Well, well, well, I have something really interesting for you. Remember the Xe HP GPU leak from a couple of days back that showcased a 4-tile 500W GPU? And remember how most people assumed this is a 512 EU affair with 4096 cores? Well, I decided to ask one of my sources working at Intel and they told me the exact ratio and it's not 128 EU per tile. The distinction, it seems, lies in between Xe LP and Xe HP - and the documents leaked were about Xe HP.
Intel Arctic Sound Xe HP GPUs have 512 EUs on a single tile, 4-tile variant to have 2048 EUs and up to 16,384 cores
Let's start by setting the record straight about something. The original leak by Digital Trends assumed a number of 128 EUs per tile. The problem is they were talking about Xe LP which is not the same thing as Xe HP. While the former is designed as a low power/performance platform and internally referred to as Xe 12.1, the latter is a high power/high-performance platform internally referred to as Xe 12.5. The EU count per-tile of Xe 12.5 or Xe HP is actually different. Xe HP actually has an EU count of 512 EUs per tile.
I had stated in my article that 4096 core count seems almost certainly incorrect considering launching a 4k count GPU with a 500W power draw would be tantamount to suicide by Intel. Since Intel Xe HP (12.5) has 512 EUs per tile, the actual EU count for the 1 Tile, 2 Tile, and 4 Tile respectively are 512 EUs, 1024 EUs, and 2048 EUs! A massive 2048 EU GPU with a 500W power draw makes a lot more sense on paper. Now considering both architectures are different they might easily have different EU-to-core ratio, but if it's the same (and that is a big if) you are looking at 16,384 cores in total. Clocked at a measly 1200 Mhz, this behemoth would be able to output 39.3 TFLOPs of raw compute while drawing only 500Ws!
Here are the actual EU counts of Intel's various MCM-based Xe HP GPUs along with estimated core counts and TFLOPs:
- Intel Xe HP (12.5) 1-Tile GPU: 512 EU [Est: 4096 Cores, 12.2 TFLOPs assuming 1.5GHz, 150W]
- Intel Xe HP (12.5) 2-Tile GPU: 1024 EUs [Est: 8192 Cores, 20.48 assuming 1.25 GHz, TFLOPs, 300W]
- Intel Xe HP (12.5) 4-Tile GPU: 2048 EUs [Est: 16,384 Cores, 36 TFLOPs assuming 1.1 GHz, 400W/500W]
Intel's ATS platform is going to be the company's first real foray into the world of high-performance computing and is essentially the prototyping platform for Ponte Vecchio that's coming later down the line. Here is where things get even more interesting by the way. I have heard that as of right now, only the 1 Tile and 2 Tile variants have been accepted (remember that the slides were in Pre-POPL2 stage and were made in early 2019) while as the 4-Tile still hasn't been accepted yet.
Considering Ponte Vecchio is slated to arrive in 2021 anyways, it might actually not make sense for Intel to continue forward with the 4-Tile monstrosity (but plans can and do change with these things). In either case, the leak (and this clarification) are about a product that is pretty much in its infancy in the product life cycle and it would do well to take everything at face value because the final product will probably look very different. The one thing I can say for sure is that if Intel does decide to go with a 4-Tile ATS GPU, it won't have 512 EUs - it will have 2048.