Valve has released its Proton 9.0 software update which adds improved support for NVIDIA GPUs and high core count CPUs.
14.04.2024 - 16:23 / wccftech.com / Muhammad Zuhair
Jim Keller has an interesting proposition on NVIDIA's Blackwell GPUs, claiming that the firm should've utilized an alternative cost-efficient interconnect method.
Well, Tenstorrent's CEO Jim Keller is known for some of *fun* his takes on tech products. Recently, he called NVIDIA's CUDA a swamp, too, so you can imagine how he pans out on social media. Jim has now criticized Team Green's R&D budget for their Blackwell GPUs, claiming that the company would've saved billions by simply changing the interconnect method. You can see it in his tweet on X below:
In terms of the Blackwell interconnect, NVIDIA has utilized their 5th Generation "NVLINK" interconnect technology. While NVLINK has significant advantages over other methods in terms of bandwidth, latency, and scalability, it lags due to its "proprietary" nature, which can't cater to any other manufacturers' hardware.
Moreover, the utilization of NVLINK is said to be a high-cost move in industry dynamics, which is why Jim Keller has criticized it, especially since he promotes an "open-source" ideology for the markets, but NVLINK is the exact opposite here.
Now, the alternative, in this case, is Ethernet interconnect, which is a more cost-effective option and destroys the proprietary state of NVLINK, ultimately providing a much better ecosystem, but in terms of the performance figures, Ethernet needs to do a look at catching up to Team Green's NVLINK combined with InfiniBand, but with UE (Ultra Ethernet) interconnect in the way, things could potentially change here, fueling the standard's adoption, since UE guarantees effective AI and HPC performance. However, it is still in the development phase.
Jim Keller believes that NVIDIA's whopping $10 billion R&D costs on Blackwell are too much, and the company could have significantly reduced it by simply changing interconnect. Jim is known to troll companies, especially NVIDIA, multiple times, but his claim does sound viable to an extent, as Blackwell's costs are simply outrageous.
But despite what you may say about NVIDIA's Blackwell AI GPU costs, there's no doubt that the demand for these chips is sky-high & even companies such as Intel and AMD are having a hard time breaking NVIDIA's dominance which has only been accelerated with the announcement of Blackwell.
Valve has released its Proton 9.0 software update which adds improved support for NVIDIA GPUs and high core count CPUs.
NVIDIA's first batch of its highly-anticipated Blackwell DGX GB200 AI servers is expected to drop into markets by 2025, starting the next gold rush in the industry.
Apart from gaming and AI, Hive Systems shows that complex passwords are no match for top GPUs such as the NVIDIA GeForce RTX 4090.
NVIDIA's GeForce RTX 40 GPUs, including the RTX 4060 Ti, are experiencing some major supply issues, leading to increased prices.
NVIDIA is expected to dominate the AI market in 2024 as they are set to sell $40 billion worth of GPU, significantly outpacing AMD & Intel.
Samsung has witnessed the signing of a massive contract for its HBM3E memory, securing $3 billion worth of supply to none other than AMD.
NVIDIA's monopoly over the AI markets is now being questioned by several analysts, who claim that increasing GPU power consumption won't be sustainable in the future.
SSD & HDD prices are expected to skyrocket, with two big manufacturers, Western Digital, and Seagate, anticipating increased demand.
NVIDIA has finally acknowledged tough competition in the AI segment, claiming that AI is the "largest computational problem in history."
AI startup TensorWave, believes that AMD's Instinct MI300X accelerator is a far better option than competitors such as NVIDIA's H100.
NVIDIA's Blackwell AI GPUs are expected to lift every other associated segment, including CoWoS (TSMC) & HBM DRAM, as the market expects millions of chips to be shipped by 2025.
Startup, Celestial AI, has developed a new interconnect solution that utilizes DDR5 & HBM memory to increase the efficiency of chiplets with AMD possibly being among the first to use such a design.