⋮    ⋮  

AMD Vega and Navi GPU Roadmap Detailed – Vega 10, Vega 11 in 2017, Vega 20 in 2018 and Navi Shipping in 2019 To Servers


AMD's major server roadmap, detailing their next generation Vega and Navi GPUs has been leaked by Videocardz. The roadmap that spans up to 2019 shows how AMD will configure their next-gen GPUs into different server segments that include applications like HPC, Deep Learning and Inferencing.

AMD Vega and Navi GPU Roadmap For Servers Leaked - Vega Arrives in 2017, Navi in 2019

The latest roadmap leak comes just a day after major specifications were also revealed by Videocardz. 2016 started off with AMD replacing most of their graphics lineup with Polaris based cards. With the exception of high-end products, the Polaris 10 and Polaris 11 cards based on FinFET architecture delivered good improvements across the board. Whether that be in terms of efficiency or performance but the most stellar selling point for these cards was their competitive pricing.

Moving on, AMD is expected to launch new high-end, enthusiast products in the market. But the products we are looking at today are not aimed at consumers or gaming enthusiasts but rather, the server market. So let's take a look what 2017 and beyond has to offer in terms of server aimed products from AMD.

AMD Radeon Pro Series To Get Radeon Pro S9 Nano

The AMD Radeon Pro series was introduced two months back and featured a range of Polaris 10, Polaris 11 and Fiji based cards. The main attractions were the Radeon Pro WX 7100 and Radeon Pro SSG. The former is the most affordable work-station card while the latter is a heavy-duty, Fiji based card that comes with a on-board SSD solution to deliver immense boost in productivity and performance on workstation suites. More on these two cards here.

AMD plans to introduce a new server-focused card in this lineup too. Based on their Fiji GPU, the card would be known as Radeon Pro S9 Nano. The "WX" stands for workstation while "S" stands for server. The solution is expected to feature the same specifications as the Radeon R9 Nano along with some clock adjustments. The new launch could happen in a matter of weeks as mentioned in the leak. Radeon Pro S9 Nano will aim the inference market and can be sighted as competition against NVIDIA's recently launched Tesla P40 solution for inferencing.

The R9 Nano, although expensive, was an impressive product with insane specs in a SFF solution.

Of course, the Radeon Pro S9 Nano won't be the only inferencing card available. AMD will also launch cheaper cards based on Polaris 10 and Polaris 11 GPUs. These will go in nicely against the Tesla P4 solution with possibly affordable price bracket.

AMD Vega 10, Vega 11, Vega 10 x2 and Vega 20 In 2017 and 2018

In 2017, AMD finally plans to upgrade all of their GPUs to the new FinFET architecture. While Polaris serves the mainstream, Vega will be aiming for the high-end market. The new graphics chip for Deep Learning will be the Vega 10, a very beefy GPU which we detailed yesterday. AMD Vega 10 is expected to launch in first quarter of 2017.

From the leaked details, we now know that Vega 10 can feature up to 4096 stream processors based on a new GCN uARCH (GFX9). The chip has as much as 24 TFLOPs of half-precision and 12 TFLOPs of single precision compute performance. The chip also features up to 16 GB of HBM2 that is clocked to operate at 512 GB/s. The cards based on this chip are rated with 225W TDP.

AMD's Vega lineup will feature immensely powerful cards with HBM2 memory in 2017. Original Image by FallenZeraphine

Now we noticed that Vega 10 isn't aiming the HPC market but instead, focused at the Deep Learning servers. It will have two parts available in 2017, a single chip and dual chip solution. Both feature double precision at 1/16th the rate of FP32. It confirms that Vega 10 is similar to GP102 which cuts down double precision compute from the more beefy GP100 GPU. The details on this will be covered in a more informative article later on.

AMD Vega 20 Will Be Their First FinFET Based HPC Chip - Launching in 2H 2018

This shows that Hawaii will still stand as the HPC muscle up till 2018. A new GPU will be developed by AMD in the meanwhile known as Vega 20 with DP rated at 1/2 of SP. This would be a solution targeted at GP100 or the new GV100 chip based on Volta which is expected to be introduced by then.

The latter is more possible since Vega 20 would not only use a enhanced built design but also come packaged in the new GloFo 7nm FF process. Global Foundries has already announced risk production of 7nm FF in 2018 so we might just get to see Vega 20 in some form in that year. With up to 32 GB of HBM2 memory (1 TB/s) and a fast 64 compute unit design, this chip will have impressive double precision capabilities for HPC platforms.

Also, do you remember that we mentioned a high-performance interconnect being developed by AMD a while ago? Looks like the solution will be known as xGMI. This peer-to-peer solution will be embedded in HPC servers that include Vega 20 and Zen based Naples processors.

AMD Next Generation Vega 10, 11, 20 and Dual GPU Graphics Card Rumored Lineup:

WCCFTechPolaris 10Vega 11Vega 10Vega 10 Dual GPUVega 20
Process14nm FinFET14nm FinFET14nm FinFET14nm FinFET7nm FinFET
Transistors In Billions5.7TBATBATBATBA
Stream Processors23042304+ (est.)409681924096
Clock Speed1266 MhzTBA1526 Mhz 1100 Mhz+ (est.)1800 Mhz+ (est.)
Performance5.8 TFLOPSTBA12.5 TFLOPS19 TFLOPS - 24 TFLOPs (est.)15 TFLOPS+
Memory8GB GDDR5TBA8GB/16GB HBM216-32GB HBM216-32GB HBM2
Memory Bus256bitTBA2048-bit (2 Stacks)4096-bit (2048-bit x2)4096-bit (4 Stacks)
PCI Express3.0TBA3.03.0 4.0
Bandwidth256 GB/sTBA512 GB/s1 TB/s1 TB/s

AMD Navi 10, Navi 11, Navi 10 x2 in 2019, Navi 20 in 2020?

Moving forward, AMD has plans to ship Navi in 2019. Now you might believe that Navi has been delayed but we don't find that as the case. This roadmap is specific to the server platform and these products have the tendency to arrive later as compared to consumer parts. Hence, Navi could still be a planned 2018 launch on the consumer side but could miss that time frame on the server side of things.

Next-Gen GPUs will leverage from more advanced memory solutions such as HBM3.

Nevertheless, Navi 10 and dual Navi cards are expected to ship in 2019. We can expect these cards to leverage from new GCN architecture, one that is revised extensively for greater efficiency leap over Vega. There's also next-gen memory (HBM3) and PCI-e 4.0 in the talks. Navi 10 will be aiming the Deep Learning market while the smaller Navi 11 core will be aimed at the Inference market.

Just like Vega, Navi won't see a HPC solution for a while. The Vega 20 GPU will be remain at the helm of AMD's HPC offerings and eventually be replaced by a powerful double Precision part in the form of Navi 20. This solution can hit the market around 2020. That's it as far as AMD's server roadmap is concerned, we will cover some specifications comparison between NVIDIA Pascal/Volta and AMD Vega/Navi GPUs in the days ahead to assess their potential in the HPC markets.

AMD Server GPU Roadmap 2016-2019

Product Segement
HPCHawaiiHawaiiVega 20Vega 20
Deep LearningPolaris 10
Polaris 11
Vega 10
Vega 10 x2
Vega 10
Vega 10 x2
Navi 10
Navi 10 x2
InferenceRadeon Pro S9 NanoVega 11Vega 11Navi 11