Ever since April 2017, AMD has been steadily slicing into Intel’s desktop performance position, both in core counts and per-core performance. After an excellent parry in the form of Coffee Lake, Intel’s overall positioning weakened. The repeated delays afflicting the 10nm process node have kept the company iterating on the same architecture, a situation that won’t change until Rocket Lake arrives late in Q1 2021.
The story thus far (in picture form, anyway).
Up until today, Intel could still argue that it held a leadership position in areas like single-threaded performance and a number of games. Today, that support line has been cut. Zen 3 is now the fastest high-end desktop micro-architecture on the market in every category we test including (broadly) gaming. Intel’s HEDT line can still sometimes compete effectively against AMD CPUs thanks to AVX-512 optimization, but Intel’s consumer hardware still lacks this feature, which arrives with Rocket Lake.
I’m going to cut right to the chase: AMD claimed it would improve IPC by 1.19x. It has. AMD claimed that its new CCX architecture would significantly improve performance in gaming and that we’d see more uplift in titles than would typically be expected in a new CPU architecture, and that’s exactly what we see. While the improvements obviously vary from title to title, the gains across the board are real.
I’ll have more to say about Zen 3 in an upcoming article, but the news-in-brief on the CPU core is this: AMD rearchitected the entire chip, to the point that Zen 3 is effectively a new iteration of the core. Mark Papermaster and other AMD engineers we spoke to were clear about the degree of work that had gone into every aspect of the new chip.
These two slides offer a concise summary of what AMD changed when it went rummaging around in Zen 2 in search of optimizations. Branch predictor bandwidth is higher and AMD claims its zero-bubble technology reduces latency in the event of a mispredict. The execution units now have larger windows, reduced latency on some operations (this is probably not a major factor), and a six-wide dispatch unit, even though the FPU is still limited to retiring four instructions per cycle. According to AMD, the ability to burst up to six instructions produced better overall utilization of available resources. I’ll have more to say about the low-level changes in the not-too-distant future.
The biggest mid-level change to Zen 3 is the shift to an 8-core CCX. AMD also now uses a unified 32MB L3 cache rather than the 2x16MB cache structure it previously employed. The effect of these changes is quite noticeable in certain benchmarks.
Performance Comparisons
We’ve included data on the Core i9-10900K, the Ryzen 9 3900XT, Ryzen 9 5900X, and Ryzen 9 5950X. We have only partial results for the Ryzen 9 3950X — with two reviews this morning, I had to make tough choices about which chips I had time to test — but I included it where I could.
The 3900XT-5900X comparison is arguably more useful in any case. Both the 3950X and 5950X are constrained by the ~140W maximum power draw of the AM4 socket, where the 12-core has a bit more room to breathe. For the delta between the XT and the X part, just knock a few percent off the Zen 2 core.
Intel’s Core i9-10900K is the company’s current top-performing consumer desktop CPU, so we’ve included it as well. Intel will next update its desktop CPU family with Cypress Cove, the backported 10nm CPU design implemented on 14nm, arriving at the end of Q1 2021.
All systems tested using Windows 10 2004 patched up to the latest version. GeForce driver 457.09 was used on all three machines, along with a Corsair MP600 and an Nvidia RTX 3080. The Core i9-10900K was tested in an Asus ROG Maximum XII Hero (Wi-Fi), while all of the AMD CPUs were tested in an MSI X570 Godlike motherboard with 32GB of Crucial Ballistix RAM across four DIMMs, and clocked at DDR4-3600.
All of the CPUs we tested were capable of sustaining a DDR4-3600 clock without instability or other problems. As a reminder, AMD is pricing these CPUs somewhat higher than their predecessors, at $550 and $800 respectively. Launch prices on the Ryzen 9 3900X and Ryzen 9 3950X were $500 and $750, so we’ll be watching to see if the chips prove their price/performance ratios.
Performance Results
7Zip
We’ll kick off with 7zip, which used to be a benchmark that AMD and Intel split, with Intel leading in compression and AMD in decompression. In what will become a theme, that’s no longer the case, and the 16-core Ryzen 9 5950X offers nearly double the performance of the Core i9-10900K.
Cinebench R20
In years past, we would see this benchmark divvied up between Intel and AMD, with Intel reliably claiming the single-threaded peak and AMD hoarding multi-threading glory. This arrangement is nearly preserved if you only look at the Core i9-10900K and Ryzen 9 3900XT, but then the 5900X and 5950X arrive and ruin things for everybody.
AMD claims a 1.17x improvement in single-threaded and a 1.16x improvement in multi-threaded at the 12-core point. The gains for the 16-core CPU are much lower in multi-threaded, possibly due to power consumption limitations. The 5950X still pulls off a 1.06x improvement over the Ryzen 3950X in MT and 1.2x in ST.
Indigo Bench
In Indigo bench, we actually see a performance regression between the Ryzen 9 3900XT and Ryzen 9 5900X — I need to re-check this result — but the Ryzen 9 5950X still outperforms all other solutions, beating Intel by nearly 2x in the “Bedroom” scene. You would typically run these workloads on GPUs, but running the render on a CPU gives Ryzen a chance to flex its muscle.
Corona Render 1.3
Corona Render is an Intel-friendly benchmark, as evidenced by the Core i9-10900K’s very strong showing against the Core i9-3900XT. The changes to AMD’s core really help here — the 5900X is now as fast as the Ryzen 9 3950X, while the Ryzen 9 5950X takes just 82 percent as long as the Ryzen 9 3950X to complete the task.
Blender Render Benchmark 2.0.5
We tested Blender v2.9.0 using Blender Render benchmark 2.0.5, in all of the scenes distributed with the application.
The 5900X has an interesting pattern of differences from the Ryzen 9 3900XT — the gap between them isn’t static, but changes with the scene. The 5900X wins every comparison, however. The Core i9-10900K is outperformed both per-core and overall.
NeatBench 5
Neat Video is a video noise remover I’ve been trying to teach myself as part of my video upscaling work / DS9 project. The company also produces NeatBench, which tests the maximum throughput of CPUs and GPUs.
Again, AMD picks up 1.18x generationally from the 3900XT to the 5900X. Since the 3900XT is a bit faster than the 3900X, we can see AMD is landing in the ~1.22x range it told us to expect.
Scaling from the 5900X to the 5950X isn’t very good here. Either the application tops out around 12 threads, or the 5900X is soaking up the power budget with just 24 threads. Overall performance is well above Intel.
Qt Compile
Our Qt compile test is performed in Microsoft Visual Studio 2019. We see significant performance uplift from the 3900XT to the 5900X, with the workload rendering in 83.5 percent the time. Scaling from the 12-core to the 16-core is nowhere near linear, but AMD ekes out an additional 10 percent.
Game Benchmarks
AMD has boldly predicted its own ability to beat Intel in various games and resolutions, and we’re nothing if not obliging.
Ashes of the Singularity: Escalation
Ashes kicks things off with wins for AMD at 1080p, 1440p, and 4K. While a 5 percent gain from the Ryzen 9 3900XT to the Ryzen 9 5900X isn’t all that large, the fact that we’re seeing a boost at that resolution and Crazy detail levels speaks to some of the benefits AMD is getting from its new core architecture.
Assassin’s Creed: Origins
ACO is a narrow win for the Core i9-10900K, and while the gap is generally within the margin of error, it’s a gap that consistently favors the Intel CPU. AMD makes a few very small improvements here, but not all that much.
Borderlands 3
The 5950X has to do stand-in duty for game performance between the Zen 2 and Zen 3 architectures, but Borderlands 3 appears to be a game that doesn’t benefit from AMD’s new architecture. AMD wins 1080p here, while Intel narrowly takes 4K.
Deus Ex: Mankind Divded
Our Deus Ex testing method uses MSAA, which might as well be a boat anchor as far as its impact on GPU performance. Even the mighty RTX 3080 can’t shake loose any meaningful differences at any resolution.
Far Cry 5
But where Deus Ex failed, Far Cry 5 comes through with shining colors. Performance at 1080p Ultra with HD textures enabled is up a whopping 1.27x. Even 1440p gains 1.19x, and we don’t see things settle down until we hit 4K and become GPU-limited. It’s difficult to predict which games will and won’t benefit from Zen 3’s improvements, but the uplift can be substantial when it happens.
Final Fantasy XV
In Final Fantasy XV, the Ryzen 9 5900X wins 1080p and ties up with the Core i9-10900K at 1440p and 4K, while the Ryzen 9 5950X is a whisper behind. Performance is up about 5 percent over the Ryzen 9 3900XT and a bit more than that over the 3900X — again, not a bad showing in gaming, where CPU-related improvements are hard to come by.
Hitman II
We have two Hitman results to show you, one from each benchmark map.
Again, a nice 1.11x uplift for the Ryzen 9 5900X and 5950X in Miami, though the Mumbai map is much tighter. Intel and AMD each take home one map here. The Ryzen 5000 family doesn’t dispatch Intel in every single title or resolution, but it blows enough holes in Intel’s hegemony to leave the company’s claim to gaming dominance in tatters.
Shadow of the Tomb Raider
Another massive uplift in SotTR actually puts AMD above Intel by 10fps. Again, the Ryzen 9 5900X maintains a 9 percent uplift over the 3900XT at 1440p, with the game becoming GPU limited at 4K.
Strange Brigade
Strange Brigade is the test we tapped to illustrate performance in the Vulkan API. The result is a dead heat between all four chips.
Warhammer II: Dawn of War
Warhammer II has always been an Intel win, as you can see when comparing the Ryzen 9 3900XT against the Intel Core i9-10900K in the built-in Skaven benchmark. Here, the uplift from Zen 2 to Zen 3 is no less than 1.34x, and again, AMD now wins the benchmark. Warhammer II truly doesn’t like AMD’s Zen, Zen+, or Zen 2 micro-architectures, because the game still picks up an additional 1.08x at 4K. AMD leads Intel in the Skaven benchmark at both 1080p and 1440p before Intel ties things up at 4K.
Add up the game results, and AMD wins six benchmarks, loses three, and ties two. That’s enough to award the company overall leadership of the segment, particularly considering the size of some of AMD’s leaps. The gains often seem to come in games where Ryzen was disadvantaged against Intel to start with, but there are also several instances of AMD outperforming Intel in absolute terms.
Conclusion: AMD Currently Offers Faster Chips and Better Values
The last facet to this conversation, of course, is price. The enthusiast market wasn’t exactly thrilled when AMD announced it would be raising prices. Do the new Ryzen CPUs justify it?
In a word, yes, at least as far as the 5900X is concerned. In Cinebench R20, the 5900X was 1.16x faster than the 3900X and costs 1.1x more. In Corona Render 1.3, using the 5900X cut rendering time by ~1.3x for an additional 10 percent on the sticker price. Gaming is harder to predict, since it varies by title, but if you are a gamer, you’ve got a shot at some very nice performance uplifts provided you’re already using a fairly high-end GPU to start with. This is not to say that we wouldn’t still see uplift on a GTX 1080 or 1080 Ti, but I haven’t evaluated that question and cannot speak to it.
The 5950X is a bit tougher. Price/performance scaling is not expected to be linear at these price points, and the ~1.1x performance improvement the 5950X turns in over the 3950X in multi-threaded code is a worst-case scenario. I’d split the difference and say that the 5900X still offers a better price/performance ratio at a higher absolute price, while the argument for the Ryzen 9 5950X is a bit more situational.
AMD now leads Intel modestly in gaming and hammers it everywhere else, in part because Intel has chosen to keep its CPU per-core pricing relatively high. The Core i9-10900K is currently selling for $488 on Amazon, versus an intended $550 MSRP for the Ryzen 9 5900X. As our benchmarks show, the Ryzen 9 5900X is considerably more than 1.13x faster than the Core i9-10900K.
Intel will have a chance to reclaim its single-threaded throne four months from now, when Rocket Lake launches, but it will not reclaim a leadership position in consumer multi-threaded performance unless it moves HEDT CPU cores into the mainstream market. It’s not even clear if that’s possible given the restrictions of the LGA1200 socket, and we’d probably know if the company was planning anything so massive.
Rocket Lake, we now know, tops out at eight cores, while AMD will continue fielding up to 16 CPU cores in the Ryzen 9 5950X. Even with a 1.2x IPC improvement outside of gaming and, say, a 1.05x improvement specifically in gaming, Intel won’t be in a position to challenge the top of the consumer multi-threaded stack. We have to assume a 1.2 – 1.25x performance uplift to even hold Rocket Lake’s overall performance at eight cores steady against Comet Lake’s 10. Intel has given us no indication that we should expect a 1.4x – 1.5x performance uplift in multi-threading, and that’s what it would take for an eight-core chip to challenge a 12-core CPU like the Ryzen 9 5900X.
Zen 3 is an unparalleled success for AMD. The company has literally redefined what kind of performance is possible within a given desktop power envelope and price point over the past 3.5 years. With this new architecture, AMD expanded its leadership position over Intel in multi-threaded and single-threaded workloads, while surpassing it in gaming overall. Intel still leads in specific titles, but AMD can claim to be leading the industry.
Now if you’ll excuse me, I need a nap.
Now Read:
- AMD Is Hitting Market Share It Hasn’t Held in a Decade
- AMD’s Mark Papermaster Dishes the Goods on Zen 3
- AMD Has Scaled Ryzen Faster Than Any Other CPU in the Past 20 Years
No comments:
Post a Comment