Today, AMD is launching its Ryzen 7 5800X3D eight-core CPU at $449. The new chip is both a straightforward improvement over the 5800X and a bigger deal than it might seem. The “3D” in the model name refers to the additional 64MB of vertically-mounted L3 cache (V-Cache) that AMD has attached on top of the CPU die. The new chip features 96MB of L3 cache in total and a slightly lower clock speed compared to the Ryzen 7 5800X. Overall, AMD is claiming a 15 percent improvement in gaming.
This is the first time any high-end semiconductor manufacturer has launched a chip with this kind of configuration. AMD is using TSMC’s CoWoS (Chip on Wafer on Substrate) technology to build the 5800X3D, but dropping a large L3 cache on top of the CPU creates some significant cooling challenges. AMD avoided trapping excess heat on the CPU die by keeping the new L3 well away from typical hotspots.
Putting the 5800X3D In Context:
Slapping a larger cache on a CPU is a tried-and-true method of boosting its performance. Twenty years ago, a large L3 cache was a major distinguishing factor between Intel’s consumer-oriented “Northwood” P4 CPUs and its “Gallatin” Xeon server platform. Intel would later recycle the Gallatin core as the Pentium 4 “Extreme Edition” to challenge AMD’s then-new Athlon 64.
It makes particular sense for AMD to explore this concept given Zen’s long-term development trajectory. Some of the most significant changes AMD has made to its flagship CPU architecture over the past five years have addressed shortcomings in its L3 cache design. Its RDNA2 GPUs distinguished themselves from their RDNA predecessors thanks to the addition of ray tracing and a massive L3. Clearly AMD has had big caches on the brain for quite some time.
There are several factors that make the 5800X3D an interesting chip to examine. First, it may be a partial preview of AMD’s future technology roadmap. We expect to see V-Cache deployed more widely with Zen 4 if the 5800X3D is successful, so measuring V-Cache performance now gives us a window into how AMD chips may respond to larger cache loadouts in the future.
Second, the 5800X3D is the first high-end CPU from any manufacturer that integrates a vertical cache. This kind of integration was historically impossible due to the difficulty of routing so many connections through the top of a CPU die and the thermal problems associated with sticking an additional heat-generating die on top of chips we struggle to cool already. In the past, top-mounted memory was restricted to very low-power devices, and even then the attached RAM was wired to the CPU package externally, not connected directly via through-silicon vias.
While vertical die placement is limited by the need to avoid aggravating existing hot spots, there are real long-term advantages to this kind of manufacturing. The freedom to mount silicon vertically offers chip designers an entirely new dimension to play with when it comes to optimal trace routing. In some cases, it may be possible for companies to reduce latency across a chip by strategically mounting some circuits above others.
The ability to mount cache vertically could pay particular dividends in chiplet-based designs. If AMD wanted to build a CPU with a 96MB L3 cache in a conventional planar layout, each of its chiplets would be much larger than today. That wouldn’t necessarily be a problem for single or dual-chiplet designs, but AMD packs up to eight chiplets per package on a Threadripper or Epyc CPU. Smaller chiplets means more chiplets per package and more cores per physical CPU.
Third, AMD’s decision to slap another 64MB of L3 on the 5800X gives us an excellent window into how bandwidth-limited an eight-core Ryzen actually is – with the caveat that the 5800X3D is not the chip we’d prefer to test for this kind of measurement. The 5800X3D provides one memory channel for every four CPU cores. Chips like the 5950X and Threadripper 3990X only offer one memory channel per eight CPU cores. We’d prefer to measure the impact of more cache on a higher core-count chip, but we may still see some interesting results from the 5800X3D.
Positioning the 5800X3D
Officially, the 5800X3D is a gaming CPU. AMD’s entire marketing push for the chip focuses on that market exclusively. The company went so far as to write in its Reviewing Guide: “We want to be clear, the Ryzen 7 5800X3D is a gaming processor, there is no uplift in content creation performance.”
This statement is stronger than it should have been. There are content creation workloads that benefit from the Ryzen 7 5800X3D’s additional V-Cache, but AMD isn’t targeting those markets with this chip.
When AMD first unveiled its V-Cache technology it demoed the L3 on CPUs with more than eight cores. AMD quietly shifted its plans at some point between its initial unveil and today’s launch, and we can hazard a few guesses as to why.
The additional 64MB of L3 cache built into the 5800X3D isn’t cheap, even if AMD can achieve equal yield between the 5800X3D and non-3D CPUs. AMD’s decision not to launch the Ryzen 9 5900X3D and 5950X3D suggests that the company doesn’t think it needs the parts in-market to compete against Intel in the 6-8 months before Zen 4 launches.
Here’s our take: AMD designed Zen 3 to be V-Cache compatible from the beginning, because it always knew it would introduce the technology in the future and it wanted the option to deploy large L3 caches on AM4 if Zen 4 / AM5 was delayed or if Intel proved to be more serious competition than anticipated. Whatever the conditions were for activating that contingency, they didn’t come to pass – but AMD still wants to launch a swan song for AM4.
AMD is targeting the eight-core market because eight-core chips are the current sweet spot for gamers between high core count and high frequency, and because it only has to bin one good L3 chiplet per physical CPU instead of two. The company’s decision to go forward with a relatively limited launch could be read as a statement on AMD’s overall yield or it could reflect the realities of the pandemic’s strained supply chains.
Launching a single chip like the 5800X3D lets AMD position the CPU as a good-bye present for gamers and consumer enthusiasts. Dropping a full launch stack out the door today and following with Zen 4 in just six months could lead to upset customers who bought into AM4 not realizing it was a dead-end platform.
Compared to the 5800X, the 5800X3D trades a modest amount of clock for a whopping chonk of cache. The fact that AMD has gone so far as to disable overclocking suggests that the company has tuned the 5800X3D’s base clocks very carefully and that attempting to exceed them could damage or destabilize the chip in short order.
The last reason AMD is focusing this chip on the gaming market, specifically, is because it values being able to claim top bragging rights in gaming just as much as Chipzilla does. A lot of boutique systems get sold on the basis of who is leading in a bare handful of gaming tests.
Is the 5800X3D “Just” a Gaming Chip?
Our review evaluation will be based primarily on the 5800X3D’s performance in games, but we expect the L3 cache to improve application performance, too. The question of where – and how much – is one topic we’ll keep an eye on in this review, both in games and in applications.
Different CPU architectures respond to cache boosts in different ways. The Pentium 4 Willamette and Northwood architectures were very sensitive to L2 cache size, while AMD CPUs of the same era were less so. To be clear – “less sensitive” does not mean that adding L2 or L3 cache did not improve performance. It means that different CPU architectures respond differently to a cache increase, and the degree of improvement often depends on whether a chip was cache-starved to begin with.
Tests like Cinebench are unlikely to respond to a larger L3 because they don’t stress cache. Other tests / applications might see a significant performance improvement if the additional L3 allows critical data to be kept closer to the CPU.
How We Test
There are two schools of thought when it comes to CPU-centric testing in games. If you want to get the GPU out of the way and examine how fast the CPU can move data in and out of the core, benchmarking the game at the lowest possible settings is a good idea. While this is academically interesting, it’s not how most people actually play.
We’ve focused our testing on 1080p, 1440p, and 4K, though I did run Borderlands 3 at 720p / Very Low detail to get a sense for how much the 5800X3D’s L3 could boost performance in that scenario. Many games we tested showed a larger delta between the 5800X and 5800X3D at 1080p with the gap shrinking or disappearing altogether at 4K, while the 5800X3D led in other titles at all three resolutions.
Borderlands 3: Tested in Badass Mode.
Hitman 3: Dartmoor: Tested in Ultra detail mode, with Motion Blur off.
Shadow of the Tomb Raider: Tested in “High” graphics mode with SMAAT2x enabled.
Far Cry 6: Tested in Ultra detail with HD textures enabled.
Horizon Zero Dawn: Tested in Ultra detail.
Warhammer III: Tested in Ultra detail with screen space reflections enabled.
Assassin’s Creed: Valhalla: Tested in Ultra Detail mode.
Let’s do the Borderlands 720p numbers first. The Ryzen 7 5800X and Intel Core i9 12900K scored 257.51 and 260 fps in 720p / Very Low detail settings, while the Ryzen 7 5800X3D leaped ahead at 317 fps. The size of that gap does not hold at 1080p but it’s an interesting example of how low-resolution gaming can particularly benefit from large L3 caches.
The Ryzen 7 5800X3D is faster than both the Ryzen 7 5800X and the Core i9-12900K, particularly at 1080p. The size of the uplift ranges from modest to substantial. At 1080p, the Ryzen 7 5800X3D wins or ties every game except for Warhammer III. As expected, the gains are not evenly distributed. Some titles gain a great deal of performance while others don’t budge.
Our tests show a 10 percent uplift for the Ryzen 7 5800X3D over the Ryzen 7 5800X at 1080p. Since gains are game-dependent, choosing a different selection of games would allow a reviewer to hit AMD’s 15 percent claim. The 5800X3D’s lead over the Core i9-12900K is smaller, but AMD still clinches a 2.5 percent overall performance uplift. That’s a tie for all intents and purposes, but it’s also enough of a lead for AMD to claim the title of “Fastest Gaming CPU” in its marketing copy.
The gains shrink as one moves through 1440p and 4K and workloads become more GPU-bound. The 5800X3D continues to offer a 1-3 fps advantage at 4K depending on the title.
Non Gaming Performance
We promised to evaluate the 5800X3D primarily on its gaming performance, but we want to consider the non-gaming angle too, if only unofficially. These results illustrate part of why AMD doesn’t want to market the Ryzen 7 5800X3D as anything but a gaming chip — but they also show more improvement than AMD implies.
Performance in Blender 3.1.2 and Cinebench R23 favors higher clocks over more cores and the 5800X wins these comparisons against the 5800X3D. The 5800X3D beats its higher-clocked sibling in our H.264 encoding tests and in one of the H.265 tests, but loses the last.
The H.264 and H.265 tests we run are primarily memory bandwidth tests, with some impact from clock and core count. The two standards do not respond identically to the 5800X3D, however. H.264 encoding benefits from the additional cache, while H.265 mostly does not. The AMD CPUs were tested with Prefetch(12) in H.264 and Prefetch(8) in H.265. This proved to be the most advantageous configuration for them. The Intel Core i9-12900K used Prefetch(12) for both H.264 and H.265.
The additional L3 cache gives the 5800X3D a 1.14x performance boost in H.264 Ultra Fast and a 1.13x boost in H.264 QTGMC Medium. In 7zip, the Ryzen 7 5800X3D is 1.11x faster in compression but slightly slower in decompression compared to the Ryzen 7 5800X.
The Core i9-12900K is much more expensive than the Ryzen CPUs. Unlike in gaming, where the 5800X3D and Core i9-12900K are often on top of each other, the Alder Lake CPU leads or ties all of these tests. This is to be expected given the price difference between the two chips.
If the Ryzen 7 5800X3D highlights some of the strengths of adding additional cache, it also highlights some of the weaknesses. The 5800X3D runs at just 89 percent of the 5800X’s base frequency and 96 percent of its boost frequency, and the extra L3 doesn’t always compensate for the diminished clock. Overall, the Ryzen 7 5800X3D offers between 95 – 115 percent the performance of the Ryzen 7 5800X. An equally clocked version of the CPU would be between 0 – 25 percent faster than the vanilla 5800X, with most applications picking up at least a small improvement.
The 5800X3D is an interesting CPU. It provides the additional oomph that AMD promised in both games and also in some applications, though there’s no way to predict which games and apps will benefit until you test them. It also gives AMD a real argument that it offers the fastest gaming CPU, with the caveat that we have not been offered the opportunity to test the Core i9-12900KS.
If you are looking to maximize your gaming value per dollar, however, you’ll want to consider other chips. The Core i9-12900K is overkill for gaming compared to either AMD option. Obviously there are also cheaper Alder Lake CPUs. If gaming is your principle workload and you want to run an Intel system, we recommend something in the Core i5 or Core i7 families.
Similarly, AMD users who do more than game may find chips like the Ryzen 9 5900X to be a better value than the Ryzen 7 5800X3D. The 12-core 5900X is selling for as little as $394 online, well below this chip’s $450 MSRP.
The performance difference between the 5800X and the 5800X3D is not large enough to justify the price gap between the two, even by boutique pricing standards. This launch is about AMD slugging it out with Intel at the very top of the market where money is no object. Gamers who want the absolute fastest gaming CPU AMD manufacturers may want to seriously consider the Ryzen 7 5800X3D. Gamers who want the excellent price/performance ratio AMD is known for will be happiest if they consider other chips.