siuba

完全錯



Memory bandwidth加1倍, 平均GPU效能加30%
qcmadness 發表於 2015-5-20 23:22

都未出, 你能否定嗎？
我只係帶出 bandwidth 高左唔等於會有好高效能

henry0

amd翻身睇尼舖

polarhei

回覆 11# siuba

等一等, GPU易失憶哦. 另外, 想版縮小一點.  但HBM能否出街已經是大問題.

燕飛

點解會出唔到? 2.5d 個yield應該唔低

Kundera

都未出, 你能否定嗎？
我只係帶出 bandwidth 高左唔等於會有好高效能
siuba 發表於 2015-5-21 16:02

don't mystify
memory bandwidth is always the bottleneck for GPUs as they are throughput oriented.
you cannot double the number of execution units without doubling the bandwidth or else the EUs will be just laying there wasting power. Which is why L1/ L2 cache, address coalescing is implemented in modern GPUs or the eDRAM in Iris graphic. They are measures to maximize throughput when GDDR5 or system RAM is plateaued.

yjfoeg

人地Nvidia降低bus width同時提升效能，AMD就靠bus width去硬爆…

Xeony

人地Nvidia降低bus width同時提升效能，AMD就靠bus width去硬爆…
yjfoeg 發表於 2015-5-21 21:01

但nvidia 行緊 7GHz 喎。amd 行GDDR5 @ 5GHz, HBM 應該行 1GHz wor!


(Sent from HKEPC iPhone app)

Xeony

實物已有，但賣唔賣得出街就唔知了。


Source from" 半準

(Sent from HKEPC iPhone app)