Brand new details regarding the Intel ARC Alchemist lineup of gaming graphics cards have been leaked by Moore's Law is Dead. The new leaks cover the performance and positioning of various GPU configurations within the ARC lineup that is launching in Q1 2022.
Intel ARC Alchemist Gaming Graphics Card Rumors: Three Xe-HPG GPU Configurations Aiming High-End & Entry-Level Segments
According to the rumor, Intel will have at least three configurations of ARC Alchemist GPUs ready for launch in Q1 2022. These will include two configurations based on the top 512 EU die and one configuration based on the 128 EU die. Although there are more GPU configs that we have seen in leaks, it looks like those may be used in future products though that cannot be confirmed. So let's start with the top-end configuration.
MLID has leaked the latest information regarding Intel's ARC Alchemist gaming graphics cards & respective GPU configs.
Intel Xe-HPG 512 EU ARC Alchemist Graphics Card
The top Alchemist 512 EU variant has just one configuration listed so far and that utilizes the full die with 4096 cores, 256-bit bus interface, and up to 16 GB GDDR6 memory featuring a 16 Gbps clock though 18 Gbps cannot be ruled out as per the rumor.
The Alchemist 512 EU chip is expected to measure at around 396mm2 which makes it bigger than the AMD RDNA 2 and NVIDIA Ampere offerings. The Alchemist -512 GPU will come in the BGA-2660 package which measures 37.5mm x 43mm. NVIDIA's Ampere GA104 measures 392mm2 which means that the flagship Alchemist chip is comparable in size while the Navi 22 GPU measures 336mm2 or around 60mm2 less. This isn't the final die size of the chip but it should be very close.
NVIDIA packs in tensor cores and much bigger RT/FP32 cores in its chips while AMD RDNA 2 chips pack a single ray accelerator unit per CU and Infinity Cache. Intel will also have dedicated hardware onboard its Alchemist GPUs for Raytracing & AI-assisted super-sampling tech.
The Xe-HPG Alchemist 512 EU chip is suggested to feature clocks of around 2.2 - 2.5 GHz though we don't know if these are the average clocks or the maximum boost clocks. Let's assume that it's the max clock speed and in that case, the card would deliver up to 18.5 TFLOPs FP32 compute which is 40% more than the RX 6700 XT but 9% lower than the NVIDIA RTX 3070.
Expected performance and FLOPs of Intel's ARC Alchemist top GPU die compared to RTX 3070 & RX 6700 XT. (Image Source: MLID)
In a speculative measure of performance, MLID states that the TFLOPs make no sense for comparison as performance scales differently respective to the architecture, not the FLOPs performance. The gaming graphics card is pretty much expected to be faster than the RX 6700 XT & RTX 3070 at this point but with work ongoing on the driver suite, the performance is expected to improve further.
Also, it is stated that Intel's initial TDP target was 225-250W but that's been upped to around 275W now. We can expect a 300W variant with dual 8-pin connectors too if Intel wants to push its clocks even further. In either case, we can expect the final model to rock an 8+6 pin connector config, The reference model is also going to look very much like the drone marketing shot Intel put out during the ARC branding reveal. That reference design was leaked a while back by MLID too. There're also talks about a custom lineup being worked upon by Intel's AIB partners.
Intel ARC Alchemist vs NVIDIA GA104 & AMD Navi 22 GPUs
| GPU Name | Alchemist DG-512 | NVIDIA GA104 | AMD Navi 22 |
|---|---|---|---|
| Architecture | Xe-HPG | Ampere | RDNA 2 |
| Process Node | TSMC 6nm | Samsung 8nm | TSMC 7nm |
| Flagship Product | ARC (TBA) | GeForce RTX 3070 Ti | Radeon RX 6700 XT |
| Raster Engine | 8 | 6 | 2 |
| FP32 Cores | 32 Xe Cores | 48 SM Units | 40 Compute Units |
| FP32 Units | 4096 | 6144 | 2560 |
| FP32 Compute | ~16 TFLOPs | 21.7 TFLOPs | 12.4 TFLOPs |
| TMUs | 256 | 192 | 160 |
| ROPs | 128 | 96 | 64 |
| RT Cores | 32 RT Units | 48 RT Cores (V2) | 40 RA Units |
| Tensor Cores | 512 XMX Cores | 192 Tensor Cores (V3) | N/A |
| Tensor Compute | ~131 TFLOPs FP16 ~262 TOPs INT8 |
87 TFLOPs FP16 174 TOPs INT8 |
25 TFLOPs FP16 50 TOPs INT8 |
| L2 Cache | TBA | 4 MB | 3 MB |
| Additional Cache | 16 MB Smart Cache? | N/A | 96 MB Infinity Cache |
| Memory Bus | 256-bit | 256-bit | 192-bit |
| Memory Capacity | 16 GB GDDR6 | 8 GB GDDR6X | 16 GB GDDR6 |
| Launch | Q1 2022 | Q2 2021 | Q1 2021 |
Intel Xe-HPG 384 EU ARC Alchemist Graphics Card
Moving on, we have the Intel Xe-HPG Alchemist 384 GPU which will be a cut-down variant of the 512 EU die. The 384 EU chip will feature 3072 cores, up to 12 GB GDDR6 memory, and a 192-bit bus interface though Intel could tune it up to 256-bit bus feature an 8 GB GDDR6 memory capacity. The card would also feature a TDP of around 200W and is expected to feature performance NVIDIA RTX 3060 and RTX 3060 Ti.
The 384 EU GPU is expected to feature a 16 MB smart cache, respectively. Since the GPU is pretty much a cut-down model, we will get to see clocks over the 2 GHz mark. Now Intel has been testing out various cut-down configurations of its top-die including a 448 EU and 256 EU variant. We don't know if there are any gaming graphics cards planned to release with those specific configs but Intel has the option to release even more SKUs if they see the market for it.
Intel Xe-HPG 128 EU ARC Alchemist Graphics Card
Then lastly, we have the Intel Xe-HPG Alchemist 128 EU parts. The top config is once again a full-fat SKU with 1024 cores, a 64-bit bus interface, and up to 8 GB GDDR6 memory. The cut-down variant will come with 96 EUs or 768 cores and a 4 GB GDDR6 memory featured across a 64-bit bus interface. The chip will also feature a clock speed of around 2.2 - 2.5 GHz and have a sub 75W power consumption which means we will be looking at connector-less graphics cards for the entry-level segment.
Performance is expected to land between the GeForce GTX 1650 and GTX 1650 SUPER but with raytracing capabilities. One big advantage that Intel could have over AMD and Intel is that with these cards, they might enter the sub-$250 US market which has been completely abandoned in the current generation of cards. The GeForce RTX 3050 series only got a laptop release so far with RTX 3060 serving the entry-level Ampere segment at $329 US while the RX 6600 is expected to be AMD's entry-level solution for around $300 US.
This GPU will be very similar to the DG1 GPU-based discrete SDV board however Alchemist will have a more improved architecture design and definitely more performance uplift over the first-gen Xe GPU architecture. This lineup is definitely going to be aimed at the entry-level desktop discrete market based on the specifications.
Intel Xe-HPG Based Discrete Alchemist GPU Configurations:
| GPU Variant | GPU Die | Execution Units | Shading Units (Cores) | Memory Capacity | Memory Bus | TGP |
|---|---|---|---|---|---|---|
| Xe-HPG 512EU | Alchemist-512EU | 512 EUs | 4096 | 16/8 GB GDDR6 | 256-bit | 225-275W? |
| Xe-HPG 384EU | Alchemist-512EU | 384 EUs | 3072 | 12/6 GB GDDR6 | 192-bit | 225-275W? |
| Xe-HPG 256EU | Alchemist-512EU | 256 EUs | 2048 | 8/4 GB GDDR6 | 128-bit | 150-200W? |
| Xe-HPG 192EU | Alchemist-512EU | 192 EUs | 1536 | 4 GB GDDR6 | 128-bit | 150-200W? |
| Xe-HPG 128EU | Alchemist-128EU | 128 EUs | 1024 | 4 GB GDDR6 | 64-bit | |
| Xe-HPG 96EU | Alchemist-128EU | 86 EUs | 768 | 4 GB GDDR6 | 64-bit |
Based on the timeline, the Xe-HPG Alchemist lineup will compete against NVIDIA's Ampere & AMD RDNA 2 GPUs since both companies aren't expected to launch their next-gen parts by the very end of 2022. NVIDIA and AMD are expected to release refreshes in early 2022 so that might give Intel's new lineup some competition but based on current performance expectations, the refreshed may not bring drastic performance differences to the lineup. The Xe-HPG ARC GPUs will also be coming to the mobility platform too and will be featured in Alder Lake-P notebooks.
https://ift.tt/3zD0cKl
Technology
No comments:
Post a Comment