This isn't an AMA on Radeon RX Vega, the consumer graphics card based on the Vega architecture, but instead the professional graphics card based on Vega - which fights Tesla/Quadro and Titan X from NVIDIA. Senior VR and Chief Architect at Radeon Technologies Group, Raja Koduri, will be hosting the AMA.
AMD detailed its upcoming Vega GPU architecture during its recent Financial Analyst Day, announcing their new Radeon Vega Frontier Edition graphics card, a new ThreadRipper 16C/32T processor, and the beast 32C/64T chip with Epyc.
But for the gamers, I think that the tease of HBCC (High Bandwidth Cache Controller) running on Rise of the Tomb Raider and seeing HBCC disabled and what I'm sure is the Radeon RX Vega running the game at 4K at 57.9FPS average, with 13.7FPS minimum. This puts the RX Vega at the speeds of the GTX 1080 11Gbps, but when HBCC is enabled - holy crap, there is a huge change.
AMD's upcoming Radeon RX Vega with HBCC enabled was running ROTT at 76FPS average (up from 57.9FPS without HBCC) but it's the minimum FPS that is where the magic happens: 49.1FPS with HBCC enabled, up from the 13.7FPS without HBCC. This is amazing, especially for enthusiast gamers like myself who want much better minimum FPS performance than brute force maximum FPS.
If you are running a 4K60 display, hitting 4K 60FPS is easy with a GeForce GTX 1080 Ti, but the minimums aren't anywhere near as good as they are with HBCC on Radeon RX Vega. The big problem that we have here is that we now need to see HBCC's performance benefits, if there are any, to other games... all games. Show me PlayerUnknown's Battlegrounds, Battlefield 1, CS:GO, Overwatch - where minimum FPS performance is very important, especially the twitch first-person shooters like CS:GO.
We've been reporting about GDDR6 technology for a while now, but SK Hynix detailed its upcoming memory technology in their recent memory catalog, with GDDR6 scheduled for Q4 2017.
SK Hynix confirms two different variants of GDDR6, with both offerings being 8Gb (1GB) modules with different frequencies: 12Gbps and 14Gbps. Both modules will use 1.35V.
I've got some details on what memory bandwidth numbers to expect, which is exciting:
- GDDR6 @ 12Gbps on 256-bit: 384GB/sec
- GDDR6 @ 14Gbps on 256-bit: 448GB/sec
- GDDR6 @ 12Gbps on 384-bit: 576GB/sec
- GDDR6 @ 14Gbps on 384-bit: 672GB/sec
AMD has just made its Radeon RX 560 graphics card official, powered by the Polaris 21 GPU and its 1024 stream processors, 64 TMUs, 16 ROPs, and 2GB/4GB configurations.
AMD has the base/boost GPU clocks at 1175/1275MHz, respectively, while the 2GB or 4GB of GDDR5 is clocked at 7Gbps on a 128-bit memory bus, and 80W TDP. Comparing this to the Polaris 11-based Radeon RX 460, with its 896 stream processors, 48 TMUs, and same 16 ROPs.
The Radeon RX 460 had its GPU clocks at 1090/1200MHz, and the same 7Gbps GDDR5 on the 128-bit memory bus and 75W TDP.
AMD has unleashed a bunch of new products today at their Financial Analyst Day, with a tease of ThreadRipper with 16C/32T, their new Naples platform and the Epyc CPU with 32C/64T, and the exciting new Radeon Vega Frontier Edition graphics card.
AMD also detailed their GPU roadmap through to 2020, with Vega arriving on 14nm - but a 14nm+ respin that I think we'll see in early 2018. The bigger tease is Navi on 7nm for 2018-2019, and a brief glimpse of 'Next Gen' sometime in 2019-2020.
AMD has announced its new Radeon Vega Frontier Edition graphics card, which isn't part of the consumer graphics card family, but instead the Radeon Pro line.
Radeon Vega Frontier Edition rocks 16GB of next-gen HBM2 memory, 13 TFLOPs of performance (NVIDIA's new TITAN Xp has 12 TFLOPs) of single precision compute performance, while we have the Vega GPU clocked at 1586MHz.
AMD detailed its Radeon Vega Frontier Edition specs with 64 next-gen compute units (nCU - 4096 stream processors), 12.5 TFLOPs of single precision compute performance (FP32), 25 TFLOPs of single precision compute (FP16), 16GB of HBM2 alongside HBC, and 8K display support.
Radeon Technologies Group boss Raja Koduri has posted something a little cryptic on his Twitter account, where he tweeted a tease on Radeon RX Vega just over 20 minutes ago.
Koduri tweeted: "Ok..complete the sentence..we stand today on the edge of a ...... (first 10 people who get this right will get something special) @radeon". I was the first to answer, saying "new frontier" by JFK, and plenty of others followed with the same answer.
He had a follow up tweet that read "Bonus points to those who can cite the author and my favorite science and space section", so I don't know if that's another tease of the space-themed Vega architecture.
It looks like the release of Radeon RX Vega is much closer than we think, and the tease of a new frontier is incredibly exciting. Bring it on, AMD!
SK Hynix has released its new Q2 2017 product databook, and inside we get some more details on their current and next-gen graphics memory products, with the company listing 12Gbps and 14Gbps modules of GDDR6, as well as 1.6Gbps HBM2 memory.
AMD's upcoming Radeon RX Vega is powered by HBM2 memory provided by SK Hynix, while NVIDIA is tapping Samsung made HBM2 memory. Radeon RX Vega should rock SK Hynix's current 1.6Gbps HBM2 that would provide up to 409.6GB/sec of memory, unless AMD were to use more HBM2 stacks, then the bandwidth and HBM2 would double. The memory bandwidth would reach a pretty damn high peak of 817GB/sec, easily beating out any graphics card in its path - and getting close to NVIDIA's new HBM2-based Tesla V100 and its memory bandwidth numbers of 900GB/sec.
Remember that AMD also has its HBCC (High Bandwidth Cache Controller) working alongside the Vega NCU, which will be able to greatly improve minimum FPS performance, and has up to 512TB of virtual address space - insanity.
AAMD will most likely release their new Radeon RX Vega graphics cards in 4GB and 8GB capacities, something that has been recently teased as the Radeon RX Vega Nova, RX Vega Eclipse, and RX Vega Core. Our world exclusive report also states that AMD will have just 16,000 or so Radeon RX Vega graphics cards to launch in the first few months, because of the scarcity of HBM2 right now.
UploadVR is in some serious trouble with the SF-based VR startup being sued by a former employee who alleges the company engaged in sexual discrimination and sex/gender harassment.
TechCrunch reports that the lawsuit was filed today with the Superior Court of California, claiming that UploadVR is plagued by "rampant sexual behavior and focus, creating an unbearable environment for Plaintiff and other female employees". The suit continues, adding: "Defendants purposefully and expressly created a 'boy's club' environment at work, focused on sex and degrading women, including female employees".
UploadVR co-founders Will Mason and Taylor Freeman "would frequently talk about much sex they were going to have at each party and how many girls they were going to have sex with". It gets worse, as the VR startup reportedly has a "kink room" that has a bed in its office, where people can have sex during parties - and I guess, whenever the hell they feel like it.
The suit adds that there is more brewing under the surface at UploadVR, as the company has "created a hostile and toxic working environment" reports The Verge. The allegations include employees that hired strippers and prostitutes, where male employees would kick out female employees so they could "use the spaces for sexual intercourse with party attendees". The suit continues, alleging widespread discrimination that saw male staff compensated more than female staff, and female staff not being reimbursed for business expenses.
Some more information on next-gen graphics cards has arrived thanks to a press release from SK Hynix about their new HBM2 tech, with SK Hynix also shipping GDDR6 memory in Q1 2018.
SK Hynix was also at NVIDIA's recent GPU Technology Conference showing off their 8Gb GDDR6 module with 16Gbps of bandwidth, up from the 11Gbps of bandwidth available on the fastest GDDR5X on GTX 1080 Ti. GDDR6 is listed as a Q4 2017 product, meaning it will enter mass production in Q4 2017 for a huge Volta launch in Q1 2018 - at least according to WCCFTech.
I think we'll see a Pascal refresh based on GDDR6 at first, and then a full Volta line up with a blend of GDDR6 and HBM2 technologies - GDDR6 for the GTX 2080/2070, and HBM2 for the Volta-based TITAN Xv, and what I'm hoping for: a dual-GPU Volta graphics card.
NVIDIA will most likely launch a GeForce GTX 2080 (or GTX 1180) with 8GB+ of GDDR6 @ 14Gbps on a 256-bit memory bus, getting close to GTX 1080 Ti performance. The full monster Volta card will probably launch with a 384-bit memory bus and 12-16GB of GDDR6 @ 16Gbps, while a TITAN Xv could unleash a full 512-bit memory bus, or shift over to HBM2 depending on the cost/scarcity of HBM2 from Samsung.