There’s a new set of rumors cropping up about AMD’s Vega. Last month, we covered an AMD press slide that stated the GPU would only arrive sometime in the first half of 2017. Now, reports indicate that the chip could drop in the first quarter — still quite late compared with Nvidia’s Pascal, which will have been in market for nearly a year at that point, but better than the late June timeframe typically hit by “H1” launch windows.
The upcoming Vega 10 chip will feature 24 TFLOPS of 16-bit precision performance, 16GB of HBM2, and 512GB/s of memory bandwidth, with 64 compute units in total, according to VideoCardz. If Vega keeps GCN’s 64 compute cores per compute unit, that puts a full-size Vega core at 4,096 cores in total. The 16GB of HBM2 is expected based on the specs we’ve seen for that standard, but the RAM bandwidth is significantly lower than what we’ve expected to date. HBM2 is designed to support up to 256GB/s of bandwidth per stack, and first-generation HBM2 solutions are expected to ship with 4GB stacks. That works out to up to 1TB/s of memory bandwidth for a full implementation — and the first-generation HBM implementations that AMD shipped were full implementations.
If these stats are accurate, it implies that AMD has kept bandwidth flat between Fiji and Vega rather than improving it further by pushing HBM2. Nvidia’s GP100, which is set to debut in 2017, offers 720GB/s of memory bandwidth. 512GB/s of bandwidth would compete well against Nvidia’s GTX 1080 and Titan X, however, and AMD may have scaled its bandwidth accordingly. TDP on the new high-end card is supposedly 225W, which would put it near Nvidia’s Titan X.
The report suggests AMD will launch Vega 10 first, with Vega 11 supposedly replacing Polaris 10 later in the year. This would imply AMD will waterfall HBM2 down to cheaper SKUs relatively quickly, but we’d still be surprised to see Vega 11 pushing into the $ 200 – $ 240 market where Polaris is supposed to live. Meanwhile, there’s supposedly a Vega 20 on the roadmap as well, with 32GB of HBM2 memory and a 7nm architecture. Current ship dates for 10nm put that process technology as available in 2017, which means we’re looking at a 2019 or 2020 delivery date for a 7nm replacement.
This is less an issue than it might seem. GlobalFoundries has said it wants to deliver 7nm and skip over 10nm entirely, but even if it manages to pull this off, we’ve already seen evidence that GPU developers are capable of delivering performance improvements even while stuck on the same process node. Nvidia may have done more of this than AMD did on 28nm, but there were real performance and performance-per-watt improvements for GCN over the same time period.
There’s no word on the specifics of Vega’s architecture, and that’s what we’d need to make any predictions about overall performance. Our sources have indicated Vega is a ground-up new architecture rather than a GCN derivative — but what that means for AMD and its relative performance position to Nvidia is still unknown. AMD’s Polaris architecture delivered some real improvements over their 28nm product, but Nvidia’s Pascal family still holds the upper hand in price/performance comparisons, thanks partly to the ongoing trouble AMD has had in delivering hardware that meets its actual MSRPs. NV has had problems of its own as well, but as we’ve discussed, not as much as AMD has had.