Another day, and another leak by Videocardz.com. The website has revealed slides from Radeon Technologies Group with information regarding the company's upcoming Vega 10 flagship GPU. These were circulated internally in the last few months of 2016. This means that while the slides appear authentic, some information, marketing names for example, will be outdated. The technical specifications however, for the most part, can be easily relied upon.
AMD preparing Vega 10 x2 in 2H 2017 and 7nm Vega 20 GPU for 2018, Vega 10 will be clocked at 1500 MHz+
The slides in question detail not only the Vega 10 GPU that is expected to release in the first half of 2017, but also Vega 10 x2 and Vega 20 GPUs. We have already seen these guys in the Vega roadmap leak for 2017-2018 before. The slides contain technical specifications of the GPUs and appear to be made for the server market.
The Vega 10 GPU is made on Global Foundries 14nm node, and will feature 24 TeraFlops of 16-bit Float (now updated to 25 TeraFlops) as well as 12 TeraFlops of 32-bit Float (now updated to 12.5 TeraFlops). Double precision performance of the Vega 10 GPU is listed as 1/16th rate and is 650 GFlops. Scaling proportionally to the other specifications, we should expect this to be around ~675 GFlops now.
With a single precision compute of 12.5 TeraFLOPs per second on a GPU with 4096 cores, and considering TeraFLOPs is a function of Clock Speed * 2 Instructions Per Clock * Cores, you are looking at a Vega 10 graphics card that is clocked at roughly 1526 MHz assuming it has the same amount of NCUs. Considering the fact that the MI25 is passively cooled and also if the consumer version has less cores, it will be clocked significantly higher than the 1526 MHz mark! However, we are talking about a completely different architectural revision here and since according to AMD this will have a higher throughput per clock, comparisons on this bases alone will have to wait till more details arrive.
The specifications mention the HBM2 stack with 512 GB/s of bandwidth but this is listed as 16 GB. We already know from the Vega Doom demos that the product will have total vRAM of 8 GB so this is probably only relevant for the server market. The card will consume 225 Watts of power.
The first slide also mentions the Vega 10 x2 GPU, which will be revealed in the second half of 2017. While there is no clear indication, it does appear to be the dual GPU vega we have been hearing a lot about. The TDP and HBM2 stacks fit nicely into this equation, the only thing that sticks out like a sore thumb is the CU count (now called NCU) which should be 128.
That said, and speaking from personal experience, AMD’s internal slides usually have technical typos so I will not be surprised if the NCU number turned out to be just that. The other alternative is that this is a newer more powerful version of the die, but this is a lower probability event in my opinion. This card will use 300 Watts of power, which is comparable to dual GPUs the company has released in the past.
For the second half of 2018, AMD promises to deliver Vega 20, which will be fabricated on Global Foundries 7nm process and feature the same core count as Vega but will have twice the 16-bit Floating Point rate and a much more appetizing, ½ rate double precision along with ECC (error correcting) memory. This means this is the entry that AMD is going to pit against Nvidia’s Tesla lineup which is designed for double precision operations that also require high data integrity.
The card will have 4 stacks of HBM2 for 1 TB/s of bandwidth and will be scalable from 16 GB to 32 GB. It will have xGMI support for P2P GPU communication and will support PCIe Gen4 host connection. The TDP of the card will be 150-300 Watts, which appears to be a configurable amount depending on the exact variant you use. Finally we have the picture of the actual roadmap, which like Videocardz mentions, could have changed since the time the slides were created.
AMD Next Generation Vega 10, 11, 20 and Dual GPU Graphics Card Rumored Lineup:
| WCCFTech | Polaris 10 | Vega 11 | Vega 10 | Vega 10 Dual GPU | Vega 20 |
|---|---|---|---|---|---|
| Year | 2016 | 2017 | 2017 | 2017 | 2018 |
| Process | 14nm FinFET | 14nm FinFET | 14nm FinFET | 14nm FinFET | 7nm FinFET |
| Transistors In Billions | 5.7 | TBA | TBA | TBA | TBA |
| Stream Processors | 2304 | 2304+ (est.) | 4096 | 8192 | 4096 |
| Clock Speed | 1266 Mhz | TBA | 1526 Mhz | 1100 Mhz+ (est.) | 1800 Mhz+ (est.) |
| Performance | 5.8 TFLOPS | TBA | 12.5 TFLOPS | 19 TFLOPS - 24 TFLOPs (est.) | 15 TFLOPS+ |
| TDP | 150W | TBA | 225W | 300W | 150-300W |
| Memory | 8GB GDDR5 | TBA | 8GB/16GB HBM2 | 16-32GB HBM2 | 16-32GB HBM2 |
| Memory Bus | 256bit | TBA | 2048-bit (2 Stacks) | 4096-bit (2048-bit x2) | 4096-bit (4 Stacks) |
| PCI Express | 3.0 | TBA | 3.0 | 3.0 | 4.0 |
| Bandwidth | 256 GB/s | TBA | 512 GB/s | 1 TB/s | 1 TB/s |
