What is NVIDIA Blackwell? All about the GPU ar­chi­tec­ture

NVIDIA’s Blackwell mi­croar­chi­tec­ture is setting new standards in GPU tech­no­logy, bringing sig­ni­fic­ant per­form­ance im­prove­ments and higher ef­fi­ciency to various areas of use. NVIDIA Blackwell offers modern tech­no­lo­gies that are specially optimised for complex cal­cu­la­tions.

What is NVIDIA Blackwell?

NVIDIA Blackwell is modern mi­croar­chi­tec­ture that was launched in 2024 and named after the math­em­atician David Blackwell. It was specially developed for use in high-per­form­ance graphics pro­cessors. Blackwell is the successor to Hopper GPU ar­chi­tec­ture and delivers sig­ni­fic­ant im­prove­ments in per­form­ance and higher energy ef­fi­ciency.

Blackwell was optimised for use in ar­ti­fi­cial in­tel­li­gence (in par­tic­u­lar gen­er­at­ive AI and large language models), machine learning, sci­entif­ic cal­cu­la­tions and gaming. The ar­chi­tec­ture is based on advanced man­u­fac­tur­ing tech­no­lo­gies and uses modern chiplet design to achieve higher per­form­ance. A special feature of Blackwell is the improved memory ar­chi­tec­ture that enables faster data pro­cessing. It also offers optimised support for ray tracing and more efficient use of Tensor Cores for AI ap­plic­a­tions. NVIDIA pri­or­it­ised optimised scalab­il­ity in the design, which makes the ar­chi­tec­ture suitable for data centres as well as high-end consumer products.

What new features does NVIDIA Blackwell have?

Compared to its pre­de­cessor Hopper, Blackwell GPUs offer a number of technical in­nov­a­tions. One of the most important is the in­tro­duc­tion of an advanced chiplet design that enables increased scalab­il­ity and ef­fi­ciency. The design makes it possible to combine several smaller chips rather than using a single mono­lith­ic chip, which sig­ni­fic­antly improves per­form­ance. Ad­di­tion­ally, Tensor Cores were optimised for ac­cel­er­at­ing deep learning models more ef­fi­ciently.

Ray tracing per­form­ance was also improved, which enables more realistic lighting and shadow cal­cu­la­tion for games and graphics ap­plic­a­tions. The memory ar­chi­tec­ture was also modified. With the new gen­er­a­tion of HBM (High Bandwidth Memory), NVIDIA Blackwell can reach much higher memory band­widths. Energy ef­fi­ciency was also improved using new man­u­fac­tur­ing tech­no­lo­gies and better cooling mech­an­isms. And finally, NVIDIA Blackwell also supports in­ter­faces such as the next gen­er­a­tion of NVlink and PCIe 5.0, which enable faster com­mu­nic­a­tion between different GPUs and CPUs.

What are the main areas of use for NVIDIA Blackwell?

NVIDIA Blackwell ar­chi­tec­ture is designed for a variety of high-per­form­ance ap­plic­a­tions:

• Ar­ti­fi­cial in­tel­li­gence and machine learning: Optimised Tensor Cores enable Blackwell to train and run very large AI models. Blackwell GPUs are equipped with Con­fid­en­tial Computing, which uses hardware-based security features to protect con­fid­en­tial data and AI models from un­au­thor­ised access. That makes it ideal for companies that work with gen­er­at­ive AI, neural networks and autonom­ous driving and want to increase both per­form­ance and security.
• Data centres and cloud computing: Blackwell GPUs are specially optimised for use in data centres and offer high computing power with re­l­at­ively little energy use. That makes them suitable for big data analysis, sim­u­la­tions and sci­entif­ic cal­cu­la­tions (e.g. weather forecasts and climate modelling).
• Gaming and 3D graphics: Thanks to more efficient AI ac­cel­er­a­tion, game de­velopers can better combine ray tracing with classic ras­ter­isa­tion tech­no­lo­gies to create more realistic scenes with optimised per­form­ance. That makes Blackwell suitable for high-end gaming computers and VR ap­plic­a­tions.

What are the ad­vant­ages and dis­ad­vant­ages of NVIDIA Blackwell?

Ad­vant­ages of NVIDIA Blackwell

NVIDIA Blackwell mi­croar­chi­tec­ture has a number of sig­ni­fic­ant ad­vant­ages that make it a great choice for pro­fes­sion­al ap­plic­a­tions and gaming. One of the biggest ad­vant­ages is the sub­stan­tially increased computing power achieved with the new chiplet design. Rather than in­cor­por­at­ing a mono­lith­ic chip, Blackwell uses a modular structure that enables better scalab­il­ity and more efficient man­u­fac­tur­ing. That means that high-per­form­ance GPUs can be produced more flexibly and optimised for different areas of use.

Another advantage of Blackwell GPUs is improved energy ef­fi­ciency. The use of advanced man­u­fac­tur­ing tech­no­logy allows for a reduction in energy use in relation to computing power. That’s par­tic­u­larly important for data centres that require high per­form­ance for the lowest possible energy con­sump­tion.

In addition, Blackwell GPUs are equipped with Tensor Cores, which were specially developed for machine learning and AI ap­plic­a­tions. They make it possible to quickly train and run complex models, making Blackwell an ideal solution for companies and research in­sti­tu­tions.

Gamers also benefit from these im­prove­ments, since improved ray tracing tech­no­logy enables a more realistic rep­res­ent­a­tion of light and shadow in modern games. New memory tech­no­lo­gies also ensure faster data pro­cessing and reduced latency.

Dis­ad­vant­ages of NVIDIA Blackwell

Despite all those advances, there are still several chal­lenges and potential downsides to Blackwell GPUs. One of the most obvious is the high price, es­pe­cially for pro­fes­sion­al models and server solutions. Its advanced tech­no­logy makes Blackwell an expensive in­vest­ment that doesn’t ne­ces­sar­ily make sense for every use case. Although energy ef­fi­ciency was improved, the power usage involved in high-per­form­ance computing is sig­ni­fic­ant, es­pe­cially when it comes to multi-GPU setups.

Another potential hurdle is the need for specific software op­tim­isa­tions. To get maximum per­form­ance from Blackwell-based GPUs, de­velopers need to adapt their software ac­cord­ingly. That can present a challenge in pro­fes­sion­al workflows if existing programs aren’t fully optimised for the new ar­chi­tec­ture.

Finally, Blackwell GPUs may not be widely available at first, as the demand for high-per­form­ance GPUs increases and new tech­no­lo­gies are often initially produced in small quant­it­ies.

Ad­vant­ages and dis­ad­vant­ages at a glance