CES: AMD Details V-Cache, Zen 4, and the Ryzen 6000 Mobile Family
CES has kicked off, with a flurry of releases from the usual suspects. Like a lot of companies, AMD uses CES to preview the products it will launch throughout the year. Things were pretty quiet for AMD in the back half of 2021, but the company has a number of interesting launches teed up for 2022. Major new CPU and SoC initiatives from AMD in 2022 include the Ryzen 6000 Mobile family, Zen 4’s debut, and AMD’s V-Cache-equipped Zen 3 cores.
V-Cache Comes to Just One CPU
When AMD announced V-Cache, we thought the company’s higher-end chips like the 5900X and 5950X would be the obvious targets for the capability. One of the best ways to improve the performance of memory-bound CPUs is to add more L3 cache, so the larger CPUs seemed like prime targets.
Instead of going that route, AMD is going to introduce just one V-Cache enabled SKU. The Ryzen 7 5800X3D will be an 8C/16T CPU with a 3.4GHz base clock and a 4.5GHz boost. That’s slightly below the base/boost clocks of the Ryzen 7 5800X, which comes in at 3.8GHz and 4.7GHz respectively. According to AMD, it made the decision to lower clock speeds slightly rather than binning a stricter variant of the chip.
The reason AMD is limiting itself to just one SKU is that the 5800X3D is a bit of a pipecleaner (that’s our term, not AMD’s). A two-chiplet design is more complex than a single chiplet and requires a second die, which would increase CPU prices and lower availability at a time when parts are already in short supply.
This is the kind of limited launch we would expect when a company is testing a new capability, and integrating a huge L3 cache directly on top of the CPU definitely qualifies. We know Milan X will debut V-Cache in servers this year and expect to see the capability on at least some Zen 4 chips going forward, though AMD has not confirmed that it will offer a 64MB V-Cache as a standard option on any specific Zen 4 parts.
The 5800X3D is specifically intended for gamers and will deliver an estimated 1.15x performance improvement on average in gaming titles relative to the 5900X. AMD claims between a 1.0x and a 1.4x speed-up at 1080p High. Productivity gains are also expected but are generally lower than 1.15x. Performance in games against the Intel 12900K at 1080p High ranges from 0.98x – 1.2x faster, in an unspecified suite of titles. Overclocking might nudge those results higher, though AMD’s CPUs aren’t known for overclocking performance these days.
The big question here is where AMD will price the CPU. Right now, the 5900X is a $550 CPU and the 5800X is $450 on paper. Amazon shows the 5800X available for $400 as of this writing, but your mileage may vary. The 5800X3D needs to keep its price around $500 to be taken seriously or else risk being outclassed by the 5900X. While the 5800X3D may be faster in gaming, the 5900X will likely still retain an advantage in any well-threaded application. We expect a launch later this spring and will have more details on price and positioning closer to that date.
Ryzen 6000 Mobile SoC Family
Next up, there’s the Ryzen 6000 Mobile family and what AMD is calling “Zen 3+.” This new family of CPUs encompasses 10 chips and spans 15W – 45W, but eight of the 10 SoCs are 35W-45W parts. These new SoCs are built on TSMC’s 6nm node. 6nm is a cost-optimized version of 7nm that offers roughly 1.18x better density but no official performance or power improvements.
Zen 3+ CPUs are power-optimized versions of Zen 3. They do not offer any architectural improvements over Zen 3 cores; the gains for Zen 3+ over Zen 3 in mobile are due to factors like better power management, higher clocks, and potential gains in power efficiency, not any architectural changes.
In AMD’s lingo, “HS” series chips are 35W parts while “H” series chips are 45W. Most of the new chips have 12 RDNA2 compute units up from eight on Vega, but the 6600H, 6600HS, and 6600U offer just six CUs. AMD’s decision to define these chips as 15W-28W as opposed to 15W may reflect higher burst power consumption despite efficiency improvements. Base clocks on the Ryzen 6000 Mobile U-series are much higher than they were on the previous 5000 series.
Given the improvements between RDNA2 and Vega we’re willing to bet that six RDNA2 core clusters should still match or exceed eight Vega CUs, but we won’t know that for certain until laptops are available for review.
AMD has increased top end frequencies in the Ryzen 6000 Mobile family and claims the new cores sustain much higher single-threaded clocks than previous chips. Performance on the 6800U is claimed to be between 1.1x and 1.3x faster than the 5800U, though this will undoubtedly depend on the cooling solutions and overall design of the comparison laptops.
AMD is claiming a range of power improvements in specific use-cases, mostly web browsing, video streaming, and video conferencing. These are all common areas where manufacturers prefer to share power consumption data because it’s easiest to show improvements here. Real-world power consumption depends on the screen, Wi-Fi radio, and other system components in addition to the SoC and GPU (if any).
AMD is claiming a 1.5x larger compute engine (that’s the bump from 8 CUs to 12), up to 1.5x more memory bandwidth (courtesy of the shift to DDR5), an L2 cache that’s 2x larger and twice the ROP throughput. Vega 8 is a 512:32:8 design (cores, TMUs, ROPs), implying that mobile RDNA2 is somewhere between 768:32:16 and 768:64:16. Clock speeds have also increased on some parts, from 2GHz to as high as 2.4GHz.
AMD argues that these improvements are collectively worth 1.8x – 2x performance in the same power envelope. There are several reasons to think the company is telling the truth. We know that RDNA2 is ~1.25x more efficient than Vega, clock for clock. A 1.2x clock increase isn’t worth much on its own, because mobile SoCs are so bandwidth-limited. With Ryzen 6000 Mobile, however, AMD is also moving to DDR5 with support for DDR5-5200 and LPDDR5-6400. Compared to DDR4-3200, DDR5-5200 offers 1.62x more bandwidth. The exact improvement will depend on what you own now versus what you buy, but a 1.5x – 1.65x increase in memory bandwidth over DDR4 is likely worth 1.2x – 1.5x depending on the game. The exact amount will depend on the specifics of the individual title. Factor in some modest improvement from clock speed and AMD’s performance claims start looking plausible. It wouldn’t be surprising if the 15W-28W APUs are more power-limited than anything else when it comes to beating the older Ryzen 5000 Mobile family.
AMD will be encouraging OEMs to use its mobile gaming GPUs and 35W+ APUs in the same system so as to take advantage of capabilities like SmartShift Max. SmartShift Max is AMD’s brand name for allocating system power flexibly between CPU and GPU to maximize overall performance. The feature has seen relatively limited uptake thus far, but AMD emphasized that it expects more than 200 premium design wins in 2022, far more than just a few years ago. If the company’s mobile GPUs are competitive (or simply available) we may see more laptop manufacturers take advantage of this capability.
AMD will also be the first laptop SoC designer to implement Microsoft’s “Pluton” core to handle security functions; systems with the feature should be available late this spring. There may also be a pair of TenSilica DSPs integrated into the SoC (a Vision Q6 and C5, reportedly), but AMD did not announce those features.
The last major announcement from AMD relates to Zen 4. AMD has confirmed that Zen 4 will ship in the back half of the year on TSMC’s N5 node. The new SoC will use the AM5 socket and will transition to a Land Grid Array (LGA).
After decades of futzing with pins, AMD will do away with them in consumer desktop systems. Some of AMD’s power circuitry is moving to the top of the chip now that the bottom will be nothing but pins, leading to this rather interesting heatspreader design:
AM5 will support AM4 coolers, making a new cooler potentially unnecessary for would-be upgraders. AMD may change some of its TDP targets, but it is unlikely to adjust them dramatically compared to AM4. Zen 4 will use DDR5 and will offer support for PCIe 5.0. It will likely take until at least the end of 2022 for DDR5 to become available at reasonable prices and there are no devices with PCIe 5.0 support shipping to customers just yet, so AMD isn’t missing much by delaying official support until the back half of the year.
It’s great to see AMD finally moving beyond Vega. After a poor initial showing in 2017, the GPU architecture redeemed itself as a mobile chip and powered multiple generations of AMD APUs. It’s time for AMD to unify its gaming architectures and RDNA2 is a worthy successor. The increased power efficiency claims and additional performance should help AMD stay competitive against Intel’s 12th Gen CPUs and mobile Alder Lake. The shift to DDR5 will do only good things for AMD’s mobile GPU and integrated GPU performance.
The 5800X3D announcement is a little disappointing if you hoped to see 16-core options with 96MB of L3 per chiplet, but there’s no guarantee that we won’t see those chips in 2022 or early 2023. It’s possible that AMD will use the 5800X3D as a pipecleaner and announce a wider range of V-Cache-equipped CPUs later this year.