AMD Next Horizon Live Blog: Starts 9am PT / 5pm UTC
by Ian Cutress on November 6, 2018 8:00 AM EST- Posted in
- CPUs
- AMD
- Enterprise CPUs
- Vega
- 7nm
- EPYC
- Rome
- Next Horizon
11:54AM EST - AMD's 2019 is set to be full of 7nm products, and on the back of AMD's New Horizon event on 2016, today it is hosting part two: Next Horizon. In CEO Dr. Lisa Su's welcome letter, it states that AMD is set to 'enter a new chapter in [its] journey to deliver the datacenter of the future'. I'm here in San Francisco to get all the details on what appears to be discussions and presentations about the next generation EPYC and 7nm Vega, as well as customer responses and deployments about AMD's current datacenter portfolio.
11:55AM EST - I'm now sat down ready to begin
11:59AM EST - I'm sat next to Paul from Tom's and Charlie from SemiAccurate
12:00PM EST - Our schedule mentions Radeon Instinct, EPYC, and ROCm
12:00PM EST - This is truly going to be a datacenter event
12:01PM EST - Opening video
12:01PM EST - Just quoted AnandTech in the video
12:02PM EST - Dr Lisa Su to the stage
12:02PM EST - 'Excited to talk about the next horizon of computing'
12:02PM EST - Two years first discussed Zen
12:02PM EST - A lot has happened since
12:03PM EST - It's all about the journey
12:03PM EST - Four key tenets
12:03PM EST - 1/ High Perf, 2/ Great Products, 3/ Ambitious Goals, 4/ Undaunted Determination
12:03PM EST - Be incredibly ambitious
12:04PM EST - Focusing on great tech and IP: High perf Graphics and high perf compute
12:04PM EST - Coming with solutions and unique differentiations
12:04PM EST - adding value to the ecosystem
12:04PM EST - There's a lot still to do
12:04PM EST - We love all of our markets - consumer and gaming too
12:04PM EST - lots of innovation
12:04PM EST - but the focus today is in datacenter
12:05PM EST - Lisa got into to Semiconductors because she liked to solve problems
12:05PM EST - Need high perf CPU and GPU to solve modern problems
12:05PM EST - The datacenter is evolving
12:06PM EST - more and more devices being connected requires infrastructure
12:06PM EST - Workloads that are driving computing are growing exponentially
12:06PM EST - An incredible need for more computing horsepower
12:06PM EST - Workloads are very different to those from 10 years ago
12:06PM EST - AMD is very excited
12:06PM EST - The DC represents an incredible opportunity
12:07PM EST - $29b market in 2021 for AMD Total Addressable Market
12:07PM EST - The GPU/Accelerator environment is growing a lot
12:07PM EST - A strong double digit growth over the foreseeable future
12:07PM EST - AMD's approach is different to the competition
12:08PM EST - Not looking at incremental changes
12:08PM EST - Some workloads require specialisation
12:08PM EST - Aggressive roadmaps. Made some important bets
12:09PM EST - Products today are results of bets and decisions made 4-5 years ago
12:09PM EST - Influence for the future
12:09PM EST - How we connect in the system and bring customers more performance and more capabilities
12:09PM EST - This is how AMD approaches the DC
12:09PM EST - 15 months since EPYC launch
12:09PM EST - First Zen processor
12:09PM EST - reintroduction to the DC
12:10PM EST - Really pleased about how customers have adopted EPYC
12:10PM EST - Can get more performance at every price point
12:10PM EST - More flexibility at every price point
12:10PM EST - Choose the right config for every situation
12:10PM EST - Bringing down TCO
12:10PM EST - How everything comes together
12:10PM EST - Also Radeon Instinct
12:11PM EST - MI25 with 25TF perf
12:11PM EST - Ideal for ML and Cloud Gaming
12:11PM EST - Also ROCm
12:11PM EST - ecosystem beyond the hardware
12:11PM EST - software ecosystem is critical
12:11PM EST - Making sure we have the guts of it all
12:12PM EST - AMD momentum on EPYC and Instinct
12:12PM EST - Great first 15 months
12:12PM EST - Every major OEM and ODM with EPYC platforms
12:12PM EST - SIs and Partners supporting CPU / GPU platforms
12:12PM EST - Also cloud deployments
12:13PM EST - Azure, Baidu, Tencent, Oracle, Dropbox
12:13PM EST - Cloud instances made available to a large number of users
12:13PM EST - GPU with Alibaba, Baidu, AWS
12:13PM EST - Momentum is clearly here
12:13PM EST - Today announcements
12:13PM EST - Technology and demos
12:14PM EST - Major new announcement
12:14PM EST - Goal with EPYC is to reach the most number of users possible
12:14PM EST - Must be partnered with cloud #1
12:14PM EST - AMD and AWS
12:15PM EST - Incredible ride with AWS
12:15PM EST - (If you're reading this on Youtube via a 'livestream', please head to AnandTech.com to get the proper story)
12:15PM EST - AWS has a long term view
12:15PM EST - Listening to AWS customers
12:16PM EST - Security, reliability and performance is important
12:16PM EST - AWS customers want choice
12:16PM EST - AWS wants to support every single workload out there
12:16PM EST - Also cost
12:16PM EST - Every customer is interested in cost
12:16PM EST - AMD is helping deliver lower pricing
12:17PM EST - R5A, T3A, and one other coming in a few weeks
12:17PM EST - Most popular instance types
12:17PM EST - Some instances available today
12:17PM EST - R5a, M5a today, T3a soon
12:17PM EST - Use the same scripts and APIs as you do today
12:18PM EST - Can instantly save 10% on compute costs
12:18PM EST - No change in software to current customers
12:18PM EST - Many customers are running large databases on M5 can use M5a
12:18PM EST - Hadoop
12:18PM EST - Exceptional Perf per dollar
12:19PM EST - Now for the next horizon
12:20PM EST - New Zen 2 High Perf core
12:20PM EST - 7nm
12:20PM EST - Modular system design
12:20PM EST - The early pieces of Zen 2 today
12:20PM EST - Not only CPU but also system design
12:20PM EST - All very innovative
12:20PM EST - Betting big on 7nm
12:20PM EST - Papermaster to do that presentation
12:21PM EST - Also David Wang will present about designing GPUs for the cloud
12:21PM EST - High-Perf compute engines
12:21PM EST - Additional specs will be coming for MI60
12:21PM EST - First 7nm GPU in the industry
12:21PM EST - For cloud applications
12:21PM EST - Also hearing from Forrest Norrod about Rome
12:22PM EST - First preview of Rome
12:22PM EST - Vision where Rome fits in the industry
12:22PM EST - Incredible capabilities
12:22PM EST - AMD's Total Commitment to the Datacenter
12:22PM EST - AMD loves to solve hard problems
12:22PM EST - It is about what we can do for our customers
12:23PM EST - Mark Papermaster to the stage
12:24PM EST - The CPU journey
12:24PM EST - Zen is a new phase of high perf
12:24PM EST - Scalability of FinFET in 14nm and 12nm
12:24PM EST - Focused a lot of energy into Infinity Fabric
12:24PM EST - Important for scalability
12:25PM EST - Zen had +52% IPC improvement
12:25PM EST - It showed that with focus, AMD can deliver
12:25PM EST - AMD committed to a strong roadmap
12:25PM EST - leapfrogging roadmap with zero product gap
12:25PM EST - That's why today is so important
12:26PM EST - Zen 2 is now sampling, Zen 3 is on track
12:26PM EST - Zen 2 to market in 2019
12:26PM EST - Always two full design efforts in flight
12:27PM EST - AMD is stable and has been executing to the called play
12:27PM EST - Bringing leadership to the most demanding emerging workloads
12:27PM EST - Focusing on execution and delivery
12:27PM EST - The roadmap is as important as the product for enterprise customers
12:27PM EST - It's a gamechanger for AMD
12:28PM EST - Looking back on AMD's high performance in the market with Zen
12:28PM EST - Ryzen, Threadripper, EPYC, Ryzen Mobile
12:28PM EST - Driving changes with Ryzen
12:28PM EST - Leveraging investments in hidden gems
12:29PM EST - Infinity Fabric and modularity
12:29PM EST - Ryzen has been awarded 700 recognitions since launch
12:29PM EST - This is a phenomenal time for transition to AMD
12:30PM EST - AMD is committed to generational improvement
12:30PM EST - Starting with 7nm
12:30PM EST - Picking the right technology is phenominally important
12:30PM EST - AMD made a huge bet on 7nm
12:31PM EST - For AMD's customers, 10nm wasn't the right note. 7nm is set to be a long sustained node in the industry
12:32PM EST - The benefit of the lower power and die area is worth it. Deep partnership with TSMC.
12:33PM EST - 7nm offers 2x density, 0.5x power (at same perf), 1.25x perf (at same power)
12:34PM EST - 7nm delivers tremendous value
12:35PM EST - 7nm is now leadership - the competition can't compete
12:39PM EST - First to market with 7nm in the DC
12:46PM EST - Our CMS is having issues
12:47PM EST - Industry most leading edge technology
12:49PM EST - TSMC has been working closely with AMD
12:52PM EST - Zen 2
12:52PM EST - World's first 7nm High Perf x86 CPU
12:52PM EST - CPU Core Execution Enhancements
12:52PM EST - 2nd Gen IF
12:52PM EST - More Security Elements
12:52PM EST - Modular Design
12:53PM EST - Improved pipeline, DOuble loading point and load store
12:53PM EST - Doubled core density
12:53PM EST - Half energy per operation
12:53PM EST - Improved branch predictor
12:53PM EST - Better instruction pre-fetching
12:53PM EST - Re-optimized instruction cache
12:53PM EST - Larger op-cache
12:53PM EST - FP with to 256-bit
12:53PM EST - Doubled load/store bandwidth
12:53PM EST - Increased dispatch/retire
12:53PM EST - Maintained high throughput modes
12:53PM EST - Security
12:53PM EST - Memory Encryption with Increased Flexibility
12:54PM EST - Hardware enhanced spectre mitigations
12:54PM EST - AMD will stay focused
12:54PM EST - Zen introduced a multi-chip approach
12:54PM EST - Enabled configurability, increased peak compute
12:54PM EST - Zen 2 Each IP is its optimal technology
12:54PM EST - 14nm IO die
12:54PM EST - 7nm CPU chiplets
12:54PM EST - Optimized IO die improves latency and power
12:54PM EST - Revolutionary new approach
12:54PM EST - Others in the industry will adopt this approach
12:54PM EST - Perf is holistic - the process technology and design all helps
12:54PM EST - Seizing the opportunity
12:54PM EST - Zen 4
12:54PM EST - Momentum is now on AMD
12:56PM EST - Now talking GPU
12:56PM EST - Radeon MI60 optimized for datacenter
12:57PM EST - World's first 7nm GPU
12:57PM EST - Flexible Vega architecture
12:57PM EST - Leading edge memory
12:58PM EST - Industry's only hardware-virtualized GPU
12:58PM EST - 13.2B transistors, 331 mm2
12:58PM EST - Double density, 1.25x perf at same power
12:59PM EST - 50% lower power at same freq
12:59PM EST - FP32 and GP64
01:00PM EST - Special ML operations for training and inference
01:00PM EST - end-to-end ECC
01:01PM EST - 32GB of HBM2
01:01PM EST - 1TB/s bandwidth
01:02PM EST - Enterprise class reliability
01:02PM EST - PCIe 4.0 !
01:02PM EST - 64 GB/s Bidirectional CPU to GPU
01:03PM EST - Infinity Fabric GPU to GPU at 100 GB/s per link
01:03PM EST - Infinity Fabric GPU to GPU at 100 GB/s per link
01:03PM EST - Without bridges or switches
01:03PM EST - Connected in a ring
01:03PM EST - Helps scaling multiGPU
01:04PM EST - Hardware virtualization without software overhead
01:04PM EST - data isolation
01:04PM EST - Many VMs per GPU
01:04PM EST - Scale up with multiple GPUs
01:04PM EST - One GPU to 16 VMs, or one VM to 8 GPUs
01:04PM EST - No licence fee
01:05PM EST - Comes with the hardware
01:05PM EST - ROCm open source software
01:05PM EST - ML framework
01:05PM EST - Open source enables community participation
01:06PM EST - ROCm works on NVIDIA GPU and Intel CPU
01:06PM EST - Announcing ROCm 2.0
01:06PM EST - Comes with MI60
01:07PM EST - Latest updated ML Frameworks and Math libraries
01:07PM EST - Dockers and Kubernetes support
01:07PM EST - Upstreamed for Linux Kernel
01:07PM EST - That's pretty important
01:08PM EST - Working with Google on TensorFlow, Baidu on PaddlePaddle, Facebook with Caffe/PyTorch
01:11PM EST - Performance
01:11PM EST - 8.8x DGEMM faster
01:21PM EST - Demo time
01:22PM EST - AMD CPU + GPU for ML difficult problems
01:22PM EST - 118 TOPS INT4
01:24PM EST - 14.7 TFLOPS FP32
01:24PM EST - 7.4 TGLOPS FP64
01:29PM EST - Ryan here: apologies for the site issues. Apparently you're all very eager to hear about Zen and Vega
01:31PM EST - Meanwhile AMD is on a 20 minute break. So here's a good time to refill your coffee
01:53PM EST - Here we go
01:53PM EST - '64 cores' - was that an announcement
01:53PM EST - 33% better memory bandwidth
01:54PM EST - Forrest Norrod on stage
01:54PM EST - video time
01:55PM EST - Going through EPYC state of affairs and use cases
01:56PM EST - Differentiated services
01:56PM EST - Advantages of EPYC with cores, perf, scalability
01:56PM EST - even single socket deployments
01:57PM EST - Azure has LV series for storage/IO intensive
01:57PM EST - Azure has Hb series for high perf
01:57PM EST - Amazon is going broad
01:57PM EST - M and T instances for general purpose
01:57PM EST - R instances for memory optimized perf
01:58PM EST - Oracle offers virtualized and bare metal
01:58PM EST - different deployment models
01:59PM EST - Up to 60% lower cost per core hour
01:59PM EST - 32 cores in a single socket is really disruptive for managed services
02:00PM EST - Baidu runs single socket EPYC replacing legacy dual socket platforms
02:00PM EST - Single socket efficiency is the right match for them
02:00PM EST - Tencent is offering AMD internally and externally for differentiated TCO
02:01PM EST - Seemless transition to AMD EPYC instances
02:01PM EST - EPYC is great for virtualized environments
02:01PM EST - AMD has best in class price/perf for virtualization
02:02PM EST - under industry standard benchmarks
02:02PM EST - VMWare Barcelona this week
02:02PM EST - VMware working closely with AMD
02:03PM EST - VMware to allow AMD customers to work with VMware licences very easily
02:03PM EST - Hyperconvergence for OpEx savings
02:03PM EST - Day-zero support for new technologies on VMware
02:04PM EST - embracing a strategy to redefine the server
02:04PM EST - Replacing the typical 2000 core datacenter deployment
02:04PM EST - Data is very consistent on this
02:05PM EST - As the infrastructure ages, perf goes down and admin costs goes up
02:05PM EST - time to reclaim perf
02:06PM EST - cost-for-cost replacement would be 2S Intel Gold 5118 - 24 core
02:06PM EST - EPYC gives you single socket EPYC 7551P 32 cores with 2TB DRAM support for same price
02:07PM EST - Distruptive TCO: 26% less admin, 29% less hardware, 63% licensing, saves 63% space
02:07PM EST - 45% lower TCO (cost per year per VM
02:09PM EST - EPYC #1 SPECrate2017_fp_peak 2P
02:10PM EST - 44% better weather modelling, 25% faster fluent, 41% faster NAMD
02:10PM EST - Cray to the stage
02:10PM EST - Cray and AMD just announced Shasta, an AMD EPYC based supercomputer
02:11PM EST - Titan with 300k Opteron cores is still #7 on the TOP500
02:12PM EST - Haas F1 team is using Cray CS500 for CFD testing
02:12PM EST - (woohoo)
02:13PM EST - Cray's customers want to run a variety of applications from AI to analytics
02:13PM EST - Shasta allows customers to do both on a single deployment
02:13PM EST - New Cray interconnect called Slingshot
02:14PM EST - NERSC / Lawrence Berkeley has $126m contract for a Shasta
02:15PM EST - 100 PetaFLOP deployment
02:15PM EST - Road to exascale goes through Shasta
02:16PM EST - (sorry, can't upload photos for some reason
02:18PM EST - 'AMD is just getting started'
02:18PM EST - Time for the next horizon announcement
02:19PM EST - Preview of Rome
02:20PM EST - Up to 64 Zen 2 cores, 128 threads
02:20PM EST - Socket compatible with Naples
02:20PM EST - Rome is PCIe 4.0
02:21PM EST - Making the commitment today
02:22PM EST - Performance of Rome:
02:22PM EST - 2x Perf per socket
02:22PM EST - 4x Floating point per socket
02:32PM EST - 8 cores per Zen 2 die
02:33PM EST - Cray benchmark: One socket Rome scored 28.1 seconds, Two 8180M 30.2 seconds
02:33PM EST - Rome air-cooled, non-overclocked, not final frequency
02:34PM EST - On track for 2019
02:35PM EST - That's all for the presetnations. Demos this afternoon
02:35PM EST - I'm going to see what's up with our SQL DB. It needs a kick
02:35PM EST - More info later today
02:35PM EST - .
104 Comments
View All Comments
Galdom - Tuesday, November 6, 2018 - link
And so it begins, the great battle of our time.nathanddrews - Tuesday, November 6, 2018 - link
When in Rome, do as the Romans.Meteor2 - Saturday, November 10, 2018 - link
If those benchmark are even half true that’s still astonishing. Zen 2 seems to blaze past whatever Intel’s latest Core uarch is.FreckledTrout - Tuesday, November 6, 2018 - link
This is not going to be a productive work day for me.The Chill Blueberry - Tuesday, November 6, 2018 - link
Haha, me neither.Chaitanya - Tuesday, November 6, 2018 - link
Its even worse for me in India. I was supposed to go to sleep about an hour back as I have to be in field at 530hrs(IST). Its going go be hard waking up or staying awake tomorrow.ToTTenTranz - Tuesday, November 6, 2018 - link
Ian is already giving us complete spoilers with those tags...Dadyal - Tuesday, November 6, 2018 - link
I hope we hear about next gen AMD high end gpus for gamers and Zen 2 cpus for desktopsSaturnusDK - Tuesday, November 6, 2018 - link
Unlikely. That will be discussed at the CES 2019 keynote.ET - Tuesday, November 6, 2018 - link
We can guess quite a lot from this, for Zen 2, at least, if we assume that the same architecture will be used for the desktop (and mobile), which seems to me like a reasonable assumption. So, 8 cores per chiplet, probably 1-2 chiplets per desktop CPU, 2x floating point performance (and the other architectural enhancements discussed), PCIe 4.