Search the Community
Showing results for tags 'hbm2'.
Found 22 results
-
Videocardz has just spotted and AMD card in CompuBench sporting 16GB of VRAM, and a clock speed of 1600Mz; along with 64 Compute Units. It looks like another Vega card has cropped up, and has double the VRAM and 400Mhz more on the clock compared to the last TimeSpy leak; where an AMD card had 1200Mhz core clock and 8GB VRAM at 700Mhz ( HBM2 is rated up to 1000Mhz). If clocked at 1600Mhz, it would indicate 13.1 TFLOPS in Single Precision which is not too shabby. https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock https://compubench.com/device.jsp?benchmark=compu20d&os=Windows&api=cl&D=AMD+6864%3A00&testgroup=info Now before anyone goes and gets a bit hyped we need to look at the VRAM for one. AMD's supplier is SK Hynix, and they have no 8GB stacks ready for production this year it seems; let alone 2.0GBps modules which two are needed of to obtain the 512GB/s Vega is supposed to have. If that's still the case right now, this card found in CompuBench is likely either an Instinct card, or a Radeon Pro; which could allow them to use 4 dies of HBM2, as opposed to only two. Cost is not an option in the professional space, as it is for normal consumers; and HBM2 is rather expensive. We've also seen Raja hold up the Vega die showing 2 HBM2 dies, but that could only be in relation to the consumer gaming cards. Source: SK Hynix Catalogue for 2017 https://www.skhynix.com/static/filedata/fileDownload.do?seq=374 It still a ways off before Computex and we get any solid information directly from AMD sadly. It's all true folks! We have out 16GB, 1600Mhz Vega Card, the Radeon Vega Frontier Edition. http://pro.radeon.com/en-us/vega-frontier-edition/
-
Source: http://www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it So basically, HBM2 alone on Vega is costing about $150, with the interposer costing another $25 which pegs the cost of the memory system in Vega at $175. This is really, really high, especially considering that on Vega 56 this would comprise nearly HALF the cost of the GPU. This also probably explains AMD's choice to go for 2 HBM stacks instead of 4 on Vega; 4 probably would've made the memory alone cost almost $400. The VRM setup probably costs another $30-$40 ish, which would make the cost of the video card excluding the gpu already a little over $200. Say the GPU die costs $75, and AMD really isn't making a huge profit on 56 or 64. However, AMD really had no choice and had to go with HBM. HBM leads to lower power consumption which in turn lets AMD push clock speeds a bit higher, and when AMD is at a significant performance/watt deficit compared to Nvidia, they need everything they can get to catch up. At the same time, it has to be questioned whether this really was worth doing considering that going HBM2 costs AMD approximately three times as much compared to GDDR5. Is it really worth paying $100 more just to get 20W more headroom which may lead to 4-5% more performance? -
It’s a quadro gp1000 for 7000$ It have 16gb HBM2 and not a single review or tear down is up? What’s goin on? https://www.nvidia.com/en-us/design-visualization/quadro-store/ -
the last couple of month i heard and read a lot vega is great for mining and thanks too this an can't really afford one. now my question is is it cause of hbm2 memory in vega or is the internal structure with calc cores simply better for mining? I think at some point i read ethereum is more memory demanding in mining than the actual gpu(?) now "if" there would be a vega with gddr5(x) would it still be so great for mining or dos the hbm2 a big thing here? If this topic would be better in another forum please move it.
-
Apparently, AMD uses two chip packaging locations for the Vega GPU for the two main SKUs -- Vega 64 and Vega 56. According to several official slides from AMD, there are minor differences in the physical packaging, depending on the country of origin: AMD uses both Samsung & SK Hynix HBM, Samsung on Vega FE & Vega64, SK Hynix on Vega56 & Vega64 South Korea packed Vega 10s don't have the filling between/around chips Taiwan packed Vega 10s have filling between/around chips The height difference between GPU & HBM is only on the South Korea packed Vega 10s, it's 40 µm, no issues when thermal paste fills the gap Taiwan packed Vega 10s are 0.1mm shorter than South Korea Packed In retail, Vega 64s are supposed to have filling, Vega 56's not Source (french): http://www.tomshardware.fr/articles/radeon-vega-package-gpu-hbm,1-65141.html
-
Well it's not like we didn't see this coming. Source: http://www.guru3d.com/news-story/upcoming-geforce-gtx-cards-use-gddr5x-not-hbm2.html Personally I don't think it's going to matter at all. While HBM is great, it's way over hyped. This is my first news post, ever. If I did something wrong please don't crucify me. If I did something wrong, my apologies.
-
hi everybody i need some help I am building a content creation pc for one of my friends and I was wondering if I should be using the frontier edition vega card or the Xp for the pc he will be using primmer pro i am guessing that vega will be better but I just wanted to make sure thanks
-
I just got my Radeon VII yesterday, and I'm really enjoying it. It replaced my 1080 Mini, and so far it seems fine. The drivers were the major worry, but my experiences have been pretty good. If anything, it's response is faster for recovering from sleep, and during restarts than my 1080. I've had no compatibility issues with games so far, and it scores well on Passmark. I was wondering if anyone had any other experiences with this card, and if yours differs from mine, how so and what should I look for? I know there aren't many of these things out there, and I was lucky to get one, but I figured someone on the forums must have one.
- 1 reply
-
- 1
-
-
- amd
- radeon vii
- (and 1 more)
-
So, AMD is going to be releasing their Vega Instinct Graphics Cards with 32 Gigabytes of HBM2. GDDR6 is cheaper than HBM2 but I've heard HBM2 is faster. Can anyone highlight the differences? i.e: Speeds Bandwidths Costs Redundancy etc. (Note: Pardon me if there is any error in my post *Edited because of mistake)
-
Samsung has now started producing its improved hbm2 8gb chips After the disappointing launch of hbm2 to the market with long delays and low clocks, it seems that hbm2 finally meets the original spec and even surpasses it. with this new generation hbm2 can reach 2.4Gbps, which is 0.5Gbps faster than the fastest hbm stacks on the consumer market. they have also been able to achieve this at a lower voltage, before they needed 1.35v to achieve 2Gbps now they only need 1.2v which might mean some improvements in overclockability. Samsung is calling this new chips Aquabolt. opinion: This would have been a great addon to a 12nm vega 10 part, or a small vega replacement part as it makes single stack products much more competitive with gddr5/6 thanks to interposer being smaller this levels of performance are higher than max overclock that can be achieved in current vega cards (they top out at around 1150-75) while being at stock at lower voltages,i'm impressed. 2 stack config changes: currently new on vega clock 950 1200 mhz mem bus 2048 2048 Bandwidth 486 614 Gbits/s source: https://overclock3d.net/news/memory/samsung_has_started_mass_producing_2nd_generation_8gb_hbm2/1 https://news.samsung.com/global/samsung-starts-producing-8-gigabyte-high-bandwidth-memory-2-with-highest-data-transmission-speed
-
I haven't seen any news regarding Volta lately, possibly because Vega was so underwhelming. But since Vega uses HBM2 memory, and supposedly it is "better" from what I've heard, will Volta utilize HBM2, or will Nivdia stick with GDDR5X for a while?
-
http://www.pcper.com/news/Processors/Intel-Launches-Stratix-10-FPGA-ARM-CPU-and-HBM2 Shipping is said to begin by the end of the year (Kirzanich, IDF 2016) Sampling has begun. Stratix 10 SOC: 4 Partly customized ARM A53 cores 4 stacks of HBM2 at 1TB/s in bandwidth (8 or 16GB total not given) Intel/Altera FPGA 5.5 Million Logic Elements HyperFlex Optimization for registers and critical paths 5x the logic density vs. Altera's previous products 70% less power Onboard ECC and end to end data encryption/integrity checks 10TFlops Single Precision DSP (can do 5 DP or can do a mix b/c it's an FPGA) 80 gigaflops per watt or 125W for the whole enchilada (misprint in the article would mean 125Gigawatts, so obviously it's flipped). Dimensions not yet provided, but we do have a graphic. It's way smaller than Fiji. I would be very angry if I was IBM or Xilinx right now. This is one serious piece of silicon. For comparison, Xilinx's top of the line FPGA today: https://www.xilinx.com/products/silicon-devices/fpga/virtex-ultrascale-plus.html 3.78 million logic elements (Intel has a 45.5% advantage here) Uses (if I'm reading this correctly) quad-channel DDR4 2666 for main memory, but has a total internal fabric bandwidth of 1.05TB/s and nearly 1Tb of ethernet fabric (It's built for high-speed derivatives trading, which is also why it costs $40,000 to buy ONE of them. It's also used for radar/sonar and other signal processing, so there are many other performance metrics to look at, but that's the gist of it. Update: 80GFlops/Watt confirmed by Altera representative. https://youtu.be/zI_thFZQzgc?t=2m9s
-
Source: http://www.anandtech.com/show/10527/sk-hynix-adds-hbm2-4-gb-memory-q3 Specifications of SK Hynix's HBM2 Memory Chips Part Number Density Data Rate Bandwidth per Stack Packaging Feature Availability H5VR32ESM4H-20C 4 GB 2.0 GT/s 256 GB/s 5mKGSD 4Hi stack, VDD/VDDQ=1.2V Q3'16 H5VR32ESM4H-12C 1.6 GT/s 204 GB/s Looking at this, we can probably assume that SK Hynix has already begun mass production of hbm2. With them stating availability is q3 (right now), they probably had to have begun mass production by now or they wouldn't have enough chips available for this quarter. It looks they they're starting off with 4gb stacks of hbm2, which is probably what will be used on the flagship vega GPU, so it is highly possible that one of these two chips will be what's used on vega. As for which one AMD will choose, we do not know yet. By utilizing four of these though AMD will be able to put 16gb of VRAM at either 816 or 1024 GB/s, which will be pretty extreme. SK Hynix will probably eventually start manufacturing 2-Hi and 8-Hi HBM2 stacks which would make a single GPU capable of having 32 gb of memory at 1024 GB/s! This is cool stuff, and it'll be interesting to see how different companies will utilize HBM2 in various different products. HBM1 wasn't much of a success, but hopefully HBM2 will be.
-
We all know that Nvidia revealed some GP100 info based on the P100 Tesla product launch, but two sites actually reviewed the graphics and diagrams and have identified... that GP100... isn't the full blown GP100. Firstly from Guru3D -> http://www.guru3d.com/news-story/nvidia-gp100-gpu-architecture-recap-full-gpu-has-3840-shader-processors.html And corroborating the above, TechPoweUp -> https://www.techpowerup.com/221641/nvidia-pascal-gp100-silicon-detailed So definitely, the next Gen Titan will most likely be a very serious beast to content with. Although it is not certain if the next Ti will be GP100 or a separate GP104 SKU, but we all hope for the GP100 version
-
Article Link: http://arstechnica.co.uk/gadgets/2016/04/nvidia-tesla-p100-pascal-details/ Article Highlights: Full fat baby! Argh! I want to see it coming out now. Damn tease. HBM 2 at least on professional cards! 7->10 teraflops of single precision improvement. 4096 Bit Memory Bus. Nice. That's an interesting surprise. There is still top end to unlock. Hmm, is this the first whisps of a future Pascal Ti ?! Looks like Pascal will have a high clock increase like Maxwell did over the previous gen. Nvidia displaying their AI and VR first approach. Sounds like a beast system for developers. All in all very exciting! Can't wait to see new systems with new architecture! Lets just hope we get to see it all in 2016.
-
Hey there, I am wondering about the performance differences between the different RAM Types. Now of course dependent on what you do, "performance" is defined pretty differently. So for the purpose of this posting I will use performance as equal to "performance in the standard Firestrike Benchmark of 3DMark". The reason for my quesion is mostly because I'd like to be able to have a sense of how much of a step GGDR5 to GDDR5X is (in the upcoming GTX 1070 and 1080) and also how much of a boost HBM2 will be in the follow up generation. Mostly "better" VRAM also comes with an increase of other specs. So the GTX 1070 and 1080 will have more differences besides of the VRAM. But can anyone estimate on what difference it would make if you just exchanged upgraded the VRAM (and it's connections that it could actually be used). And what about HBM? Well, HBM2 is not used yet from what I know. But what is the estimated boost from HBM2 over HBM? If you leave out that you could put more RAM in the same space. In this instance, would you compare the R9 380X (which has 4GB of RAM) to the R9 fury for example and say that the difference is basically the boost due to HBM Memory? Of course there is hardly an exact answer. But I'd appreciate your best guesses and extrapolations from the know examples we have with current cards and data. Will a cheap HBM2 card blow a top GDDR5X card away just because of the VRAM? Or is the boost and the role of this step rather limited and only plays a small role in overall performance boost? -
From Altera Not to familiar with this FPGA SOC stuff so don't know what to comment on. Still, it's nice to see technology going forward. http://wccftech.com/intel-altera-offering-hbm2-fpga-sip-stratix-10/
-
source: https://translate.google.com/translate?hl=en&sl=auto&tl=en&u=http%3A%2F%2Fwww.hwbattle.com%2Fbbs%2Fboard.php%3Fbo_table%3Dnews%26wr_id%3D15876 this rumor comes straight after Samsung officially announced they have begun mass production of 4GB HBM2 stacks: http://linustechtips.com/main/topic/530231-samsung-begins-mass-production-of-4gb-hbm2-stacks/ the interesting part about this: according to a quote from Digitimes Korea, Hynix hasn't yet started mass producing HBM2 and are expected to do so in August I know, google translate .. don't blame me for it if this news is to be believed and AMD sourced HBM2 only from SK-Hynix, their video cards will be late to the party - I would not want to be in AMD's shoes if the rumors are true AMD's exclusivity deal with Hynix turns and bites them in the arse .. would be hilarious -
source: http://news.samsung.com/global/samsung-begins-mass-producing-worlds-fastest-dram-based-on-newest-high-bandwidth-memory-hbm-interface the 4GB stack is manufactured in 20nm and was achieved by stacking four 8 gigabit cores on top of a buffer die; they are vertically interconnected by TSV (Through Silicon Vias) holes and microbumps - over 5000 per single 8gb die Samsung's DRM package offers 256GBps of bandwidth - twice as much as HBM1 and ~7 times more than a single 4Gb GDDR5 chip Samsung also plans to introduce 8GB HBM2 stacks, specifically to be used in GPU production, during this year - no concrete date --- if nVidia was planning to source 8GB HBM2 from Samsung, looks like they'll have to wait a bit longer but this doesn't mean they can't use multiple 4GB stacks - depending on how exactly they engineered the memory controller inside the GPU
-
Nvidia is getting ready the big guns for 2016, featuring the top of the line big Pascal GPU with HBM2 memory and ultra-fast proprietary NVLink multi-GPU interface. The model shown at the GTC presentation is intended for high-density HPC applications. Source: http://wccftech.com/nvidia-pascal-nvlink-200-gbs-interconnect-hbm2-stacked-memory-1-tbs-bandwidth-powering-hpc-2016/
-
source: https://www.jedec.org/news/pressreleases/jedec-updates-groundbreaking-high-bandwidth-memory-hbm-standard --- who is JEDEC? Solid State Technology Association, formerly known as the Joint Electron Device Engineering Council (JEDEC), is an independent semiconductor engineering trade organization and standardization body JEDEC was founded in 1958 as a joint activity between EIA (Electronic Industries Alliance) and the National Electrical Manufacturers Association (NEMA) to develop standards for semiconductor devices. NEMA discontinued its involvement in 1979 --- what is HBM? High Bandwidth Memory is a type of DRAM jointly developed by AMD and Hynix as a competing product to Intel's and Micron's Hybrid Memory Cube HBM has been adopted as a standard by JEDEC in 2013 under JESD235 nomenclature the 1st products to utilize HBM were the AMD's Fury lineup utilizing the Fiji GPUs
-
source: Fudzilla via PCPer --- 17bil transistors, that huge, it's little more than double the count for Titan X with it's 7,1 billion, with a B any bets? who's gonna be the 1st to market with a HBM2 product: AMD or nVidia? also, 32GB of VRAM? yeah .. that's not gonna be a desktop product, rather a Tesla accelerator or a Quadro
