RAM chip producer Micron has just lately been making some fascinating claims about its subsequent era of ultra-fast reminiscence for graphics playing cards, GDDR7. In comparison with what’s at present getting used (GDDR6X and GDDR6), Micron says its forthcoming tech is “anticipated to realize higher than 30% enchancment in frames per second for ray tracing and rasterization workloads.”
That is a exceptional efficiency enhance, regardless of the way you take a look at it, and it is usually the protect of the numerous architectural modifications a brand new GPU design brings in. Nonetheless, whereas it is definitely true that the information switch fee and bandwidth of GDDR7 will likely be at the least 30% higher than the quickest GDDR6/6X on provide proper now, it is a totally different story altogether in terms of precise video games and functions.
Micron provides all of the GDDR6X chips utilized in Nvidia’s graphics playing cards and the quickest it affords is rated to 24 GT/s (24 billion transfers per second). It additionally sells 18 GT/s GDDR6, although Samsung has 20 GT/s GDDR6 on the market. With regards to GDDR7, although, solely Micron is offering any specs in the intervening time and the 2 fashions obtainable for sampling are rated at 28 and 32 MT/s.
Take 24, enhance it by 30%, and also you get a price of 31.2 so it is apparent the place Micron is getting its efficiency claims from. However as an example you may take that GDDR7 and add it to a present graphics card (ignoring the truth that it really would not work as a result of the GPU cannot use it)—would video games and benchmarks be 30% quicker, like Micron says?
I ran by just a few exams on an RTX 4080 Tremendous, altering the clocks on its VRAM throughout the widest vary I might handle. Every thing else remained the identical, the efficiency variations seen are purely from altering the reminiscence clocks. First up had been two 3DMark exams, Metal Nomad and Velocity Approach. The previous makes use of conventional rasterisation methods for all of the graphics, whereas the latter shoehorns in a good quantity of ray tracing.
It wasn’t doable to realize a 30% distinction in clock speeds however 18% is sufficiently sufficiently big to have the ability to extrapolate how a lot distinction quicker VRAM would make. As you possibly can see above, the soar of 18% between the slowest and quickest VRAM speeds solely produced a 5% and seven% acquire in body charges in 3DMark exams.
It was the same story in two video games I checked out, Cyberpunk 2077 and Returnal, each at 4K. With no ray tracing enabled, the most important enhancements had been within the minimal body charges with 12% for Cyberpunk 2077 and 11% for Returnal. Enabling ray tracing (path tracing for CP2077), together with DLSS Balanced and Body Era to get some playable body charges, dropped these slightly—11% and seven% positive factors for the 2 video games.
For an 18% enhance in VRAM velocity, the most important acquire I noticed was 12%, which was fairly good. Nonetheless, that was just for the minimal body charges, and the averages solely improved by 6% at greatest. This means that if the RTX 4080 Tremendous could possibly be geared up with 30% quicker VRAM, one definitely would not see that type of a rise within the common body charges of these video games and exams.
Sure, it is solely a tiny fraction of all of the video games one might probably take a look at, however given how graphically demanding they had been, they had been good candidates for any {hardware} modifications that present a efficiency enhance.
So does that imply Micron is mendacity by its again tooth over these claims? Not likely as a result of I strongly suspect that there are some “ray tracing and rasterization workloads” that may present a big enhance in body charges when utilizing quicker VRAM, and as soon as GDDR7 is out in the true world, there’s sure to be at the least one sport that is used a advertising instrument to spotlight the positive factors that GDDR7 brings.
However it’s price remembering that the overwhelming majority of builders should not going to make a sport that has 100% memory-limited rendering efficiency. It is going to rely way more on the variety of shader models and the structure underlying all of it.
At this level, you may be questioning what is the level of getting GDDR7. The reply lies in the truth that present GPUs are partially designed round one of the best obtainable reminiscence know-how—there is not any level in making a GPU with a gazillion shaders when it would not be bodily and financially doable to produce it with sufficient VRAM bandwidth to feed all these shaders.
Nonetheless, GPUs of the long run will have extra shaders than these we are able to at present purchase, and the extra highly effective fashions would require quicker VRAM to maintain them fed with information. The following RTX 5090 might properly be 30% quicker in video games than the present RTX 4090, nevertheless it will not actually be all the way down to GDDR7 being 30% quicker than GDDR6X. It will be all the way down to the same old issues: Extra shaders, increased clock speeds, and extra cache.