AMD Vega to launch with slower HBM2, faster HBM2 later

AMD is expected to launch its Vega graphics cards in the next couple of months, with an unveiling in May and a launch in June to be expected - but now SK Hynix have updated their memory product catalog, giving us a better look at the bandwidth numbers around HBM2.

SK Hynix updated their graphics databook, showing off HBM2 for Q4 2016 with a 4-stack 4GB HBM2 product with two different speeds: 2Gbps and 1.6Gbps, with 256GB/sec bandwidth and 204GB/sec bandwidth, respectively. This has now changed, with a shift in availability.

The company is now listing the slower 1.6Gbps (204.8GB/sec) for Q1 2017, while the faster 2Gbps (256GB/sec) HBM2 now stuck with a 'TBD' for availability. This could lead AMD to launch a slower HBM2-based card in Q2, and a faster Vega graphics card later in the year.

AMD should launch a performance Vega 10-based graphics card with 8GB of HBM2, and if SK Hynix's new specifications are correct - we could be looking at 409.6GB/sec of memory bandwidth. This is lower than the GDDR5X-based Titan X from NVIDIA with its 480GB/sec memory bandwidth, and slower than AMD's older Radeon R9 Fury X which has HBM1 and 512GB/sec bandwidth (but was limited to just 4GB).

We could see AMD using GDDR5 and/or GDDR5X with Vega, keeping the HBM2-based models at a higher price, because yields are also going to be an issue, just like the Fiji-based Fury graphics cards in 2015.

Remember, that AMD's next-gen Vega GPU architecture includes the impressive sounding - but still mystical and mostly unknown (performance benefit wise) HBC (High Bandwidth Cache). So the lowered memory bandwidth numbers might not mean much if HBC can pick up the slack in a big way, and we still don't know much about the Vega NCU itself - which will have its own surprises in store, I'm sure.

I reached out to my good friend and awesome analyst Anshel Sag, Associate Analyst at Moor Insights & Strategy, regarding the news of SK Hynix's moves on HBM2. He said: "I believe that HBM2 is going to allow AMD to close the gap between themselves and NVIDIA. However we stop don't know what performance profiles AMD is looking at and what priced they will hit. Additionally, we still don't know if HBM yields will be an issue again".