Again in 2018, AMD launched the MI50 and MI60, Vega-based accelerators constructed on TSMC’s 7nm course of node. Now, there are rumors of an enormous new chip coming in that household, in a comparatively svelte energy envelope. “Arcturus” is a codename that’s been floating round for some time, having first been talked about by an AMD staffer in late 2018, however it’s by no means been clear if the GPU was based mostly on Navi or Vega. So far, AMD has maintained a divide between its Vega merchandise, that are constructed for AI and the HPC market, and Navi, which is constructed for gaming.
This rumor means that Arcturus is a Vega-derived GPU, although it appears as if it’d use the Vega refinements additionally coming down the pipe for the Radeon Cellular 4000 collection. I’ll get into why after we speak concerning the specs. The leak, by @Komachi_Ensaka, signifies a 32GB HBM card (similar most RAM loadout as the present Intuition household), with as much as 8192 GPU cores, a lift clock of 1.33GHz, and a base clock of 1GHz.
Base Clock : 1090Mhz.
Increase Clock : 1333Mhz.https://t.co/ZquA0jrej3
— 比屋定さんの戯れ言@Komachi (@KOMACHI_ENSAKA) February 9, 2020
All of this, supposedly, in only a 200W TDP.
Is it doable to cram this sort of efficiency into an information heart GPU? The reply could be sure. Let’s have a look at a couple of items of proof.
First, we all know AMD has gotten a lot better at Vega clocking. One of many shocking issues concerning the Ryzen Cellular 4000 announcement was that the GPUs in these elements are clocked a lot larger than the outdated 12nm APUs have been — as much as 1750MHz on the Ryzen 7 4800U. It’s doable that AMD’s newest Vega architectures on 7nm are extra environment friendly than the preliminary designs, though the corporate hasn’t technically moved to a brand new node.
Second, we all know that binning can yield vital enhancements. The Radeon Nano was a 28nm card with higher energy effectivity than the 16nm Polaris playing cards AMD launched after it. AMD improved the Nano as a lot because it did in comparison with the common Fury X by decreasing its clock barely and binning for the perfect playing cards. CPU and GPU energy consumption isn’t linear — sometimes the previous few hundred MHz value far more, when it comes to energy consumption per efficiency gained. By falling again to an earlier candy spot on the curve, AMD might be able to maximize the facility effectivity of the chip.
I imply, heck. I actually simply published a review of the Ryzen Threadripper 3990X, a CPU that doubles the 3970X’s core rely and retains it working in the identical 280W energy envelope by… decreasing the clock pace modestly. So clearly we’re not coping with some loopy concept right here. With that mentioned, yanking energy consumption down by a full third is just about the utmost we’d anticipate AMD to have the ability to obtain. That’s a really aggressive energy enchancment and not using a node shift.
It could be good to have some sense of which clients AMD is promoting these merchandise to. Navi OpenCL help continues to be MIA on Linux as of late December according to Phoronix, and AMD has to offer compatibility with CUDA merchandise by way of efforts like ROCm. It’s not clear if ROCm is definitely being utilized in any tasks — AMD may be very quiet about this facet of the enterprise.
We’ve suspected Google had so much to do with AMD’s resolution to commercialize the MI50 and MI60 product households, however a chip like MI100 would clearly be developed with totally different targets in thoughts. Reminiscence bandwidth is unsure; the chip might persist with the identical 1TB/s of reminiscence bandwidth because the MI50 and MI60 or push as much as 1.2TB/s with higher-speed HBM2.
It’s unlikely we’d ever see Arcturus come to the patron market. AMD’s Huge Navi is predicted to reply the wants of these clients, whereas a card like this can be reserved for information facilities or HPC.