Tuesday, February 22nd 2022
Intel Arc Alchemist Graphics Card Lineup Detailed
SiSoftware put out the mother lode of information on Intel's upcoming Arc "Alchemist" gaming graphics card series, along with OpenCL compute performance of the entry-level Arc A380. The Arc series model numbering is "A" (Alchemist) followed by a number series. The A300 series makes up the entry-mainstream; the A500 series makes up the mid-performance segment; and the A700 series leads the pack with high-end SKUs. The "Alchemist" GPUs are built on the 7 nm silicon fabrication node at TSMC, the N7.
The A300 series is based on the smaller "Alchemist" series dies, with 128 EUs (execution units), which work out to 1,024 programmable shaders. The A500 series and A700 series appear to be carved out from the larger silicon. The A500 series has roughly 384 EU or 3,072 shaders. The top-dog A700 series has all 512 EU or 4,096 shaders enabled. Intel is tapping into industry-standard GDDR6 for dedicated graphics memory. The A300-series SKUs typically have 6 GB of 14 Gbps-rated memory across a 96-bit wide memory bus, for 192 GB/s of bandwidth. The A500 series parts have 12 GB of 16 Gbps-rated memory across a 192-bit bus, for 384 GB/s of bandwidth. The top A700 series maxes out the 256-bit memory bus with 16 GB of memory at 16 Gbps data-rate, for 512 GB/s bandwidth.
Several other interesting features were revealed. Firstly, "Alchemist" supports real-time ray tracing, and meets the full DirectX 12 Ultimate specification. At this point it's not known just how much fixed-function hardware is used to achieve this. The GPU has XMX (hardware for matrix-multiplication). A note on the SIMD components: Xe HPG, or "Alchemist," lacks double-precision floating-point (FP64) acceleration. It only supports FP16 and FP32 (which is all you need for gaming and consumer apps).
SiSoftware tested an entry-level Arc A380 graphics card (1,024 shaders) through its SANDRA GPGPU benchmark that uses native OpenCL code. Here, the A380 was found to be offering performance comparable to a GeForce GTX 1660 Ti "Turing," and hot on the heels of the RTX 3050 "Ampere" and RX 6500 XT RDNA2, in the Mandelbrot half-precision and floating-point tests.
Intel is looking to debut the Arc Alchemist series within the first half of 2022.
Source:
SiSoft
The A300 series is based on the smaller "Alchemist" series dies, with 128 EUs (execution units), which work out to 1,024 programmable shaders. The A500 series and A700 series appear to be carved out from the larger silicon. The A500 series has roughly 384 EU or 3,072 shaders. The top-dog A700 series has all 512 EU or 4,096 shaders enabled. Intel is tapping into industry-standard GDDR6 for dedicated graphics memory. The A300-series SKUs typically have 6 GB of 14 Gbps-rated memory across a 96-bit wide memory bus, for 192 GB/s of bandwidth. The A500 series parts have 12 GB of 16 Gbps-rated memory across a 192-bit bus, for 384 GB/s of bandwidth. The top A700 series maxes out the 256-bit memory bus with 16 GB of memory at 16 Gbps data-rate, for 512 GB/s bandwidth.
SiSoftware tested an entry-level Arc A380 graphics card (1,024 shaders) through its SANDRA GPGPU benchmark that uses native OpenCL code. Here, the A380 was found to be offering performance comparable to a GeForce GTX 1660 Ti "Turing," and hot on the heels of the RTX 3050 "Ampere" and RX 6500 XT RDNA2, in the Mandelbrot half-precision and floating-point tests.
Intel is looking to debut the Arc Alchemist series within the first half of 2022.
37 Comments on Intel Arc Alchemist Graphics Card Lineup Detailed
A decent product can only be good news for the consumer...especially if their deep pockets can undercut AMD and Nvidia by a margin.
I'm looking forward to the mayhem....
cdn.wccftech.com/wp-content/uploads/2018/06/Micron-GDDR6-Memory_4-Custom-740x416.png
On Windows its a bit different but a change Nvidia have the best and Intel is on 2nd.
AMD is out of anything, the drivers on Linux and on windows are garbage,
maybe the steamdeck will push them a little bit but i dont think so, AMD will only push the drivers for the IGP.
www.guru3d.com/articles-pages/geforce-gtx-1050-3gb-review,3.html
and finally for 160bit only cut down Nvidia GP106 (GDDR5) but again it could just be 192bit with mixed capacities (I couldn't find any 5GB PCB photo) Of course my impression could be wrong since Nvidia chips is like a black box and maybe since pascal generation they had a way to disable one memory 32bit channel somehow, but in any case, we are talking for only 2 Nvidia models for all the GDDR era , so not exactly plenty....
My guess is looking at die shots the memory phys are usually on the sides and symmetry wise it makes sense to make the sides similar-ish - like having 2 phys on top+ 2 bottom for 128, or 3 + 3 + 2 on the side for 256 for example.
Also, I think you're taking this as chip capacity for the intel 96bits, when it's likely only a specific model. Previous "leaks" all talk about 2 chip variants, one with 512EU and 256bits bus and one with 128EU and 128bits bus, both chip will give X card variants (dunno what intel will do, but looking at CPUs I think they'll segment the heck out of it) - both 96 and 192 will be cut down variants like nvidia and amd do.
Especially interesting is the fact that it can hit 2.45GHz with just 75W power consumption, this bodes very well for alchemist architecture:
www.tomshardware.com/news/intel-arc-gpu-lineup-specs-revealed
If one cant get real game tests, its just paper card.
If you cant buy it, its paper card.
Im expecting to be seriously disappointed by performance of that thing, cause in such state they can only surprise me positively. Thats if we live long enough to actually see it.