AMD Radeon RX Vega Unveiled With 8GB HBM2, Up To 27.5 TFLOPs, Starting At $399
Additional Vega Architecture Details And Pricing
We have previously covered a number of Vega’s new features and architectural details. We’d strongly recommend taking a look at this article from a few months back for some information regarding AMD’s Infinity Fabric, Vega’s next-gen Compute and Geometry Units, the High Bandwidth Cache Controller, Primitive Shader, Draw Stream Binning Rasterizer, and the philosophy that drove AMD during Vega’s design. We do, however, have a few more tid-bits to share on Vega RX.
AMD talked a bit more about Vega’s memory management. Vega uses a simpler page-based memory management, which removes the need for more complicated memory management schemes. Large resources are not required to remain in local memory, when not in use. Active pages will have priority residency in the HBC and inactive pages will be marked for migration to slower memory.
Page size can also be optimized for the particular use case. Smaller pages can be used for sparse accesses and larger pages for dense, or larger sequential accesses. Also, note that Vega’s HBCC (High Bandwidth Cache Controller) can address system memory and on-board video memory, and the caching is programmable, so it can be inclusive or exclusive caching.
The SIMD in Vega’s new Compute Unit can process IEEE-compliant FP16 ops at double data rate (FMA, ADD, MUL, MIN, MAX), and also handle integer operations (MAD, ADD, MUL, MIN, MAX, SHL, SHR). The NCU can do mixed-precision ops as well (MAD_MIX) and there is a full set of pack, unpack, and convert ops available. There is also an extended ISA with 40 new instructions for Vega, which will be made public at some point, though a date wasn’t mentioned.
Vega has a new micro-controller on-board that handles its advanced power management, and the company also revealed that there is a dedicated clock domain for the Infinity Fabric. It can switch to use a secondary static GPU clock generator for deep-sleep state, to improve idle power. And there is also an ultra-low frequency mode for the HBM2.
There are features in Vega that help improve Virtualization as well. Vega is equipped with a security processor which enables hardware validated boot. It has Trusted Memory Zone support and Single-Root I/O for virtualization. Single-Root I/O is a standard developed by the PCI-SIG allows network traffic to bypass the software switch layer of the Hyper-V virtualization stack. Vega also has VCE hardware encode and UVD now enabled when virtualized, and it supports 16 VM guest containers.
As you can see, AMD claims minimum framerates are competitive with the GeForce GTX 1080 and / or the GeForce GTX 1080 Ti, depending on the resolution and application. AMD showed these numbers in an effort to underscore that Vega’s performance falls into the “FreeSync sweet spot” more often than not versus competing solutions.
The goal with the Radeon Packs is to not only discourage coin miners from gobbling up Radeon RX Vega cards, but also incentivize gamers to upgrade their systems and improve their gaming experience.
There are three Radeon Packs in the works. And instead of explaining how this all works ourselves, we’re going to let AMD do it. There was some confusion during our briefings as to how consumers could take advantage of the Packs, so AMD send this information along…
“We designed the Radeon Packs to make it easier for people to upgrade to an enthusiast gaming experience. We believe that there are many gamers out there who are in need of a system-wide upgrade: they want a VRR-capable, high resolution, high refresh monitor, and a performant, enthusiast-class CPU and motherboard combination for a truly high-end gaming platform.
Built around the new Radeon RX Vega graphics cards, Radeon Packs are designed to put all of those technologies at gamers’ fingertips, allowing them to make these upgrades much more cost-effectively than were they to buy each component alone. Radeon Packs include a $200 discount on the 34” Samsung CF791 curved ultra-wide FreeSync monitor, and a $100 discount on select Ryzen 7 processors and X370 motherboard combos -- $300 in combined hardware savings. We're also throwing in $120 worth of new games (titles vary by region) which every purchaser of a Radeon Pack will receive.
These discounts are applied at the time of purchase, and gamers can take or leave pieces as they see fit. So if you already have a great system and just want RX Vega, a new monitor and the games, you can do that. If you only want RX Vega and the free games, you can do that too. For those that want it all, we're giving gamers $300 in discounts off of bleeding-edge components that will transform their system into a dream machine.”
For now, the Radeon Packs are the only way to score one of the Radeon RX Vega 64 Liquid Cooled card or the limited edition Radeon RX Vega 64 air-cooled card with the metal shroud. Although, once the Packs are no longer offered, it is possible those cards will be sold separately. AMD also says their board partners will have the ability to offer the other cards at their discretion, assuming supplies aren’t exhausted.
Also, this just in, there's apparently a Nano card in the works as well (AMD is obviously full of surprises)...
That’s all we’ve got for you at this point. Radeon RX Vega availability is slated for August 14th availability, so mark your calendars
AMD talked a bit more about Vega’s memory management. Vega uses a simpler page-based memory management, which removes the need for more complicated memory management schemes. Large resources are not required to remain in local memory, when not in use. Active pages will have priority residency in the HBC and inactive pages will be marked for migration to slower memory.
Page size can also be optimized for the particular use case. Smaller pages can be used for sparse accesses and larger pages for dense, or larger sequential accesses. Also, note that Vega’s HBCC (High Bandwidth Cache Controller) can address system memory and on-board video memory, and the caching is programmable, so it can be inclusive or exclusive caching.
The SIMD in Vega’s new Compute Unit can process IEEE-compliant FP16 ops at double data rate (FMA, ADD, MUL, MIN, MAX), and also handle integer operations (MAD, ADD, MUL, MIN, MAX, SHL, SHR). The NCU can do mixed-precision ops as well (MAD_MIX) and there is a full set of pack, unpack, and convert ops available. There is also an extended ISA with 40 new instructions for Vega, which will be made public at some point, though a date wasn’t mentioned.
Vega has a new micro-controller on-board that handles its advanced power management, and the company also revealed that there is a dedicated clock domain for the Infinity Fabric. It can switch to use a secondary static GPU clock generator for deep-sleep state, to improve idle power. And there is also an ultra-low frequency mode for the HBM2.
There are features in Vega that help improve Virtualization as well. Vega is equipped with a security processor which enables hardware validated boot. It has Trusted Memory Zone support and Single-Root I/O for virtualization. Single-Root I/O is a standard developed by the PCI-SIG allows network traffic to bypass the software switch layer of the Hyper-V virtualization stack. Vega also has VCE hardware encode and UVD now enabled when virtualized, and it supports 16 VM guest containers.
Expected Performance
We are not going to speculate on Radeon RX Vega 64’s performance just yet, but AMD did mention some interesting numbers we wanted to share. AMD is focused on the user experience with Vega and made some minimum framerate comparisons versus a few NVIDIA cards.
As you can see, AMD claims minimum framerates are competitive with the GeForce GTX 1080 and / or the GeForce GTX 1080 Ti, depending on the resolution and application. AMD showed these numbers in an effort to underscore that Vega’s performance falls into the “FreeSync sweet spot” more often than not versus competing solutions.
Pricing And Radeon Packs
At this point, we’re sure you’re all wondering how much the Radeon RX Vega 64 and RX Vega 56 will cost. The standard, air-cooled Radeon RX Vega 56 will have an SEP of $399 and the standard air-cooled Radeon RX Vega 64 will sell for $499. AMD, however, is also proposing a handful of “Packs” or bundles that offer discounts on games, select Ryzen processors and motherboards, and a Samsung FreeSync monitor.
The goal with the Radeon Packs is to not only discourage coin miners from gobbling up Radeon RX Vega cards, but also incentivize gamers to upgrade their systems and improve their gaming experience.
There are three Radeon Packs in the works. And instead of explaining how this all works ourselves, we’re going to let AMD do it. There was some confusion during our briefings as to how consumers could take advantage of the Packs, so AMD send this information along…
“We designed the Radeon Packs to make it easier for people to upgrade to an enthusiast gaming experience. We believe that there are many gamers out there who are in need of a system-wide upgrade: they want a VRR-capable, high resolution, high refresh monitor, and a performant, enthusiast-class CPU and motherboard combination for a truly high-end gaming platform.
Built around the new Radeon RX Vega graphics cards, Radeon Packs are designed to put all of those technologies at gamers’ fingertips, allowing them to make these upgrades much more cost-effectively than were they to buy each component alone. Radeon Packs include a $200 discount on the 34” Samsung CF791 curved ultra-wide FreeSync monitor, and a $100 discount on select Ryzen 7 processors and X370 motherboard combos -- $300 in combined hardware savings. We're also throwing in $120 worth of new games (titles vary by region) which every purchaser of a Radeon Pack will receive.
These discounts are applied at the time of purchase, and gamers can take or leave pieces as they see fit. So if you already have a great system and just want RX Vega, a new monitor and the games, you can do that. If you only want RX Vega and the free games, you can do that too. For those that want it all, we're giving gamers $300 in discounts off of bleeding-edge components that will transform their system into a dream machine.”
Also, this just in, there's apparently a Nano card in the works as well (AMD is obviously full of surprises)...
https://t.co/4KfVROXkFV confirms @AMD #VEGA NANO. Feast your eyes!! pic.twitter.com/7oJhPXZB5f
— Drew (@Reihen6) July 31, 2017
That’s all we’ve got for you at this point. Radeon RX Vega availability is slated for August 14th availability, so mark your calendars
