Both the processors and the dedicated graphics for data centers have nothing to do with the models available for the domestic market, not only in terms of performance, but also in terms of price. An example of what we are talking about is found in the processor AMD’s most powerful server server that the company has just presented exceeds $14,000.
If we talk about GPUs for servers we have to talk about NVIDIA, a company that dominates the domestic and server market. Companies that opt for the new generation of NVIDIA server graphics, DGX B200you are going to have to dig deep into your pocket considering the price that has been made public and which exceeds $500,000.
More than $500,000 for a Blackwell DGX B200 system
It is important to note that this price does not correspond only to a graphics card but to an entire set consisting of, in addition to NVIDIA graphics, an Intel processor, storage and memory. The Broadberry company, one of the companies that sells these AI platforms, already has it for sale for $515,410.43 (plus taxes) with the following specifications:
- 8 NVIDIA Blackwell GPUs
- GPU memory 1,440 GB total – 64 TB/s bandwidth HBM3e
- Dual Intel Xeon Plantium 8570 processor with 56 cores and 112 threads each.
- 4TB DDR5 memory
- 4TB (2×1.92TB) NVMe M2.
- 32TB (8×3.84TB) NVMe SSD.
As we can read in the description, the DGX B200 is capable of offering up to 72 petaFLOPS of performance. AI training and up to 144 petaFLOPS of performance inferencewhich is a significant improvement compared to the previous generation with Hopper architecture.
According to the information that NVIDIA offers on its website, the new DGX B20 offers up to 15 times higher performance than DGX H100 in real-time large language model inference, while the improvement with AI training performance is 3x.
Although NVIDIA officially presented this new generation of server graphics, they are not yet available on the market, although the first batch has already been delivered to OpenAI.
NVIDIA’s product aimed at the broadest segment of the market is the DGX B200, however, it is not the only one it plans to launch on the market. It is likely that, for Microsoft, Meta and OpenAI, there will be other more complete models that will never be released and that will cost even more money than this one, which we could consider as the entry model.
Such has been the expectation of this new generation of graphics for servers with Blackwell architecture that, according to Morgan Stanley, NVIDIA has all the fish sold until the end of 2025, who claims that demand has been much higher than the previous generation.
AMD presented the accelerators last week MI325X and by 2025 the new ones are expected MI335Xand according to the company, they have little or nothing to send to the NVIDIA B200. These new accelerators in combination with the new EPYC processors may endanger NVIDIA’s reign in AI.
