Video Cards & GPUs - Page 301

Apple has officially released its new Mac Pro into the wild, a new PC that when configured with the best internals you can buy will set you back over $53,000 -- but at the same time AMD has announced the new Radeon Pro W5700X workstation graphics card.

The new Radeon Pro W5700X is a higher-end version of the Navi 10-based Radeon Pro W5700, except this time it features the full Navi 10 die. This means it rolls out with 2560 shading units, versus the 2304 found in the regular Radeon Pro W5700. It also doubles the VRAM count up to 16GB of GDDR6, versus just 8GB of GDDR6 in the regular Radeon Pro W5700.

AMD's new Radeon Pro W5700X is better specced across the board with 9.5 TFLOPs of single-precision compute performance, versus 8.89 TFLOPs on the regular Radeon Pro W5700. We have 160 texture units versus 144 on the regular W5700, while both cards still maintain 64 ROPs. The Navi 10 GPU is clocked lower on the new Radeon Pro W5700X with 1855MHz boost, down from 1930MHz on the normal W5700.

The next GPU architecture we get from NVIDIA will be Ampere, but we're already hearing about the GPU architecture after that in the form of Hopper.

Hopper will reportedly be an MCM design or multi-chip module -- think of the multi-chip GPU design as the chiplet technology that has allowed AMD to leapfrog Intel in the CPU market... But for next-gen GPUs.

A recent patent unearthed by 'kopite7kimi with NVIDIA choosing Hopper to represent the late Grace Hopper, who was a computer scientist who popularized the idea of machine-independent programming languages. Her work led the world into the creation of COBOL, the high-level programming language.

Intel will have much more to share about its new Xe GPU architecture in 2020, but before then the hype train just spat out a little more steam with Raja Koduri teasing a new high-performance Xe GPU.

Koduri recently said the team designing the high-performance Xe GPU calls it the "faster of all silicon", and said it rubs shoulders with some of the largest silicon ever designed. Koduri recently met with the engineering team based in Bangalore, India -- where he tweeted out the Xe GPU was "the baap of all", which Wccftech's Usman Pirzada says translates to "the father of all".

With Raja referring to this Xe GPU as some of the largest silicon in the world, Pirzada thinks Raja is talking about a GPU with a die size pushing 750-800mm2 -- this would make Intel's new Xe GPU one of the largest GPUs ever designed. With this, we should expect some great performance but we're still a while out from that.

Qualcomm is really pushing desktop-level features with the Snapdragon 865, with unparalelled amounts of power that provides huge performance for unreleased smartphones -- but also some important features lke upgradeable GPU drivers.

The new smartphones coming in 2020 powered by Qualcomm's new Snapdragon 865 will support updateable GPU drivers, with Qualcomm working with Google to have Adreno GPU Driver updates freely available on the Google Play Store.

This small (but very significant) change in Snapdragon 865 with updateable GPU drivers brings mobile users into the world of the PC -- as we've had this for decades. Qualcomm made a point that the GPU updates will come out "well after" the phone itself launches.

I've been told by NVIDIA directly that there will "not be a GeForce RTX 2080 Ti SUPER" graphics card, but that hasn't stopped the team from TecLab from making a Frankenstein-style RTX 2080 Ti SUPER themselves.

The team ripped apart two separate GALAX GeForce RTX 2080 SUPER graphics cards, taking their faster 16Gbps GDDR6 modules and using 11 x 16Gbps modules onto an RTX 2080 Ti. The stock GeForce RTX 2080 Ti has 11GB of GDDR6 at 14Gbps, but the RTX 2080 SUPER ships with faster 16Gbps modules.

NVIDIA, if it were to launch a GeForce RTX 2080 Ti SUPER, would make it much beefier than just increasing GDDR6 bandwidth. We should expect more CUDA cores as well as maybe a wider memory bus, with faster GDDR6 as well.

There have been whispers that some big job losses were coming to Silicon Valley, and in that time I've reached out to some industry sources who have corroborated such things -- but now, it seems it's beginning.

Chris Hook and Heather Lennon, former Radeon Technologies Group marketing staffers, have left Intel. In the last couple of years they both left the crumbling team of RTG where most of them found a new home at Intel, and now Hook and Lennon have left the odyssey. I reached out to both of them and had the news confirmed independently, too.

Given that Intel are really starting to see their wheels wobble in the consumer market with huge market share losses to their consumer and now HEDT processors, AMD is really putting the pressure on them. Intel did a shady trick of having an NDA of their latest 10th-gen HEDT chips just before AMD's new Ryzen Threadripper 3950/3970X launched -- and then within a couple of days this news breaks out.

NVIDIA GeForce GTX and GeForce RTX gamers will want to jump on the latest GeForce Game Ready 441.41 drivers, which include optimizations for Halo: Reach that is about to drop on the PC, and offers optimal support for the new Quake II RTX v1.2 update.

Game Ready for Halo: Reach means that you get the latest performance optimizations, profiles and bug fixes from NVIDIA for the game -- while you get optimal support for the latest version of Quake II RTX, which provides even better quality graphics, water reflections, and ray tracing goodness.

There's also some issues fixed with Red Dead Redemption 2 when SLI is enabled and you're running Vulkan, while both Shadow of the Tomb Raider and Forza Horizon 4 get some tweak love in the new GeForce 441.41 drivers. The big note here is the strange "known issue" about V-Sync which NVIDIA outright says in the new drivers: "V-Sync does not work". Alrighty.

NVIDIA very quietly updated its Tesla range of graphics cards in the last 24 hours, introducing the new Tesla V100s graphics card that packs a faster Volta GPU and quicker HBM2 memory.

The new Tesla V100s (note the small 's', I guess this means 'SUPER' in the Volta sense) is pretty much just as faster version of the original Tesla V100, with the same GV100 GPU on the 12nm node and the same 5120 CUDA cores -- but it's the clock speed and memory bandwidth that are impressive.

NVIDIA's original Tesla V100 has its GV100 GPU boost clock at up to 1367MHz resulting in 14 TFLOPs of compute performance, while the new Tesla V100s has its GV100 GPU at up to 1601MHz boost clock which results in 16.4 TFLOPs of compute performance.

NVIDIA has been pretty quiet when it comes to multi-GPU rigs with their latest Turing-based GeForce RTX and GeForce RTX SUPER graphics cards, but is all of that about to change? Quite possibly.

The creator of SLI Compatibility Bits, a super-useful tool for SLI and NVLink users, recently spotted that NVIDIA silently added Multi-GPU Checkerboard Rendering in their latest GeForce drivers. You'll need to manually enable this feature inside of the NVIDIA Inspector Profile tool by changing some of the settings including SLI compatibility bits (for DX10, DX11, and DX12) as well as the SLI rendering mode options.

Checkerboard rendering isn't something new as it's a method of rendering that's used in some console games, including the PlayStation 4 Pro and its upgraded 4K graphics in some games. But when we're talking multi-GPU configurations the kinda doors that Checkerboard Rendering can open are extensive -- like multi-GPUs working in DX10, DX11, and DX12 titles even if the game itself doesn't directly support it.

NVIDIA told me at E3 2019 during the GeForce RTX SUPER unveiling that there would NOT be a GeForce RTX 2080 Ti SUPER, but according to the latest rumors it appears that the RTX 2080 Ti SUPER is really coming and it'll be a BEAST.

The latest rumors are coming from kopite7kimi on Twitter, who has leaked out other information on the previous RTX SUPER and GTX 16 graphics cards from NVIDIA. The new rumor pegs up some specs on the card with NVIDIA reportedly using the full Turing TU102 GPU with a mean 4608 CUDA cores, 576 tensor cores, 72 RT cores, 288 texture units, and 96 ROPs and even-faster 16Gbps GDDR6 memory.

We should expect NVIDIA to go above and beyond with the amount of GDDR6, with the non-SUPER GeForce RTX 2080 Ti having 11GB of GDDR6 the new purported RTX 2080 Ti could have 11-16GB of GDDR6. I would like to see it with 12GB minimum, but rather 16GB -- it provides a nice buffer between the insane 24GB of GDDR6 on the flagship TITAN RTX.