If you’re a fan of AMD’s GPUs, the past few years have been a bit rough. Polaris has offered competitive performance in the midrange consumer market, but AMD’s Vega fell short of overall expectations. In the wake of Vega’s launch, AMD announced that it would build a 7nm GPU based on Vega and aimed at the machine intelligence market. AMD hasn’t formally shipped MI60 yet, but it announced the part today and confirmed that it would ship before the end of the year.
The new MI60 will use PCIe 4.0 for additional bandwidth and offers up to 1.25x more performance within the same power envelope or 50 percent reduced power at 14nm clock speeds. That 25 percent gain is in-line with what we predicted for potential 7nm performance on Vega, and the power reduction at same-clock is also as expected — GPUs tend to draw much more power when pushed out of their clocking ‘sweet spot.’ It’s the first GPU to use Infinity Fabric over the PCIe bus and packs 13.2B transistors in a 331mm2 package, compared with 12.5B transistors in 486mm2 for the original Vega.
The MI60 will pack 32GB of HBM2 with a massive 1TB/s of memory bandwidth. That’s significantly higher than any GPU currently on the market; Nvidia’s V100 GPUs top out at 900GB/s. Those of you who were holding out some hope that 7nm Vega would debut in any kind of personal SKU or consumer product, we’ve got bad news — AMD explicitly confirmed on stage that this card is solely intended for enterprise markets.
Performance in DGEMM and SGEMM benchmarks is said to be directly comparable to Nvidia’s Tesla V100 and substantially faster than the MI25 (shown above). But of course, these are vendor-provided tests and should be taken with a grain of salt until we know more about the conditions under which the tests were run.
Overall, the MI60 seems to be an aggressive high-end part for this market, but it’s not clear what kind of adoption it will see. Nvidia has a significant lead over AMD in this space, and while offering a high-end 7nm solution for AI/ML workloads is absolutely required for AMD to compete, it will take time for enterprises and companies to evaluate and adopt the solution. AMD aggressively talked up being the first company to ship a 7nm GPU and the first to support PCIe 4.0, but software compatibility and ease-of-porting will probably be more critical to overall uptake.
AMD confirmed that it will ship the card before the end of the year and that it’s already working on MI-Next, but gave no details on any upcoming consumer cards based on Navi.
Now Read: AMD launches Radeon Instinct GPUs to tackle deep learning, artificial intelligence, Epyc Achievement: AMD Now Available for Oracle Cloud Compute Instances, and AMD Moves All 7nm CPU, GPU Production to TSMC