Samsung is now mass producing new higher density GDDR6 memory built on its 10nm-class process technology that it claims offers twice the speed and density of its previous 20nm GDDR5. Samsung's new GDDR6 memory uses 16 Gb dies (2 GB) featuring pin speeds of 18 Gbps (gigabits-per-second) and is able to hit data transfer speeds of up to 72 GB/s per chip.
According to Samsnug, its new GDDR6 uses a new circuit design which allows it to run on a mere 1.35 volts. Also good news for Samsung and for memory supply (and thus pricing and availability of products) is that the company is seeing a 30% gain in manufacturing productivity cranking out its 16Gb GDDR6 versus its 20nm GDDR5.
Running at 18 Gbps, the new GDDR6 offers up quite a bit of bandwidth and will allow for graphics cards with much higher amounts of VRAM. Per package, Samsung's 16Gb GDDR6 offers 72 GB/s which is twice the density, pin speed, and bandwidth than that of its 8Gb GDDR5 running at 8Gbps and 1.5V with data transfers of 32 GB/s. (Note that SK Hynix has announced it plans to produce 9Gbps and 10Gbps dies which max out at 40 GB/s.) GDDR5X gets closer to this mark, and in theory is able to hit up to 16 Gbps per pin and 64 GB/s per die, but so far the G5X used in real world products has been much slower (the Titan XP runs at 11.4 Gbps for example). The Titan XP runs 12 8Gb (1GB) dies at 11.4 Gbps on a 384-bit memory bus for maximum memory bandwidth of 547 GB/s. Moving to GDDR6 would enable that same graphics card to have 24 GB of memory (with the same number of dies) with up to 864 GB/s of bandwidth which is approaching High Bandwidth Memory levels of performance (though it still falls short of newer HBM2 and in practice the graphics card would likely be more conservative on the memory speeds). Still, it's an impressive jump in memory performance that widens the gap between GDDR6 and GDDR5X. I am curious how the GPU memory market will shake out in 2018 and 2019 with GDDR5, GDDR5X, GDDR6, HBM, HBM2, and HBM3 all being readily available for use in graphics cards and where each memory type will land especially on the mid-range and high-end consumer cards (HBM2/3 still holds the performance crown and is ideal for the HPC market).
Samsung is aiming its new 18Gbps 16Gb memory at high performance graphics cards, game consoles, vehicles, and networking devices. Stay tuned for more information on GDDR6 as it develops!
Also read:
Go go Samsnug
Go go Samsnug
30% gain in manufacturing
30% gain in manufacturing productivity over gddr5? nice!
but all a moot point if miners keep the GPU prices up!
May as well have HBM2 and the
May as well have HBM2 and the miners paying the R&D and initial plant and equipment cost for HBM2 production so that HBM2’s inital costs can be fully amortized making HBM2 cheaper in the long run as HBM2 production costs get lower. The more HBM2 volume the better for that economy of scale reason alone. GPUs are in such demand by the miners for hashing compute so GPUs are going to be higher regardless of the VRAM used. So now while GPU prices are high the percantage cost of including HBM2 on the GPU is not as high relative to the overall GPU’s cost compared to the time when GPUs where not in demand by the miners and GPUs where selling for close to MSRP.
Any GDDR6 cost savings is not going to be passed on to the consumer anyways by any GPU makers and the retailers are sure not going to be pricing lower with all that GPU demand in the market currently. The Miners want the HBM2 for that power savings as well as the bandwidth and the best Hashes/Power consumed metrics, because the Hashing algorithms love the bandwidth.
And the Miners do not want mining focuesd GPUs because the miners want a larger used GPU resale market that includes gamers as well as other miners. So even the miners are calculating the used resale demand of GPUs into their overall mining costs recovery equation when it comes time to update their GPU mining eguipment with more powerful and power efficient GPUs. Miners want a larger market of gamers/other miners when it comes time to unload their older GPUs and get back some resale value on their GPU hardware investment.
That’s quite something. You
That’s quite something. You could then, theoretically, have a 4GB 64-bit card with 144 GB/S of memory bandwidth, on the cheap cheap.
Volta 1160 in Oct?
Volta 1160 in Oct?
Anyway to mix ddr6 with some
Anyway to mix ddr6 with some hbm2 as cache? And if that would be of any benefit?
Well the discrete mobile Vega
Well the discrete mobile Vega variants are coming with 4GB of HBM2 which should be fine for bandwidth. And with Vega’s HBCC able to use the HBM2 as the HBC(High bnadwidth Cache) that one stack of HBM2 as 4GB of last level GPU cache and the remainder of the VRAM made virtual and paged out to regular DDR4 syetem memoryor SSD. Those single stacks of JEDEC HBM2 are subdivided into 8, 128 bit indipendent channels with the VEGA HBCC able to swap pages to and from HBM2 and System memory in the background. JEDEC HBM2 even has a pseudo 64 bit channel mode where each of those 8, 128 bit channels can be subdivided into 2, 64 bit pseudo channels for fine grained memory access.
So Really DDR4 memory should be enough with the HBC(HBM2 at 4GB) managed by Vega’s HBCC and that’s the Job of Vega’s HBCC to hide any latency of system memory transfers for even VRAM paged out to SSD/Hard-Drive as Virtual VRAM page swaps done in the background. With Vega any HBM2 can become like a last level GPU cache and 4GB of HBM2/Cache should be fine leveraging 8 or 12 GBs of Virtual VRAM paged to regular system DRAM. Most Games are tuned for 8GB of VRAM anyways but Vega’s HBCC IP should be able to do fine using 4GB of HBM2 as last level GPU cache(HBC) and the rest paged out to system DRAM/SSD.
Hopefully that Samsung Aquabolt HBM2 can be offered in 4GB variants and dependeing on the Vega nCU counts on any Vega discrete mobile SKUs that single stack of HBM2/4GB will offer enough bandwidth for most laptop gaming. That’s the beauty of Vega’s HBCC/HBC IP for discrete mobile gaming as it will be of more use for mobile than desktop where the desktop GPU SKUs usually get 8GB of HBM2 so they rarely need to make use of Vega’s HBCC/HBC IP as the games are mostly not using more than 8GB of VRAM anyways.
Now Graphics Application(Non Gaming) can use pleny more Virtual VRAM than games so even for 8GB of HBM2 Vega’s HBCC/HBC IP can come in handy.
I love to see AMD getting some Zen/Vega APU variants out there with at least a single stack of HBM2 announced so there can be that option by 2019. At least fore an APU’s graphics not having to be starved for bandwidth even if the laptop OEM only uses a single memory channel to DIMM based system DRAM.
I think that more folks need to fully understand how cache works and how the JEDEC HBM2 standard and those 8, 128 bit channels to each HBM2 stack can independently operate to service the GPU’s Memory controller and the HBCC’s background memory swaps at the same time while the HBCC on Vega is a direct client of the L2 Cache on Vega GPUs. So even the latency out to slower DDR4 system memory can be hidden from the GPU with the GPU feeding mostly from its L1(I$, D$), l2 cache and out to HBM2(HBC) and that’s why Cache was invented in the first place to hide the latency of slower system memory and keep the most often used pages in Cache and swapping out things to and from slower memory in the background while the processor’s, CPU cores or GPU Shader cores, work mostly from cache.