Here are some of the first pictures of the "TU104," the chip at at heart of NVIDIA GeForce RTX 2080, in its top-spec TU104-400A-A1 variant. While the package itself isn't bigger than that of its predecessors, the GP104 and GM204, the die of the TU104 is huge! If it's built on existing silicon fabrication processes such as 14 nm or 12 nm, the TU104 could end up with transistor-counts higher than those of even the GP102.

You'll recall that we posted bare-PCB pictures of this board recently. In its production (or close-to-production) variant, the board draws power from a combination of 6-pin and 8-pin PCIe connectors, conditioning it with an 8+2 phase VRM. The GPU is surrounded by eight GDDR6 memory chips. Display connectivity includes three DisplayPorts, an HDMI, and a USB type-C (VirtuaLink).

Later today (20th August), NVIDIA will formally unveil its GeForce RTX 2000 series consumer graphics cards. This marks a major change in the brand name, triggered with the introduction of the new RT Cores, specialized components that accelerate real-time ray-tracing, a task too taxing on conventional CUDA cores. Ray-tracing and DNN acceleration requires SIMD components to crunch 4x4x4 matrix multiplication, which is what RT cores (and tensor cores) specialize at. The chips still have CUDA cores for everything else. This generation also debuts the new GDDR6 memory standard, although unlike GeForce "Pascal," the new GeForce "Turing" won't see a doubling in memory sizes.

NVIDIA is expected to debut the generation with the new GeForce RTX 2080 later today, with market availability by end of Month. Going by older rumors, the company could launch the lower RTX 2070 and higher RTX 2080+ by late-September, and the mid-range RTX 2060 series in October. Apparently the high-end RTX 2080 Ti could come out sooner than expected, given that VideoCardz already has some of its specifications in hand. Not a lot is known about how "Turing" compares with "Volta" in performance, but given that the TITAN V comes with tensor cores that can [in theory] be re-purposed as RT cores; it could continue on as NVIDIA's halo SKU for the client-segment.

Here are some of the first pictures of the bare printed circuit board (PCB) of NVIDIA's upcoming GeForce GTX 1180 graphics card (dubbed PG180), referred to by the person who originally posted them as "GTX 2080" (it seems the jury is still out on the nomenclature). The PCB looks hot from the press, with its SMT points and vias still exposed. The GT104 GPU traces hint at a package that's about the size of a GP104 or its precessors. It's wired to eight memory chips on three sides, confirming a 256-bit wide memory bus. Display outputs appear flexible, for either 2x DisplayPort + 2x HDMI, or 3x DisplayPort + 1x HDMI configurations.

The VRM setup is surprisingly powerful for a card that's supposed to succeed the ~180W GeForce GTX 1080, which can make do with a single 8-pin PCIe power input. The card draws power from a combination of 6-pin and 8-pin PCIe power connectors. There is a purportedly 10-phase VCore side, which in all likelihood is a 5-phase setup with "dumb" phase-doubling; and similarly, a 2-phase memory power (which could again be doubled single-phase). The SLI-HB fingers also make way. There's a new connector that looks like a single SLI finger and an NVLink finger arranged side-by-side. NVIDIA still hasn't given up on multi-GPU. NVLink is a very broad interconnect, in terms of bandwidth. NVIDIA probably needs that for multi-GPU setups to work with not just high resolutions (4K, 5K, or even 8K), but also higher bit-depth, higher refresh-rates, HDR, and other exotic data. The reverse side doesn't have much action other than traces for the VRM controllers, phase doublers, and an unusually large bank of SMT capacitors (the kind seen on AMD PCBs with MCM GPUs).

A Vietnamese online store put up the first listing of a GeForce GTX 1180 based ASUS ROG Strix graphics card. The store even put out some specifications of the card, beginning with it being based on the "GT104" silicon, based on the "Turing" series. With "Turing" NVIDIA appears to be forking its GPU architectures on the basis of chips that feature DPFP (double-precision floating point) cores and Tensor cores, and those that lack both (and only feature SPFP cores). "Turing" is probably a fork of "Volta" that lacks both DPFP CUDA cores and Tensor cores; and sticks to the cheaper GDDR6 memory architecture, while "Volta" based GPUs, such as the TITAN V, implement pricier HBM2 memory.

Among the specifications of the GeForce GTX 1180 are 3,584 CUDA cores, and 16 GB of GDDR6 memory across a 256-bit wide memory interface. The memory is clocked at 14 GHz (GDDR6-effective), which works out to 409.6 GB/s of memory bandwidth. Pre-launch prices, just like most specifications, tend to be bovine excrement, which in this case converts to a little over USD $1,500, and isn't really relevant. What is, however, interesting is the availability date of September 28.