GDDR 5 vs GDDR 6 | GDDR vs HBM vs GDDR 6

| |

What is GDDR and HMB ?

Next to the GPU, memory is the most important component of a graphics card, because the data must be fed to the GPU as quickly as possible. Via the PCI Express interface they first reach the graphics memory and from there can be loaded into the GPU at a speed of 510 GB/s and more. Over the years, memory technology for graphics cards has developed rapidly. The demand for more and more bandwidth has also led to a two-track development. DDR has become GDDR and in the future we will see more and more graphics accelerators relying on HBM.

GDDR stands for Graphics Double Data Rate and, along with High Bandwidth Memory (HBM), is an important memory standard on current graphics cards. As with DDR mainboard memory, there are several generations of GDDR memory. Together with Micron, NVIDIA developed a faster variant of GDDR5 called GDDR5X, which was used on some Pascal cards. The double data rate of DDR memory, and hence GDDR, is achieved by transferring at both the rising and falling edges of the clock signal.

Over the generations, the bandwidth of GDDR memory has been massively increased. At the same time, power consumption has decreased significantly. With GDDR, we are talking about a memory bandwidth of 25.6 GB/s on a 256-bit wide memory interface. With GDDR6, 320 GB/s are already achieved today and even faster variants are planned for the future. The clock rates have been increased over the generations from 166 MHz to now 1,750 MHz and more.

For the new GeForce RTX cards, NVIDIA relies on the new GDDR6 memory manufactured by Micron and Samsung. Depending on which model you get here, it uses either Samsung or Micron memory. There is no difference in performance.

The competition to NVIDIA

While NVIDIA’s competitors made the switch to HBM in the desktop market, they had to make some compromises in terms of availability and storage size. HBM also has the disadvantage of having to design the memory interface quite wide, which needs space and therefore causing development costs of the GPU. On the other hand, HBM and the GPU are in the same package, which allows a compact design of the graphics card overall.

NVIDIA offer graphics cards with DDR graphics memory, i.e. GDDR5(X) and GDDR6, as well as HBM (High Bandwidth Memory). HBM was developed to meet the demand for ever greater memory bandwidths. HBM2 now achieves data rates of up to 1 TB/s, while GDDR6 has reached the end of the line at 672 GB/s.

NVIDIA currently uses HBM almost exclusively on the GPU accelerators in the server area. An exception are desktop hermaphrodites such as a Titan V. GDDR6 will continue to play an important role for NVIDIA for the foreseeable future and will meet high memory bandwidth requirements.

Memory bandwidth is a technical specification. It is supported by methods of compressing the data in memory. This saves space in the memory itself, but also accelerates the transfer of data. Delta color compression has been used in NVIDIA GPUs for several generations. NVIDIA is the 5th generation of such a compression method.

It is important to note that this is a lossless compression method. This means that no data is lost and developers can rely on the method without having to develop specially adapted versions.

NVIDIA uses Delta Color Compression for memory compression. Only the base pixel value is stored and only the difference (the delta) is stored for the surrounding pixels in an 8×8 matrix. Since the delta is a much smaller value, it can be stored faster and less memory space is required. So less data has to be written to and read from the VRAM. But it is also possible to compress the single color value, so that also here memory space or better memory bandwidth can be saved.

An example of compression is a complete black and white whose value is usually stored in memory as {1.0, 0.0, 0.0, 0.0} or {0.0, 1.0, 1.0, 1.0}. In a simple procedure, however, the values 0.0 or 1.0 are sufficient to describe this unambiguously.

NVIDIA has improved the procedures for detecting compressible image content. The already known 2:1 ratio can therefore be applied more quickly and can also be used for a larger data set. New additions are compressions by a factor of 4:1 and 8:1.

Thus, in order to increase the memory bandwidth due to the faster memory, it is possible to reduce the amount of data that has to be transferred, which further increases the effectiveness of the memory interface.


If an HBM is used instead of the classic GDDR memory, the more compact GPU package allows an equally compact PCB design. On desktop graphics cards, however, saving space on the PCB only plays a minor role. This is more interesting for compact systems such as laptops. Here, every square millimeter plays an important role, the GPU and HBM design is advantageous here.

In future, it will have to be shown what role HBM will play in the gaming segment. By using GDDR6 memory on the current GeForce RTX cards, NVIDIA has shown that less complex memory can be a good and, above all, fast alternative to expensive HBM.

GDDR 5 vs GDDR 6

A fast and large graphics memory is characteristic for a high-end graphics card. SK Hynix announced the new GDDR6 at Nvidia’s graphics trade fair GTC 2017. This is to go into production in 2018. What did the new memory bring?

Samsung has started to produce the graphics card memory of tomorrow.The mass production of the first 16-Gigabit (Gb) Graphics Double Data Rate 6 memory, i.e. GDDR6, began 2018. These are used not only in gaming graphics cards and devices, but also in artificial intelligence and automotive systems.

The new GDDR6 memory is manufactured by Samsung using the 10-nanometer process. Compared to the 8Gb GDDR5 memory manufactured using 20-nanometer technology, this doubles the memory density. The pin-speed speed of GDDR6 is specified by Samsung as 8 gigabits per second (Gbps) and the data transfer rate as 72 gigabytes per second. So here too, performance has doubled compared to GDDR5. GDDR6 requires 1.35 volts and is supposed to offer 35 percent less energy consumption than GDDR5 with 1.55 volts.

Due to the increased speed, GDDR6 is interesting for next generation graphics chips. But also for the areas of 8K video processing, virtual reality (VR), augmented reality (AR) and artificial intelligence.


In addition to fast shaders, rendering and texture units, the graphics memory is important for gaming – and how it is connected. Gigantic textures and environment information must be quickly available for the computing elements, otherwise there will be a data jam and falling frame rates. Especially high resolutions beyond 1080p with sophisticated anti-aliasing push the memory of mid-range cards to their limits. Current graphic accelerators mainly rely on GDDR5, more expensive Nvidia models on faster GDDR5X. Meanwhile, AMD’s competitor architecture “Vega” is in the starting blocks with the second generation of High-Bandwidth Memory (HBM2) – whereby alleged production difficulties are delaying mass production.

How do the memory types compare?

In this respect, the data as reported by SK Hynix at the GTC 2017 is particularly interesting when compared to the other major graphics memory of the year – HBM2. In the best case, both memory types draw the level and reach the famous TByte per second, which flows back and forth between memory and controller. The capacity is on a different page. AMD has announced the Vega Frontier Edition, a non-gamer card with 16 GByte memory for the end of June 2020. However, since the size of the GDDR6 modules is supposed to remain identical to GDDR5, the memory manufacturers will probably not be able to crack the 12 GByte per card achieved so far.

HBM2 will be limited to 16 GBytes for the time being, but still has some room for improvement – literally. This is because the 3D memory is currently mostly used in the version with 4 gigabytes per stack (“4-Hi Stack”), but offers the possibility of 8 gigabytes per stack. In this respect, HBM2 graphics cards could reach up to 32 GBytes. The Frontier Edition will offer two 8-Hi Stacks.

Therefore HBM2 graphic cards are faster but also more expensive than the competitor GDDR 6.

DDR4 vs DDR3

DDR4 memory is not completely new. Some mainboards (Socket 2011-3) have been supporting RAM for years, and current entry-level boards are available from around 50 USD.

Difference between DDR3 and DDR4

The DDR4 memory differs from DDR3 in memory density, clock rate and voltage: In principle, DDR4 can hold more GBytes and achieve faster clock rates. However, the timings are usually higher with DDR4 than with DDR3.

Important to know: The latency, which is important for gamers, is calculated from timing and clock frequency. The lower the latency, the faster the system transfers data from the main memory. So here the higher clocking of DDR4 is opposed to the better timing of DDR3. For this reason DDR3 can be faster than DDR4 in practice. You can usually roughly tell which timings your RAM has by the number behind the abbreviation “CL” (Column Address Strobe Latency) – the smaller, the better.

In theory, our DDR3 bar with its short timings should be a bit faster than the DDR4 competitor, despite the lower clock frequency. As it turns out, the memory hardly has any influence on games in this case, though: Both the synthetic benchmarks (3DMark) and the refresh rates are almost identical. There are slight differences in the system tests (PCMark).

Do I need a graphics card ?

Without a graphics chip your screen will remain black. A graphics card is therefore absolutely necessary for communication with the PC. It controls the image display and provides connections for monitors. It also relieves the processor when you want to play games or watch movies.

The demands on graphics chips are increasing: games, high-resolution films and modern operating systems require a lot of computing power, which the processor on the mainboard alone can no longer handle. As early as the 1980s, graphics chips took over the calculation of lines and areas, from which Windows, for example, was the first operating system to benefit. In the 1990s, Doom founded the genre of first-person shooters and demanded a powerful processor and a graphics card with 3D acceleration.

A graphics card processor (GPU) is a highly specialized computing chip for image processing and quality improvement. In office mode, the graphics chip ensures flicker-free images in any resolution. A graphics card does the hardest work when calculating elaborate game scenes, which you can also hear in the increasing fan noise. Current cards like the GeForce RTX 2080 Ti calculate high-demanding games and software and improve the picture quality.

Do I need a fast graphics card?

A new purchase is worthwhile if it allows applications to run fast. Especially current games benefit from fast graphics cards. If you edit your own films and photos and only play games occasionally, you can choose the inexpensive mid-range models. For the office and for playing movies, a starter card or an integrated graphics solution on the mainboard is sufficient.

What is the Graphics cards’ GPU for?

The most important hardware component on a graphics card is the GPU. This is not a bare chip on the PCB (Printed Circuit Board), but in a GPU package. This GPU package consists of a carrier material, usually also a PCB, which enables the chip to be connected to the graphics card itself via a BGA (Ball Grid Array). However, there are also GPUs that are connected directly to the PCB of the graphics card via BGA. So the answer is once again: It depends on how the GPU package is structured.

If you take a closer look at a typical GPU package, the actual GPU can be recognized centrally, but that’s why the first SMD components, which are usually resistors, are already there. This GPU package in turn is connected to the PCB of the graphics card via a BGA. The graphics memory is connected externally in this case and is located outside the GPU package.

NVIDIA also manufactures GPUs in which the graphics memory is positioned in the immediate vicinity in the form of the HBM. GPU and HBM are connected via an interposer. The interposer is also a semiconductor material. Vertical and horizontal conductor paths are inserted into it using various methods to create the connection between the GPU and HBM.

Advantage of HBM

The advantage of HBM is the extremely wide memory interface, which means that extremely high memory bandwidths are also possible. However, such a connection is only possible via an interposer, because 1,024 bits or at least 1,024 traces must be implemented per memory chip. With two or four memory chips we are therefore already talking about more than 4,000 individual connections.

The production of an interposer is not quite easy and above all more expensive than putting a simple GPU package on a PCB via BGA. In addition, it is no longer sufficient to have the GPU manufactured by a contract manufacturer and then placed on the PCB. Other companies must be involved to bring GPU and HBM together on the interposer.

This is also one of the reasons (besides the availability and cost of HBM) why HBM is not yet used on all current graphics cards. Currently, NVIDIA is only using HBM – and thus a corresponding GPU package – on the Titan V. The Tesla product line now relies completely on the faster memory. Here, however, the costs play a less important role and the corresponding applications also depend on the highest possible memory bandwidth.

Power supply and Graphics card

Power Supply of Graphics card

The current and voltage supply plays an important role on modern graphics cards. NVIDIA’s recent reference implementations have been very positive. PCB designs and power supply designs for a GeForce GTX 1080, NVIDIA Titan V and now the GeForce RTX 2080 (Ti) are extremely efficient and well thought out.

The supply of GPU, memory and other components is a point that is extremely important, but is usually neglected. We are talking about the supply of components with up to 20 billion transistors from a 12 nm production in several voltage levels, which have to be tuned extremely precisely. In addition, we are not talking about a continuous supply, but rather one that must be adapted to load changes. Another factor is that a power supply should not become the actual consumer of a graphics card, but should work efficiently.

The Voltage Regulator Modules (VRM) play the most important role within a current and voltage supply. The VRMs make sure that the 12 V coming from the PC’s power supply become about 1 V, which is necessary to supply the GPU and memory.

Many manufacturers advertise with the number of voltage phases. But here, only at first glance “the more, the better” applies. Basically it can be said: The higher the Thermal Design Power, i.e. the consumption of the card, the more voltage phases are necessary for the supply.

In principle, the more phases are installed, the better the supply at higher currents. However, it is also apparent that a majority of required phases pushes the range of highest efficiency further and further upwards. Many phases have high losses during switching.

The more phases, the higher are the unwanted losses. NVIDIA has therefore developed a power supply for the GeForce RTX 2080 and GeForce RTX 2080 Ti that can dynamically switch phases on and off depending on how much is currently required by the card. This ensures that the power supply is always within an ideal range. The GeForce RTX 2080 has an 8-phase power supply that can be dynamically switched on or off between one and the full eight voltage phases. With the GeForce RTX 2080 Ti there are 13 phases.

Problems with balancing power supply

Some will certainly remember the discussion about the power supply for the Radeon R9 Fury X, which sometimes drew significantly more current from one of the two additional 8-pin connectors than is usually the case. At that time, AMD apparently had problems correctly balancing the power supply over the two connectors plus the PCI Express slot.

NVIDIA is also talking about improvements in this area for the GeForce RTX 2080 Founders Edition and GeForce RTX 2080 Ti Founders Edition. This is especially true for the two additional 8-pin or 6-pin connectors, which should now provide better or more evenly distributed power. A circuit on the cards’ PCB ensures this, and NVIDIA has developed new power controllers for this purpose, among other things.

Additional ports for Graphics card

Apart from the mentioned display connections and those for the supply of the graphic card, there are also some more, which partly don’t play a major role anymore. For the operation of an SLI the exchange and synchronization of the frame to be output is necessary. This works with NVIDIA via SLI or NVLink connectors.

Even though the Scaleable Link Interface or SLI for short did not play a major role in the Pascal generation, the connections were still present on cards from a GeForce GTX 1070 and higher, thus allowing the interaction of two cards in a multi-GPU setup. With the Turing GPUs, NVIDIA still sticks to SLI, but uses a transfer technology that was previously only used with the GPU accelerators in servers and some Quadro cards: NVLink. However, currently only the GeForce RTX 2080 Ti and GeForce RTX 2080 models are capable of running in a multi-GPU system via NVLink.

The future of data transmission in Graphic cards

The classic SLI as Alternate Frame Rendering (AFR) will in future be run by NVIDIA via the NVLink connector – the data of the frame buffer of the second card will be transferred to the first card via the NVLink connector. In initial tests, an SLI with the new cards shows amazingly good results, but also depends on the corresponding software profiles. Whether multi-GPU systems will continue to play a role in the future remains to be seen.

NVLink has, as mentioned, a clear advantage with regard to the bandwidth of the interface. The first generation SLI bridge reaches 1 GB/s, the following SLI-HB bridges reached 4 GB/s and made an SLI with 4K resolution with more than 60 Hz possible. One NVLink of the GeForce RTX 2080 reaches 50 GB/s, while the GeForce RTX 2080 Ti with two NVLink ports reaches 100 GB/s.

Are you interestet in more hot tech deals? Subscribe to our Email-list and we let you know when the best deals come up.

Leave a Comment