediately—and the HP EliteBook, which took more than a minute to ramp its fans up to max and remained much quieter than the MSI throughout the run.
The battle continues
While Intel didn't specifically tell us what conclusions we should draw from the performance delay in Zen 2 laptop CPUs versus the instant-on performance from Tiger Lake, the company was clearly hoping for something in between "AMD is gaming the benchmarks" and "it turns out, Intel was the winner all along."
We don't think there are any such cut-and-dried conclusions to draw here. Intel's findings regarding the slow performance ramp of the AMD Zen 2 laptop CPUs is, obviously, correct in the facts—we had no trouble confirming it, and it does explain why many of Intel's preferred benchmarking techniques show larger performance deltas in favor of Team Blue than the more widely used industry benchmarks like Cinebench, PassMark, and so forth.
But this ignores the greater efficiency of the AMD systems, above and beyond the delayed shift to maximum performance (and battery consumption) states in the CPU. When we run Cinebench R23 for five full minutes, a Ryzen 7 Pro 4750u system renders more scenes than the Intel i7-1185G7, and it does so with less total power consumed. There's no clever trick to explain that away.
We also believe there's a tuning argument to be made on both sides. Intel's more rapid shift to the highest performance state carries some real-world benefits with it, but we're not certain they're as compelling as the charts make them seem. In practical terms, we've spent quite some time now with both Zen 2 and Tiger Lake laptops—and the Tiger Lake systems don't really feel faster in terms of a seat-of-the-pants subjective experience. This argues strongly that there frequently isn't much point in ramping up CPU power profiles that quickly—if the human piloting the system doesn't notice the latency improvement, it's probably better to conserve the battery instead.
... and continues
Some exceptions exist—the most notable likely being boot time. Tiger Lake systems boot—and resume from suspend—phenomenally quickly, and we suspect their willingness to instantly boost performance to maximum has a lot to do with it. One developer we spoke to speculated that JavaScript just-in-time (JIT) compilation might be another short workload that nevertheless was easily human-perceptible.
The best news for consumers, we suspect, is that the "which system is better" argument is so difficult to conclusively answer in the first place. This level of competition means neither team gets to rest on its laurels, and consumers are less likely to end up buying systems nobody would want, if fully informed about the differences.
The Link LonkNovember 24, 2020 at 04:07AM
https://ift.tt/3m0vpky
AMD laptops have a hidden 10-second performance delay. Here’s why - Ars Technica
https://ift.tt/2ZDueh5
AMD
No comments:
Post a Comment