AMD Vega 10 GPU Launching In 2017 – Features 4096 Cores & 16GB HBM2, Greenland Reincarnated
AMD’s highly anticipated Vega 10 flagship 14nm FinFET GPU, previously known as Greenland, is launching in 2017 with up to 32 GB of HBM2. AMD unveiled its long-term GPU roadmap at Capsaicin yesterday for the first time, debuting its 2017 and 2018 architectures Vega & Navi. This is the very first time the company has ever discussed its graphics roadmap beyond Polaris, its extremely power efficient 14nm graphics architecture launching this summer. Raja Koduri, AMD’s Chief Architect of the Radeon Technolgies Group, revealed that the company is working on two GPU architectures to succeed Polaris.
The first is Vega. Which will be arriving next year and is the first from AMD to support second generation High bandwidth Memory. In 2018 AMD plans to introduce yet another graphics arachitecture, code named “Navi”. With Navi Raja & the Radeon Technologies Group is planning to utilize vertically stacked HBM memory throughout its graphics product lineup. This is what’s being referred to with “scalability”.
By 2018 we’re potentially looking at third generation HBM, which hasn’t been announced nor named yet. Whether it’s HBM3 or something slightly or completely different, it’s too early to tell. else no one can really tell However, what we know for sure is that whatever memory standard comes after HBM2 will still employ vertical stacking in some form or fashion.
AMD Vega 10 / Greenland Flagship HBM2 GPU Specs
The specs for Vega 10, previously known as Greenland, have been spotted on the LinkedIn profile of an AMD engineer. The GPU will feature 4096 GCN 4.0 cores, second generation HBM and a 4096bit memory interface.
Which sounds identical to the R9 Fury X but are far from it. The 4th generation GCN cores are ~50-55% faster than their predecessors and extraordinarily more power efficient according to several patents filed by the company that we’ll share in an upcoming piece. Considering that the 14LPP process clocks considerably higher than 28nm we’re looking at upwards of 60-70% performance improvement with Vega 10/ Greenland over the R9 Fury X.
Vega 10 Specs
|GPU||Polaris 10||Vega 10||GP102||GP104||GP102|
|Graphics Card||RX 480||TBA||GTX TItan X Pascal||GTX 1080||GTX 1080 Ti|
|Transistors In Billions||5.7||TBA||12||7.2||12|
|Performance||5.8 TFLOPS||12 TFLOPS||11 TFLOPS||9 TFLOPS||10 TFLOPS|
|Memory||8GB GDDR5||16GB HBM2||12GB GDDR5X||8GB GDDR5X||12GB GDDR5X|
|Bandwidth||256 GB/s||512 GB/s||480 GB/s||320 GB/s||480 GB/s|
AMD Vega 10 / Greenland Flagship HBM2 GPU Launching In 2017
Graphics generation code names aside, perhaps the most important piece of information that we got out of this is as follows : AMD’s roadmap confirms that Vega will be the company’s first architecture to feature second generation HBM. In other words, it confirms that Polaris will in all likelihood make use of GDDR5/X memory instead.
If you’ve been paying close attention to the industry you will quickly point out that this makes sense in so many ways. Firstly there’s the cost advantage of using GDDR5/X over HBM. GDDR5/X chips area cheaper to make because they don’t have to go through a meticulous process to stack them on top of each other. As a GPU maker you also don’t need an interposer with GDDR5/X like you do with HBM. Both of these factors give GDDR5/X a very real cost advantage. Secondly there’s the question of HBM2’s readiness. HBM is still ramping and it would take sometime to for HBM2 to catch up.
Thirdly, there’s Raja’s revelations last year that AMD has two 14nm GPUs coming out this year, not three, two.
Patrick Moorhead, Semiconductor Financial Analyst – Forbes
Raja also talked about how Advanced Micro Devices’ RTG will need to execute on their architectural designs and create brand new GPUs, something that Advanced Micro Devices has struggled with lately. He promised two brand new GPUs in 2016
That is Polaris 10 and Polaris 11. Both of which AMD has actually shown to journalists, we’re talking about the actual physical dies. Those who have seen them – we’ve only seen Polaris 11, but AMD has shown a Polaris 10 die to visitors of its suite at CES – reported that neither of the dies sported an interposer or HBM like Fiji.
We later confirmed with AMD that indeed the Polaris 10 and Polaris 11 GPUs showcased were configured for GDDR5.
AMD helped lead the development of HBM, was the first to bring HBM to market in GPUs, and plans to implement HBM/HBM2 in future graphics solutions.
At this time we have only publicly demonstrated a GDDR5 configuration of the Polaris architecture.It’s important to understand that HBM isn’t (currently) suitable for all GPU segments due to the current HBM cost structure. In the mainstream GPU segment, GDDR5 remains an extremely cost-effective, efficient and viable memory technology.
This doesn’t necessarily exclude that either of these chips can end up with HBM. After-all AMD might have been trying to be really coy and just showed off GDDR5 configured die. But our money is that they weren’t trying to be coy and that both of these chips will end up GDDR5/X. For the simple fact that any GPU would have to be designed from the very beginning with one memory standard in mind. As a designer you either choose to equip your GPU with an HBM memory controller or a GDDR5/X memory controller. Not both, that would be utterly wasteful to the extremely limited chip area you have.
Designing a GPU with both types of memory technologies is theoretically possible. But as the GPU will only end up with one memory standard or the other, either HBM or GDDR5/X, not both at the same time. One of the two memory controllers will repetitively end up as useless as white space on a die. So it just doesn’t make any sense to design any such product.
With all of that said, this wasn’t actually unexpected. AMD has actually hinted to us back in December at Sonoma, when we were first introduced to Polaris, that there are indeed plans for HBM2 in the future. Yesterday those plans have been made public. HBM2 once again ,just as first generation HBM did, will debut at the high-end. Nvidia likely has very similar plans to AMD for 2016 and 2017. And that is to launch two Pascal GDDR5/X GPUs and introduce GP100 with HBM2 next year.
AMD’s 2.5X Polaris Power Efficiency Jump Is More Impressive Than You Think
This is perhaps one of the less expected realizations to come to after yesterday’s Capsaicin event. A 2.5X – 250% – power efficiency jump is impressive in its own right. But actually, it’s even more so when you consider the fact that Polaris achieves this without HBM.
The power savings from HBM were the primary driver behind Fiji’s considerable power efficiency gains over Hawaii and the R9 290X. In fact HBM played a pivotal role in allowing AMD to design a GPU that not only had 45% more GCN cores than the R9 290X – 4096 vs 2816 – but, also burned less power on average. According to AMD, Polaris GPUs will deliver 2.5x the perf/watt of AMD’s 2015 28nm GPUs. Those include the Radeon 300 series and the R9 Fury series.
It’s quite difficult to pin-point exactly how far ahead each architecture is of the one before it on AMD’s roadmap. But Navi looks to be sitting at double Polaris’s power efficiency, making it 5x as power efficient as the current lineup of 28nm GPUs. Vega sits roughly in the middle between Polaris and Navi, slightly closer to Polaris. So we might be looking at a 3X improvement over 28nm GPUs, which puts it 20% ahead of Polaris in perf/watt. That’s about what you’d expect from HBM2.
But does this mean that AMD’s fabled Vega 10 / Greenland flagship GPU will be three times more power efficient than the Fury X. I’d say possible, but more likely we’re looking at a 3X efficiency jump over the R9 390X. In case you missed our exclusive report a couple of months back , Vega 10 is Greenland. The ~18 billion transistor 32GB HBM2 behemoth that’s been the subject of many leaks over the past year.
AMD Radeon Graphics Architectures
|GPU||Process||Transistors In Billions||Performance||Memory||Bandwidth|
|Southern Islands||2012||HD 7970 Ghz||Tahiti||28nm||4.3||4.1 TFLOPS||3GB GDDR5||264GB/s|
|Volcanic Islands||2013||R9 290X||Hawaii||28nm||6.2||5.6 TFLOPS||4GB GDDR5||320GB/s|
|Caribbean Islands||2015||R9 Fury X||Fiji||28nm||8.9||8.6 TFLOPS||4GB HBM1||512GB/s|
|POLARIS||2016||RX 480||Polaris 10||14nm||5.7||5.2 TFLOPS||8GB GDDR5||256 GB/s|
|VEGA||2017||RX Vega 64||Vega 10||14nm||12.5||13 TFLOPS||8GB HBM2||480GB/s|
|NAVI||2019||TBA||Navi 10||7nm||TBA||TBA||Nextgen Memory||TBA|
No HBM2 This Year?
It’s true that we will have to wait until next year for HBM2 GPUs from either Nvidia or AMD. Nvidia’s GP104 powered GTX 1080 and GTX 1070 cards – not official names – are reportedly configured for GDDR5X. And there’s virtually no chance of neither Nvidia or AMD risking production of large GPUs like GP100 or Vega 10 this early in the nodes’ lifecycles. Neither 16nm or 14nm will be mature enough for large GPU launches this year. This is all assuming HBM2 production has already sufficiently ramped up, which hasn’t yet.
Despite that, what we’re getting this year is still going to deliver a very significant jump in performance, performance per watt and performance per dollar. This is thanks to the much needed arrival of FinFET manufacturing technology as well as GDDR5X.
This update to the GDDR5 memory spec is simple enough. Double the capacity and double the data rate of GDDR5 with minimal alterations to the protocol. That in turn means existing GDDR5 memory controllers will only need minor design updates to be compatible with GDDR5X. Which saves Nvidia and AMD a lot of engineering effort and cost. Double the bandwidth also means GDDR5X will do more than an adequate job of keeping faster, bigger, next generation FinFET GPUs happy and well fed until HBM2 arrives in 2017.
At double the speed of GDDR5, a GPU configured with a 256bit memory interface can have access to up to 448GB/s of bandwidth and 8GB of memory. That’s double the bandwidth available to the GTX 980 and twice the capacity. It’s also 33% more bandwidth than what’s available to the GTX Titan X. With that in perspective GDDR5X is clearly going to be more than enough for Nvidia’s GP104 GPU and AMD’s Polaris 10 GPU.
It’s not every day that we get GPUs with 2.5x better perf/watt, in fact not ever. This is the biggest jump from one generation to the next that we have ever seen. But it’s not all about efficiency, Raja is also promising the “most revolutionary jump in performance so far”.
AMD announced that Polaris graphics cards are going to be released mid 2016, so this summer. AMD’s CEO Lisa Su then elaborated further, saying that Polaris graphics cards are to be expected on shelves and in notebooks before the back to school season this year. There’s also been a lot of buzz around Nvidia potentially demoing Pascal at GTC in April, with GTX 1080 and GTX 1070 – not official names – graphics cards launching in the summer. AMD has already demoed Polaris 11 at CES and Polaris 10 yesterday at GDC, so we can’t wait to see Pascal in action. Hopefully next month.