AMD has patented a brand new GPU design that takes a radically new strategy to constructing a graphics chip. The patent particulars an AMD GPU design that accommodates 14 chiplets, that are divided into a number of sections, enabling it to all act as one single GPU, and even a number of GPUs.
The most effective graphics playing cards from AMD, such because the Radeon RX 7800 XT, are already divided into a number of chiplets beneath the hood, however the GPU itself continues to be multi function chip – what we name a monolithic die. It’s solely the cache that’s separated out into additional chiplets on these GPUs. Conversely, this AMD patent successfully particulars 9 separate GPU dies in a single chip package deal.
The figures within the patent present the GPU being divided into three sections, every of which comprise a entrance finish chiplet and three shader engine dies, the latter of which might comprise all of the stream processors used for graphics processing. Every of those clusters is linked to one another, and can talk with two additional dies – a mode selector and a multimedia die.
All of it appears to be like remarkably just like the leaked design of the AMD Radeon Navi 4c chip, which was rumored to be coming with AMD’s forthcoming RDNA 4 gaming GPUs, such because the Radeon RX 8900 XTX later this yr, however has since been reported to be cancelled.
Nonetheless, this patent, which was noticed by Tom’s {Hardware}, exhibits that AMD has positively been engaged on this concept, so it’s now not only a rumor. It’s a sophisticated design for positive, however taking this strategy would have plenty of advantages for AMD. Advanced monolithic dies could be costly to fabricate, and also you get restricted portions of them on a wafer.
Separating it out into a number of smaller parts can’t solely enhance yields with regards to manufacturing (a producing error on one small chiplet die is way more preferable to at least one on a complete GPU), but in addition allows AMD to combine and match totally different manufacturing processes.
As with AMD’s Ryzen CPUs, which use totally different processes to make the CPU cores and the I/O hub, AMD might make its shader engine dies on the newest TSMC course of, the place pace and effectivity are paramount, however might then produce the multimedia die and mode selector dies on an older, cheaper, and extra readily-available course of.
This strategy would additionally make it a lot simpler for AMD to scale its merchandise, because it wouldn’t have to manufacture a number of monolithic dies. As an alternative, it might produce an affordable GPU with one cluster of dies, for instance, a mid-range one with two clusters, and a high-end one with all three. All of the fashions could possibly be based mostly on precisely the identical silicon, simply with totally different numbers of dies beneath the hood.
Certainly, AMD states as a lot within the patent summary, saying “a configurable variety of GPU chiplets are assembled right into a single GPU, such that a number of totally different GPUs having totally different numbers of GPU chiplets could be assembled utilizing a small variety of tape-outs and a multiple-die GPU could be constructed out of GPU chiplets that implement various generations of know-how.”
In fact, AMD hasn’t launched any particulars about RDNA 4 but, however this concept has clearly been within the works for some time. Though the patent publication date is 13 June, 2024, it was initially filed on 8 December, 2022, earlier than even the Radeon RX 7900 XTX had been launched.
We’ll have to attend and see whether or not AMD does certainly make a multi-chiplet gaming GPU, however the indications in the intervening time are that RDNA 4 will purpose on the mid-range GPU market, providing RTX 4080-level efficiency for half the value, and depart the excessive finish to Nvidia.
Within the meantime, try the current AMD Zen 5 announcement to see what the corporate is planning for its new vary of Ryzen 9000-series CPUs, that are as a consequence of be launched in July 2024.
