b*********f
发帖数: 1585 | 1 Today’s opening keynote at NVIDIA’s GTC had a number of interesting
announcements (we’ll get to those in other articles) but the “star” of
the show was the upcoming Tesla K20.
Destined to spend its life in HPC systems and supercomputers, the K20 is
based off on a Kepler GK110 core which is billed as the “largest, fastest”
processing core ever designed. It features a threefold increase in double
precision performance over a comparable Fermi core and also incorporates
high level dynamic parallelism, and Hyper Q technology.
Hyper Q allows these new Kepler parts to process up to 32 concurrent work
queues (versus the single work queue for Fermi), allowing its massively
powerful parallel abilities to stay fully engaged, optimizing efficiency.
From the information that Hardware Canucks has seen, the GK110 core is
geared strictly towards high performance computing solutions rather than the
desktop GeForce product space since it incorporates functionality that won
’t be necessary for gaming or low level GPGPU compute. But that doesn’t
mean it can’t be scaled according to NVIDIA’s needs in other markets (
namely for GeForce and Quadro products) but for the time being it looks like
NVIDIA is targeting their flagship at high margin niches.
According to NVIDIA, this Kepler derivative will be the first highly
virtualized GPU delivered to market, allowing it to render frames and stream
them instantaneously without being connected to an external display. This
means it can be used in cloud computing applications, possibly allowing for
game streaming over a broadband internet connection to a remote device. As
with the GeForce-branded chip, this one will be very efficient, making it
possible to deploy Tesla-based clusters on much larger scale.
Stay tuned for more info as we can properly dissect the various nuances of
this new chip. | b*********f
发帖数: 1585 | 2 Nvidia is announcing two products today, two Kepler based HPC / GPGPU cards
- K10 and K20.
NVIDIA Tesla K10
Tesla K10 dedicated to seismic analysis where most valuable resource is
bandwidth. It has 3x single precision of today’s fermi tesla and 1.8x the
memory bandwidth.
NVIDIA is preparing two Tesla GPUs, one is Tesla K10, based on dual-GK104
chips, which is to be devoted to high efficient computing in oil and gas
exploration and defense industry. Card is based on dual-GK104 GPUs and has a
memory bandwidth of 320 GB/s.
The Tesla K10 packs two Kepler-based GK104 GPUs and 8 GBs of VRAM.
Second Tesla, K20 is focused on double precision: 3x double precision of
fermi, includes hyper q, dynamics parallelism, for stuff like physics,
quantum chemistry, computational finance.
GK110:*Rumor* Actual chip is to feature 16 SMXs (streaming multiprocessors)
with 192 CUDA cores each. This gives 3072 CUDA cores in total. *need a
secondary source to confirm*
GK110 GPU which is coming in Q4 of this year. | b*********f
发帖数: 1585 | 3 期待GK110的geforce版，估计要等明年了，我要好好攒钱去买kepler卡皇。 | b*********f
发帖数: 1585 | 4 GK110: 71亿晶体管 2880个CUDA 15个SMX
另外值得看看AMD的反击。 | s****c
发帖数: 11300 | 5 够呛 这东西游戏性能可能不咋地 双精度对于游戏基本是没有用的 散热决定了频率也
跑不了太高
记得有一个超算build的时候买游戏显卡都要降频跑
nv走了几年的弯路 以为一个芯片就能通吃 看来是不对的
【在 b*********f 的大作中提到】
: 期待GK110的geforce版，估计要等明年了，我要好好攒钱去买kepler卡皇。
| f*******e
发帖数: 8974 | 6 前两天不是有人发nvidia开普勒游戏卡阉割了gpgpu能力吗？两个市场完全分开了 | b*********f
发帖数: 1585 | | a***e
发帖数: 27968 | 8 看来工艺还是不能保证16个都好
7970 42亿晶体管，7990估计就相当的DP应用了
【在 b*********f 的大作中提到】
: GK110: 71亿晶体管 2880个CUDA 15个SMX
: 另外值得看看AMD的反击。
| h****n
发帖数: 131 | 9 本来就只有15个呢？ 5X3
【在 a***e 的大作中提到】
: 看来工艺还是不能保证16个都好
: 7970 42亿晶体管，7990估计就相当的DP应用了
|
|
