Ever since AMD refreshed most of its Ryzen product range last year using GlobalFoundries' 12LP process—an enhanced version of the foundry's 14 nm FinFET node with higher-performance front-end-of-line transistors but only optional density improvements—and expanded that refresh to Radeons with the RX 590, I sort of expected the company to do the same with its Vega 10 graphics chip at some point. Instead, AMD surprised us at CES by introducing a consumer version of its Vega 20 data-center GPU on board a new graphics card: the Radeon VII. Taking Vega 20 out of the data center means AMD is first to market with a gaming GPU fabricated on TSMC's cutting-edge 7 nm FinFET process.
Vega 20 initially made an appearance in the Radeon Instinct MI50 and MI60 accelerators introduced late last year. Instead of making the biggest chip possible as quickly as possible like Nvidia has with its compute accelerator chips of late, Vega 20 only modestly expands the processing capabilities of the Vega 10 GPU to support some reduced-precision data types useful in deep learning applications. TSMC's 7-nm process is a real node shrink with all of the areal scaling such an advance implies, and it does most of the work in delivering a generation-on-generation performance improvement.
Vega 20 takes most of the existing Vega 10 GPU and shrinks it to 331 mm², down from 495 mm² for Vega 10. By shrinking the die in this way, AMD made room on Vega 20's underlying silicon interposer to add on two more stacks of HBM2 RAM. Vega 20 now talks to its on-package memory over a 4096-bit wide bus, and with HBM2 speeds of 2 Gb/s per pin, that works out to a whopping 1 TB/s of theoretical memory bandwidth. Compare that to the 484 GB/s that Vega 10 enjoyed on board the RX Vega 64, and we should expect a significant performance increase from this move alone.
TSMC has claimed that its first generation of 7 nm transistors can deliver up to 20% higher performance or 40% power savings at the same performance versus its 16 nm node. AMD pushed the power-savings-versus-performance-increase slider all the way to the right for the Radeon VII. The company has already said that this card would deliver its improved performance at the same power as the RX Vega 64.
As a result, the Radeon VII's board power rings in at 300 W, up 5 W over even the notoriously power-hungry RX Vega 64. In return, AMD clocks the Radeon VII's Vega 20 chip at a 1750 MHz typical boost speed, up about 13% over the Radeon RX Vega 64 in its air-cooled configuration. As we'll see from some back-of-the-napkin calculations to come, that clock-speed increase more than makes up for the fact that AMD has had to disable four of Vega 20's 64 compute units, presumably for yield reasons. Along with those compute units, the Radeon VII appears to lose 16 of its texture units. We're confirming that figure with AMD and will update our numbers as soon as we can.
In its descent from the data center, the Vega 20 GPU loses some features that will be of little use to enthusiasts. For example, Radeon Instinct accelerators using Vega 20 support both PCI Express 4.0 and coherent card-to-card communications using the Infinity Fabric interconnect, but the Radeon VII sticks to the far more common PCIe 3.0 standard and drops the data-center class inter-card connection. Vega 20 also offers support for FP64 (aka double-precision) data types at up to 1/2 the FP32 rate, and unusually, some of that FP64 processing capability survives on the Radeon VII in a form that may be useful to folks who need it.
This morning, AMD told Anandtech that it's ultimately decided to enable 1/4-rate FP64 on the Radeon VII after earlier statements suggesting that 1/8 would be the rate of choice. Some more back-of-the-napkin math suggests that in its final form, the Radeon VII could crunch through FP64 workloads at 3.35 TFLOPS. That's quite a bit better than the 1.6-TFLOP peak rate that the Kepler-powered GTX Titan offered back in the day. GTX Titan cards have endured as relatively cheap founts of FP64 performance for those who need it, so AMD might sell some Radeon VIIs to scientific-computing folk who want useful FP64 performance alongside an otherwise much more modern graphics card.
Here's a table comparing the Radeon VII's basic specs to several of today's most common pixel-pushers:
|RX Vega 56||1471||64||224/112||3584||2048||410 GB/s||8 GB|
|GTX 1070||1683||64||108/108||1920||256||259 GB/s||8 GB|
|RTX 2060 FE||1680||48||120/120||1920||192||336 GB/s||6 GB|
|RTX 2070 FE||1710||64||120/120||2304||256||448 GB/s||8 GB|
|GTX 1080||1733||64||160/160||2560||256||320 GB/s||8 GB|
|RX Vega 64||1546||64||256/128||4096||2048||484 GB/s||8 GB|
|Radeon VII||1750||64||240/120?||3840||4096||1 TB/s||16 GB|
|RTX 2080 FE||1800||64||184/184||2944||256||448 GB/s||8 GB|
|GTX 1080 Ti||1582||88||224/224?||3584||352||484 GB/s||11 GB|
|RTX 2080 Ti FE||1635||88||272/272||4352||352||616 GB/s||11 GB|
|Titan Xp||1582||96||240/240||3840||384||547 GB/s||12 GB|
|Titan V||1455||96||320/320||5120||3072||653 GB/s||12 GB|
From this chart, a couple things should stand out. One is that the Radeon VII offers by far the highest theoretical memory bandwidth of any card we've ever had our hands on. At $699, the Radeon VII is also the cheapest way to get 16 GB of RAM on a graphics card yet. We typically don't see that much VRAM until one gets into pro visualization products, but with the advent of the Radeon VII, AMD is pushing 16 GB as a new enthusiast standard. Since memory capacity is one of the Radeon VII's clearest spec wins over the RTX 2080 and its 8 GB of RAM, it's no surprise that AMD is making hay of this feature. The company suggests a number of current and upcoming games can benefit from having more than 8 GB of video memory available for gaming at 4K and maximum settings, and it also highlighted the value of that memory for prosumer workloads like 4K and 8K video processing with Adobe Premiere.
On a gut level, this push for higher memory capacities on cards destined for 4K gaming makes sense. We know that texture sizes are growing and seem likely to grow further as gamers demand higher and higher quality visuals, and I know a couple video producers who might appreciate having more video memory available without paying through the nose for it. AMD says Far Cry 5 can occupy as much as 12.9 GB of VRAM at 4K and max settings, and the company provided us a scary-looking frame time graph that suggests major inconsistency could arise if an RTX 2080 runs over its available memory pool. Far Cry 5 is one of the games we use in our test suite, so that's an easy point to return to for further analysis.
|RX Vega 56||94||330/165||5.9||10.5|
|RTX 2060 FE||81||202/202||5.0||6.5|
|RTX 2070 FE||109||246/246||5.1||7.9|
|RX Vega 64||99||396/198||6.2||12.7|
|GTX 1080 Ti||139||354/354||9.5||11.3|
|RTX 2080 Ti||144||473/473||9.8||14.2|
As a hotter-clocked Vega chip, the Radeon VII offers modest but meaningful theoretical performance improvements versus its RX Vega 64 predecessor in every one of our usual theoretical peak figures. Really, there are no surprises here. The basic resource allocation of Vega 20 didn't change as part of the move to the 7 nm node, and AMD's widening of the chip's memory bus largely overshadows the clock-speed-related gains in the table above. Graphics tends to be a memory-bandwidth-bound workload, and making sure the beast is fed is unsurprisingly a priority for GPU architects. Perhaps the addition of that wider memory bus speaks to what sat atop the list of fixes AMD envisioned for Vega after its first time out on shipping products.
The Radeon VII card itself should be a well-known quantity by this point. Yours truly wasn't on the ground at CES, but the card was, and the press enjoyed extensive access to it during the show. The Radeon VII takes the basic bead-blasted, diamond-cut aluminum shroud design introduced on limited-edition Radeon RX Vega 64 cards and turns it into a home for three fans blowing onto an open-style heatsink more typical of custom cooler designs from board partners.
We're hesitant to tear down and repaste coolers that need to go back on large GPU packages with HBM RAM like Vega 20, but the fine folks at Gamers Nexus have no such qualms, and they've already torn down their card ahead of today's embargo lift. From GN's work, we know that the Radeon VII cooler uses a massive, complicated copper vapor chamber to wick heat away from the Vega 20 package and into an aluminum fin stack. Perhaps thanks to that vapor chamber, AMD is able to keep the Radeon VII reference design within the boundaries of two slots, although the cooler shroud does jut slightly beyond the edge of the expansion-card bracket. The 300 W board power we alluded to earlier manifests itself in this card's pair of eight-pin PCIe plugs.
It's interesting to see AMD follow in Nvidia's Founders Edition footsteps for Turing by introducing an open-air cooler on the Radeon VII despite its much higher board power than any GeForce RTX card. The RX Vega 64's blower-style heatsink may not have been the best thing going for absolute noise levels, but it did dump the copious heat that card produced out the rear of a PC. Systems with a Radeon VII inside, on the other hand, will need to be well-ventilated to remove the hot air exhausted by its open-style cooler, although that demand is no different than the challenges posed by Nvidia's latest reference coolers.
AMD stickers the Radeon VII at $699, or the same suggested price as GeForce RTX 2080 partner cards. The company has drawn a bead on Nvidia's best-performing Turing card under $1000 over and over in the run-up to this morning. Let's see if AMD hit its target now.