Nvidia could join AMD and Intel in developing a multi-chiplet architecture for its next generation of GPUs, and could reap major performance gains.
Nvidia is the last of the Big Three chipmakers to still use a single wafer of silicon for the processors of its top graphics cards. So it’s rather surprising that rumors have started circulating that the company is finally moving to multi-plus adaptable solutions. chiplet module (MCM) design with its next-generation Nvidia Blackwell architecture.
The report comes from a well-known hardware leaker @kopite7kimi on X, who said that Nvidia’s commercial-grade GB100 GPU will feature MCM for the first time.
After the GA100 and GH100 dramas, it looks like the GB100 will finally use MCM.September 18, 2023
The Nvidia Blackwell architecture is expected to power both Nvidia’s next-generation commercial GPU products, used by data centers and industrial-scale users, as well as its consumer graphics cards, the Nvidia RTX 5000 series.
Even though both will use the Blackwell architecture, it is currently unclear whether the MCM change will also extend to Nvidia 5000 series graphics cards. If so, however, it could provide the next generation of Nvidia’s graphics cards deliver transformational performance that was often lacking in some of its more recent RTX 4000 cards.
Chipsets of an MCM design, when interconnected into a single processor, promise significantly faster performance compared to a monolithic silicon wafer. As Tom’s material explains that a single silicon chip is limited by the physical dimensions of the equipment used to make it. Currently, the process used by Nvidia can only produce 26mm by 33mm (858mm²) pieces of silicon. at mostand Nvidia’s commercial-grade GPUs already struggle with this maximum size.
And as it has become exponentially more difficult to further reduce the size of a transistor, the electronic switch inside a chip that produces the logic functionality of a computer, the only way to increase the number of transistors in your GPU to increase performance is to enlarge the chip. than the physical manufacturing process allows.
This is where chiplets come in. If you can produce two or more smaller chiplets, but using special links called interconnects to connect them together so that they act as a single unit, you can effectively build a larger chip than the manufacturing process can support. and significantly improve performance. With an MCM design for its GPUs, Nvidia might be able to deliver across its entire portfolio of Nvidia 5000 series cards the kinds of gains that many were hoping to see with the 4000 series, but that Nvidia hasn’t been able to. able to deliver consistently.
Obviously this is still very speculative and based on rumors, but there’s a reason why AMD and Intel have both opted for MCM in their GPUs and CPUs, and Nvidia would be very smart to follow suit, or else risk losing themselves. let yourself be left behind.
Switch to MCM, Nvidia, it’s the only way to go
The problem that chipmakers have faced for years is the end of Moore’s Law, Intel co-founder Gordon Moore’s famous prediction that the density of transistors on a chip would double roughly every two years.
This has been the case for 50 years, but because we now measure the size of transistors relative to the diameter of individual atoms in silicon, cutting the size of a transistor in half is simply no longer possible.
But consumers and industry have gotten used to faster computers every couple of years, and no one really wants to hear that the party’s over. If you’re a chipmaker looking to keep selling more processors, you need to find another way to get the performance gains the market expects, Moore’s Law be damned.
The answer to this question is to use multiple chips in conjunction with each other to achieve these performance goals. We’ve been doing this for over a decade, as Nvidia knows well.
There was a time when there was no such thing as a GPU, there was only the main CPU which was supposed to handle the graphics along with all the other operations.
However, as graphics became more advanced, the CPU was so heavily stressed that something had to be done before the calculation of 3D scenes consumed 99.9% of the CPU clock cycles or the CPU limits itself does not stop the progress of the infographic.
The solution was to shift all but the most basic graphics processing work to a second processor, the GPU, which was designed specifically for the task and has continued to power the modern era of computer graphics. Nvidia knows this because it created the world’s first GPU, the Nvidia GeForce 256, in 1999.
Then we’ve come full circle, and the GPUs are so overwhelmed by the workloads assigned to them that they can’t keep up and we can’t squeeze more performance out of the same size silicon. It’s then time to split geometry, rasterization, ray tracing, machine learning, and other GPU workloads into different mini-processors that can be specifically designed to perform these tasks faster and more efficiently than we do. are currently doing so.
Nvidia’s main competitor, AMD, is already doing this and has had very positive results so far. And while early attempts at pioneering MCM engineering may not be the revolution that the first GPU was when it arrived over 20 years ago, future attempts will take us to where we want – and Nvidia needs – to be, so Nvidia probably should. get to work on that.