@(Deep Learning)[Engineering]
2017年舊文
Deep Learning Helps Medical Image Diagnosis
More and more work are showing deep learning can help medical images diagnosis, which not only saves health-care costs but also accelerate health-care democratization especially in developing and undeveloped regions.
- Lung Nodule Detection
- Breast Cancer Screening
- Cancer Metastases Detection
- ...
DL Medical Image Diagnosis brings new challenges to platform
Medical image analysis is different from well-studied image analysis problems like ImageNet/Microsoft COCO.
-
Medical images are often bigger
-
The objects are smaller and often subtle
Based on above, Medical AI is not only a computation intensive but also a memory intensive workload.
- Higher data dimension calls for higher-dimension analysis, like 3D model for lung nodule detection.
- Higher data resolution poses higher memory requirements.
The Limitation of GPU-centric solution
For GPU-centric solution’s memory limitation, currently researchers must make compromises, and these compromises finally hurts effects.
- Down sample images to fit GPU memory
hurt algorithm effect
- Split images to fit GPU memory
- Use very small batch (e.g. 1~4) to fit GPU memory
CPU-centric Platforms unleash Medical AI Explorations
Intel Xeon & Xeon Phi supply best capability to handle computation- & memory-intensive workloads and make best flexibility on Medical AI explorations.
Case Study
Resolution Matters in Medical AI
- Batch Size Matters in Medical AI
*Nitish Shirish Keskar, Dheevatsa Mudigere, etc. On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima, 2017.2, https://arxiv.org/abs/1609.04836
A bigger picture
Segmented AI Workloads Call for Diversified Platforms
As AI workloads are getting more and more segmented and thus more and more performance characteristics are disclosed, more and more diversified platforms need to be considered in order to get best effect/performance/cost.
- Medical AI – Memory & Computation Intensive
- Training: Xeon Phi
- Inference: Xeon / Xeon Phi
- Big Data Analytics – IO & Memory Intensive
- Xeon/Xeon Phi (+ high-bandwidth memory + high-performance storage)
- Sparse Machine Learning
- Xeon (+ FPGA)
- ...
AI Cloud Services Can and Should Leverage a Broader Platform Portfolio
CSPs have convergence power to handle the broader platform platform. It's an opportunity. In PC era, it's hard to do that because every user need handle the diversity of the HWs and distributions, now CSPs can handle it and deliver an unified service to users transparently.
后記 (2022/11/08)
5年之后再回頭看這個PPT，整個領(lǐng)域似乎變了，又似乎沒變。
變了的是，NV GPU在內(nèi)存帶寬和內(nèi)存容量上持續(xù)改善，到H100已經(jīng)演變成聚合帶寬為3.35 TB/s的80GB大容量HBM3了。這些改變，其實是因應(yīng)推薦系統(tǒng)、3D分析、大規(guī)模語言模型(LLM)這些應(yīng)用的需求，逐漸拓寬GPU的對這些領(lǐng)域的適用性區(qū)間的努力。變了的還有，在短短5年的時間里，CUDA生態(tài)已經(jīng)發(fā)展成深度學(xué)習(xí)甚至是高性能計算領(lǐng)域的by-default，現(xiàn)在很多data scientists也會讀一些、寫一些甚至改一些CUDA代碼了，圍繞著這個生態(tài)也催生了Triton這種更利于小白data scientist的CUDA代碼生成工具。傳統(tǒng)的圍繞CPU編程的高性能計算生態(tài)圍墻被跨過了。可見生態(tài)從來是power的附庸，只要你有壓倒性的power，整個生態(tài)就愿意去適應(yīng)你、完善你，最后變成你的生態(tài)。Intel因其龐大的組織和決策結(jié)構(gòu)，拖累了其在新領(lǐng)域快速創(chuàng)新并建立護城河的能力，被顛覆性創(chuàng)新者推進了創(chuàng)新者的窘境，成為創(chuàng)新者的窘境的另一個生動實例。沒有變的是，application還在那兒，application的需求還在那兒，NV似乎正在從屠龍少年變成龍，我們正見證舊瓶新酒的商業(yè)宣傳，關(guān)于“GPU can do all”，這個故事Intel也講過，當(dāng)年的主語是CPU。這符合一雞多吃的商業(yè)利潤最大化的動機，是個無可厚非的商業(yè)故事。
但回到技術(shù)的語境里，當(dāng)我們拋棄其他的上下文，有時候我們會覺得目前NV的努力可能只是緩解措施，并沒有從根本上解決問題。在Xeon Phi被Intel cancel之后，業(yè)界依舊認可"scale + vector + tensor + spatial"的聚合微架構(gòu)是有前途的最終解決路徑之一。之前，我們寄希望于Xeon Phi用"Atom + AVX512QFMA"及其后續(xù)演進在原編程模型的框架下完成這次新計算IP的納入，最終完成統(tǒng)一與收斂，維護昔日帝國的輝煌。而如今，新的先驅(qū)如Tenstorrent以及Esperanto們，以"RISC-V + domain specific extensions"為新的武器，繼續(xù)前行在這條道路上。
如果我們是牧村浪潮(Makimoto’s Wave)的信徒，那我們會相信converge，是誰摘得桂冠，我們需要耐心等待。
是為記。
