TweakTown NewsRefine News by Category:
It was only earlier last month that NVIDIA had reportedly taped out the GP100 process on TSMC's 16nm process, but now we're hearing that we should expect two GPUs from this, both featuring HBM2.
The first GPU will use 8-Hi stacks, while the other will have 4-Hi stacks, both featuring the Pascal architecture, 16nm process, a 4096-bit memory bus and HBM2. The 4-Hi variation of the GPU will be clocked at 1GHz, which will end up as a consumer-orientated GeForce GPU of some sort (the GeForce GTX 1080 if NVIDIA continues along its naming path) while the second variation has four HBM2 stacks at 1GHz, each with 8-Hi.
For those not in the know, the x-Hi (x representing 4 or 8) denotes the number of stacked DRAM dies, so we'll see the GP100 GPU with 8-Hi HBM2 stacks aimed at the professional market (think Quadro, Tesla), while the 4-Hi HBM2 will be the next-gen consumer GeForce card, expected in 2016.
One of the most exciting GM200-based video cards is nearly here, with EVGA all systems go to release their GeForce GTX 980 Ti KINGPIN video card tomorrow, priced at $850.
EVGA is offering its new card with pre-binned GPUs, which allow consumers to choose their GPU ASIC quality before the purchase is made. Inside, we have the same GM200 GPU that powers the reference GeForce GTX 980 Ti, with the 2816 CUDA cores, 176 TMUs and 96 ROPS. But the Base Clock has been increased from 1000MHz on the reference design, to 1203MHz. This results in the Boost Clock being increased from 1075MHz to 1304MHz. With these clocks, it's certainly not the fastest GTX 980 Ti out there, an award that goes to the ZOTAC GeForce GTX 980 Ti AMP! Extreme Edition.
Powering the EVGA GeForce GTX 980 Ti KINGPIN card are two 8-pin PCIe power connectors joined by a 6-pin PCIe power connector. This will provide the board with up to 450W of power, which is simply insane. EVGA has deployed an extended PCB on the card with a beautiful custom design sporting extra copper filling for smoother power delivery. 12 layers have been stacked together to form the PCB, with 14+3 phase fully digital VRM will provide the card with all of the power it needs to keep it stable during those overclocking sessions.
Where the GPU ASIC quality comes into play, is that consumers can opt for four different models. The first, is the $850 model with 72%+ ASIC quality, $900 for 74%+ ASIC quality, $1000 for 76%+ ASIC quality, or a huge $1050 for 80%+ ASIC quality. These cards aren't for the faint-hearted, but for the true enthusiasts.
Under LN2, KINGPIN himself has clocked the EVGA GeForce GTX 980 Ti KINGPIN card to a huge 2100MHz on the Core, while the RAM has been increased to 2053MHz (up from 1750MHz), providing 400GB/sec memory bandwidth.
Before we exclusively revealed AMD was opting with the Radeon R9 Fury X name for its next-gen, HBM-based video card, we reported on the purported Radeon R9 490X that would arrive in 2016. We knew it would be based on the 16nm process, and use HBM2, but it looks like we were wrong with the name.
During AMD's recent earnings call, President and CEO Lisa Su talked about the company playing around with a couple of its first FinFET designs. Su said that FinFET is going to be very important for the company as we move into the new year. But, she also reiterated that a good design and architecture is just as, if not more important than the manufacturing process - something we definitely agree on.
AMD will be releasing its 'Arctic Islands' architecture next year, its next-generation GPU architecture based on the 16nm manufacturing process, and using the next-gen HBM2. HBM2 will deliver 8GB of HBM2 on a video card, with memory bandwidth of 1024GB/sec - up from the 512GB/sec found on the Radeon R9 Fury X. But what we need to know is: what will AMD call this card? Surely it can't be the Fury X 2, as I'm sure the dual-GPU card coming out later this year will take up the Fury X2 (no space) name. Could AMD release it as the Fury XX? Because I can't wait to see the third iteration: Fury XXX.
On the CPU side of things, we should see FinFET-based Zen CPUs next year, something that will combat the Kaby Lake architecture from Intel that is arriving 2016. AMD's new Zen CPUs will be based on the even-smaller 14nm process, which makes for some incredibly exciting times over at AMD.
AMD will be launching one of its most exciting video cards ever next month, with the High Bandwidth Memory-powered Radeon R9 Nano to launch mid-August, according to company CEO Lisa Su.
Su announced the release window during AMD's Q2 2015 earnings call, where she said: "Fury just launched, actually this week, and we will be launching Nano in the August timeframe". We don't know exactly what to expect, but we do know it's going to be a super-small video card with a TDP of just 175W.
Inside the Radeon R9 Nano, we know that HBM will be used, and that it will have 2x the performance-per-watt compared to the R9 290X, as well as 2x the performance density of the 290X. This will see AMD be the Mini-ITX champion, so if you've got a super-small PC that you need to buy a video card for, the R9 Nano should be the one.
PowerColor (TUL Corporation) has added to its R9 lineup with this 4GB Fury HBM model, said to deliver 60% more memory bandwidth over GDDR5 thanks to this technology advancement.
Tailored towards 4K and VR applications, the PowerColor R9 Fury offers cooling through three 90mm fans, six 8mm heat pipes and a black NI plating clip.
Complete with support for AMD LiquidVR, DirectX 12, Vulkan APIs, AMD CrossFire and FreeSync technology, this card offers a 4096bit memory interface, three DisplayPort outputs and a single HDMI option too. With the clock sitting at 1000MHz, this card looks quite sleek.
One of the main benefits of AMD's Radeon R9 Fury and R9 Fury X cards is their use of High Bandwidth Memory (HBM), and thanks to AMD's relationship with SK Hynix, AMD has 'priority access' to HBM2, according to WCCFTech.
HBM2 is going to be limited in supply at first, the same problems associated with HBM1 right now, and with NVIDIA set to use HBM2 on their upcoming 16nm-based Pascal architecture, AMD will have the one-up on its competitor. AMD should be able to secure most of the HBM2 produced at the time, which is exactly the position AMD wants to be in, come 2016.
If this is true, it is going to force NVIDIA into a corner. NVIDIA needs to jump to HBM2 with its shift to 16nm, and its first GPU based on the Pascal architecture. But, if NVIDIA finds it hard to secure HBM2 because AMD has 'priority access' to the first batches of the super-fast VRAM, this could cause a big issue for NVIDIA's next-gen cards. HBM2 is expected to drive memory bandwidth to 1TB/sec, up from the 512GB/sec possible on the Fury X, and 334GB/sec on the GTX 980 Ti.
Another impressive addition by this video card brand to add to the wish-list is the newly announced Colorful NVIDIA GeForce GTX 980 Ti iGame Ymir-X. Set to offer a big range of features, this card is cooled by a dual-slot, three-fan setup complete with a twin aluminum fin-stack heat sink and six 8mm-thick nickel-plated copper pipes to help draw out heat.
The three fans change speeds with temperature and come coupled with a red LED Colorful iGame logo. Complete with a custom PCB, this 8-pin PCIe powered card uses a 14-phase VRM and silver contact points, doing away with the regular copper offering.
Factory overclocking is just the beginning, with this GeForce 980Ti running at a core speed of 1127 MHz and a 1216 MHz boost with 7 GHz of memory. If all of this isn't enough for you, there's even a toolkit included in the box - but no pricing just yet.
While we were the first site in the world to post triple 4K benchmarks at 11,520x2160 on our Fury X cards in CrossFire, we didn't get our hands on four of them to test. Well, DGLee at IYD.KR has gotten his hands on four of the Fury X cards and thrown them into a 4-way GPU battle against four NVIDIA GeForce GTX Titan X cards in SLI at 4K.
DGLee was running Windows 8.1, an Intel Core i7-5960X, 8GB of DDR4, all on a Rampage V Extreme motherboard. But how were t hose results? Well, take a look below, for all of the 'average relative performance' of the 12 titles run. At 2560x1440, the Fury X cards in CrossFire are able to barely beat the Titan X cards in SLI, but at 4K, the 3- and 4-way Fury X cards really come out into their own and beat the Titan X cards. Even though it's by 15% in 4-way, CrossFire scaling is great on Fury X.
We start off with 4K:
Next, we have 1440p:
We noticed the annoying pump noise on both of our AMD Radeon R9 Fury X samples, but with cards going to consumers with this noise, it's not good. What is good however, is that AMD is finally addressing the issue, shipping cards with 'sound baffling adhesive' that will make the card quieter.
AMD said in a statement: "We have received feedback that during open bench testing a small number of Fury X cards emit a sound from the high speed liquid cooling pump that, while not loud, is bothersome to some users. While the vast majority of initial Fury X owners report remarkably quiet operation, we take this feedback seriously, as AMD's mission is to always deliver the best possible experience to our Radeon customers".
"While not loud" is quite the statement, as it is incredibly annoying. The question is, how did any of these cards make it out of AMD like this? Did they not test the cards at all, and just send them out? The pump noise is very apparent, to anyone sitting near the Fury X when its under load. AMD continued: "AMD Radeon R9 Fury X customers demand and deserve the best, so adjustments in the sound baffling adhesive compound were applied in the assembly of the high speed cooling pump to address the specific sound a few end users experienced as problematic. This improved the acoustic profile of the pump, and repeat testing shows the specific pitch/sound in question was largely reduced through adjustments to the sound-baffling adhesive compound in the pump".
AMD has just launched a new video card for the professional market, with the new FirePro S9170 Accelerator. This new card replaces the FirePro S9150, and has a few new tricks up its sleeve.
For starters, the new FirePro S9170 Accelerator includes the Grenada XT GPU, 32GB of GDDR5, and is aimed at DGEMM (Dense General Matrix) and sever workloads, with OpenCL 2.0 support to boot. With 32GB of GDDR5 on-board, it is one of the most VRAM packed compute cards on the market. NVIDIA's closest card is the Tesla K80 with 24GB of GDDR5.
Spec wise, the AMD FirePro S9170 rocks 2816 stream processors, 44 compute engines, 176 TMUs, 64 ROPs, and a 930MHz on its Grenada XT GPU. The 32GB of GDDR5 is clocked lower than usual, sitting at 5GHz like most professional cards. We have a 512-bit memory bus, resulting in 320GB/sec of available memory bandwidth. FP64/FP32 compute performance sits at 2.62 and 5.24 TFlops, respectively.