Jim Keller, the man who gave both the world and AMD the Zen CPU architecture in 2017, recently said that the fifth generation of the architecture, Zen5, would produce a major leap in performance, over the current Zen4 architecture. Keller, who is currently the CEO of Tenstorrent, an AI hardware firm, shared his insights on this during the RISC-V keynote that took place in Bengaluru, India, earlier this month.
Keller was talking about his company’s Ascalon processor cores, and was showing its performance versus the range of other datacentre CPUs on the market, including Intel’s Sapphire Rapids, Amazon’s Graviton 3, NVIDIA’s Grace, and AMD’s Zen5. Now, it should be mentioned that AMD had said that its next-generation Zen architecture would be making its debut in 2024 during its Financial Analyst Day in June 2022 but since then, the chipmaker has produced no tangible proof of performance for it and is focusing on its current Zen4 architecture.
https://www.youtube.com/watch?v=SupjLqPbFpY&ab_channel=Tenstorrent
Keller gave a prediction that AMD’s Zen5 architecture will likely be 30% faster than the current Zen4, in terms of integer workloads. That is a massive leap in performance on a generation-to-generation scale, especially when you consider that Zen4 already provides 15% more IPC over the previous generation Zen3 architecture. Again, he’s referring to server CPUs, and not so much consumer CPUs.
As far as clockspeeds go, Keller says that Zen5 processors would likely hit the 4GHz mark, which would make them the first server-grade CPUs to hit those speeds. Of course, the title belt of fastest datacentre CPU of the generation is still up in the air, given that there is still no word on Intel’s upcoming Emerald Rapids, the successor to the current Sapphire Rapids that also just made its debut after lengthy delay.
Again, Keller’s predictions about AMD’s Zen5 architecture are simply that: speculations. So, we ask that you simply treat his ramblings with some degree of skepticism.
(Source: Tenstorrent via YouTube, Techspot)
