AMD's Future Chips & SoC's: News, Info & Rumours.

[goldstone77]

AMD turned a $25 million operating profit in Q2 2017

https://venturebeat.com/2017/07/25/amd-posts-a-small-q2-loss-despite-its-most-competitive-chips-in-a-decade/

Juan from the same link you posted

“AMD blew away actuals for Q3 and raised expectations for the future,” said Patrick Moorhead, analyst at Moor Insights & Strategy, in an email to VentureBeat. “This came primarily on the back of the first full quarter of Ryzen desktop sales and also on some graphics improvements. AMD has a lot to look forward to as none of their actuals incorporates sales of Epyc server parts, only limited sales of the new Radeon Vega and none of the Ryzen notebook parts. Positive operating profit was icing on the cake and was expected.”

 

[juanrga]

AMD turned a $25 million operating profit in Q2 2017

https://venturebeat.com/2017/07/25/amd-posts-a-small-q2-loss-despite-its-most-competitive-chips-in-a-decade/

Juan from the same link you posted

“AMD blew away actuals for Q3 and raised expectations for the future,” said Patrick Moorhead, analyst at Moor Insights & Strategy, in an email to VentureBeat. “This came primarily on the back of the first full quarter of Ryzen desktop sales and also on some graphics improvements. AMD has a lot to look forward to as none of their actuals incorporates sales of Epyc server parts, only limited sales of the new Radeon Vega and none of the Ryzen notebook parts. Positive operating profit was icing on the cake and was expected.”

My reply to jaymc was to demonstrate that AMD signed another net loss in this quarter, not a profit.

If we want to analyze the quarter, then we can mention that RyZen sales are low (as some have been saying for a while) and that the extra cash did come from GPU sales due to seasonal mining demand.

AMD has been stuck around the breakeven level for a while, but the new Ryzen chips launched as early as March didn’t really help the company much so far. That’s the bad news.

Vega and EPYC will sign similar weak sales (even AMD predictions for rest of year show this), and Zen based Notebooks will meet with improved products from competence.

 

[jaymc]

AMD Epyc server/Ryzen CPUs would be extremely profitable for Rendering/VFX/3D Modeling/Industrial Design:

"They have failed to make a connection to a specific, rapidly growing industries that could potentially bring AMD to knew heights, and that is CGI, game design, VFX, architectural rendering, server rendering, and industrial design.
I am part of this industry (3D Modeler/Rendering Artist), and I have specifically invested my money into AMD after being blown away by the specs/price of the new CPU/GPU/Servers."

https://www.reddit.com/r/AMD_Stock/comments/6q91no/amd_epyc_serverryzen_cpus_would_be_extremely/

"Something else to think about as well is how AMD offers both GPU and CPU, and rendering with Vray now offers a Hybrid method that utilizes both. (Which is an amazing innovation). If EPYC comes equipped with better GPU than most Intel servers, then you would be better off getting that even if Intel fully dropped their prices to match AMD (Which would be a really good move by them).
The reason they are still more expensive despite have the same specs is because they know that can still get away with their brand name.
If AMD stock bombs, and Intel wins, just know that you can learn rendering and build a workstation with AMD that will be more powerful for less money, and you can charge the ever growing companies that need rendered content between 300-1500 dollars an image. Its a good investment to buy their products now."

 

[juanrga]

AMD Epyc server/Ryzen CPUs would be extremely profitable for Rendering/VFX/3D Modeling/Industrial Design:

"They have failed to make a connection to a specific, rapidly growing industries that could potentially bring AMD to knew heights, and that is CGI, game design, VFX, architectural rendering, server rendering, and industrial design.
I am part of this industry (3D Modeler/Rendering Artist), and I have specifically invested my money into AMD after being blown away by the specs/price of the new CPU/GPU/Servers."

https://www.reddit.com/r/AMD_Stock/comments/6q91no/amd_epyc_serverryzen_cpus_would_be_extremely/

"Something else to think about as well is how AMD offers both GPU and CPU, and rendering with Vray now offers a Hybrid method that utilizes both. (Which is an amazing innovation). If EPYC comes equipped with better GPU than most Intel servers, then you would be better off getting that even if Intel fully dropped their prices to match AMD (Which would be a really good move by them).
The reason they are still more expensive despite have the same specs is because they know that can still get away with their brand name.
If AMD stock bombs, and Intel wins, just know that you can learn rendering and build a workstation with AMD that will be more powerful for less money, and you can charge the ever growing companies that need rendered content between 300-1500 dollars an image. Its a good investment to buy their products now."

I did read the whole thread and I have not still decided if he is a very-confused artist or if this is some kind of very elaborated hoax. Something that get me confused is that he posted in the AMD_Stock subreddit.

He claims to work with V-Ray, and he uses buckets (~ threads) to estimate performance. He then initially claimed that his FX 8-core at home gives him 8 buckets, that his 16 core Xeon at work gives him 36 (sic) buckets, and that 16 core Threadripper would give 512 buckets. He then goes further and claims that 32-core EPYC CPU would be even more powerful and reduce rendering times from hours to seconds. He then continues with his anti-Intel and pro-AMD rant and about how the industry of rendering will be revolutionized by EPYC. Well... NO.

As others mentioned in the reddit thread, his computations are incorrect. The top Threadripper chip would provide only 32 buckets (32 threads); i.e. the same number than the old Xeon that he has at work, and the top EPYC chip would give 64 buckets. His claims about reducing rendering times from hours to seconds are also incorrect, and far from reality by many orders of magnitude, even if we ignore that performance is not a linear function of the number of threads that a CPU can process, because factors such as frequency and IPC affect total performance of a CPU. There is a part of the thread where he even claims that GHz matter little:

Dynamic splitting might be dividing it up, but if its processing a smaller area, then ghz doesnt matter as much as threading and core count.

And you can find people in the thread explaining him that GHz and IPC are so important as number of cores and threads. People also explains him that rendering (a throughput-like workload) is a special case for Zen chips:

The rendering is the best case-scenario for AMD Ryzen multi-die architecture connected through Infinity Fabric (low latency interconnect between CCX modules and dies).

This "artist" also compares the last AMD CPUs to former Xeons and then uses that information to claim "This is absolute PROOF" that AMD gives the same performance by about one tenth of the price of an Intel system. Again NO.

Since he uses V-ray, let us check V-Ray benchmark and costs of the CPUs

• 
  ■00:20.517 AMD EPYC 7601 (64 buckets) $4200
  ■00:20.640 Intel Xeon E5-4669 v3 (36 buckets) $5400
  ■00:24.104 Intel Xeon E5-2696 v4 (44 buckets) $1750

The Broadwell Xeon takes 17% more time on rendering than EPYC, but does at a fraction of the cost of the EPYC chip. Where is the revolution?

 

[aldaia]

"Over the past two weeks"... STH having access to last firmware now, doesn't mean that potential customers didn't have access to final firmware months ago. Just as STH having now access to EPYC hardware doesn't mean that customers didn't have access to EPYC hardware months ago.

I forgot that your imaginary friends had access to newest firmware months before it was released by AMD.

 

[gamerk316]

Dynamic splitting might be dividing it up, but if its processing a smaller area, then ghz doesnt matter as much as threading and core count.

Adding cores is simply a multiplyer to base per-core performance. Assuming they get loaded, you get a performance equation that looks like this:

Performance = Single Core Performance * Number of Cores

The number of cores is just as important as single core throughput.

 

[juanrga]

ThreadRipper specs

 

[letsrun4it]

People are really underestimating the 1950X. The alienware computer is garbage. If mobos, psu's, memory matter at all (and isn't this why everyone says we need to build our own PCs) then the ones we make ourselves are going to be an embarrassment of riches.

Here is what I've ordered for parts for my build, will be happy to post my results once I've got this thing put together:

PCPartPicker part list / Price breakdown by merchant

CPU: AMD - Threadripper 1950X 3.4GHz 16-Core Processor ($999.00 @ Amazon)
CPU Cooler: NZXT - Kraken X62 Rev 2 98.2 CFM Liquid CPU Cooler ($159.99 @ Amazon)
Motherboard: MSI - X399 GAMING PRO CARBON AC ATX TR4 Motherboard ($379.99 @ Amazon)
Memory: Corsair - Dominator Platinum 32GB (4 x 8GB) DDR4-3600 Memory ($460.99 @ Amazon)
Storage: Crucial - MX300 1.1TB M.2-2280 Solid State Drive ($274.89 @ Amazon)
Storage: Seagate - Barracuda 3TB 3.5" 7200RPM Internal Hard Drive ($81.78 @ Amazon)
Storage: Seagate - Barracuda 3TB 3.5" 7200RPM Internal Hard Drive ($81.78 @ Amazon)
Video Card: EVGA - GeForce GTX 1080 Ti 11GB FTW3 GAMING iCX Video Card (2-Way SLI) ($779.89 @ Amazon)
Video Card: EVGA - GeForce GTX 1080 Ti 11GB FTW3 GAMING iCX Video Card (2-Way SLI) ($779.89 @ Amazon)
Case: Phanteks - Enthoo Luxe Tempered Glass (Gray) ATX Full Tower Case ($179.99 @ Amazon)
Power Supply: FSP Group - 1000W 80+ Platinum Certified Fully-Modular ATX Power Supply ($199.99 @ Amazon)
Optical Drive: Pioneer - BDR-209DBK Blu-Ray/DVD/CD Writer ($66.99 @ Amazon)
Total: $4445.17
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2017-08-07 03:09 EDT-0400

 

[jaymc]

The Shape Of AMD HPC And AI Iron To Come

www.nextplatform.com/2017/08/08/shape-amd-hpc-ai-iron-come/

 

[8350rocks]

https://www.reddit.com/r/Amd/comments/6sg03p/1950x_geekbench_3_results_41_ghz_all_cores/

Threadripper geekbench scores

 

[juanrga]

If it is a genuine leak then the multiscore at stock settings would be ~51270, which is a 23% higher than the score of the i9-7900X and agrees with the 24% gap that AMD reports in slides for a Blender bench. Thus Intel maintains a ~30% performance gap per core.

 

[aldaia]

STH did a review of EPYC 7061 and more results are coming
https://www.servethehome.com/dual-amd-epyc-7601-processor-performance-and-review-part-1/

 

[-Fran-]

That is an interesting review. Already waiting on part 2.

Cheers!

 

[gamerk316]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

 

[-Fran-]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The street goes both ways this time. It's going to be very interesting to see where the shoe tights each company. I know in mine we're on a per-socket basis, so EPYC options are making the rounds in the internal emails 😀

Cheers!

 

[juanrga]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The dual socket EPYC performed more or less the same than dual socket Gold Xeon on a relevant bench as compiling. The Xeon is cheaper, and it can save additional licensing costs due to having 77% less cores/threads

And single socket Gold Xeon is not doing it bad either

 

[-Fran-]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The dual socket EPYC performed more or less the same than dual socket Gold Xeon on a relevant bench as compiling. The Xeon is cheaper, and it can save additional licensing costs due to having 77% less cores/threads

Spoiler

And single socket Gold Xeon is not doing it bad either

Spoiler

Thank you for graciously ignoring my second paragraph where I say the exact same thing.

Cheers!

 

[juanrga]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The dual socket EPYC performed more or less the same than dual socket Gold Xeon on a relevant bench as compiling. The Xeon is cheaper, and it can save additional licensing costs due to having 77% less cores/threads

Spoiler

And single socket Gold Xeon is not doing it bad either

Spoiler

Thank you for graciously ignoring my second paragraph where I say the exact same thing.

I don't get as your second paragraph contradicts the first paragraph saying "Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them."

I replied this part giving a benchmark that includes an 1P 6132 Xeon which is about one half cheaper than the EPYC chip, despite having better performance per core and better performance per thread on that bench.

What is more, the 6150 Xeon used in the 2P server is significantly cheaper than the 7601 EPYC used in the other 2P server.

 

[-Fran-]

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The dual socket EPYC performed more or less the same than dual socket Gold Xeon on a relevant bench as compiling. The Xeon is cheaper, and it can save additional licensing costs due to having 77% less cores/threads

Spoiler

And single socket Gold Xeon is not doing it bad either

Spoiler

Thank you for graciously ignoring my second paragraph where I say the exact same thing.

I don't get as your second paragraph contradicts the first paragraph saying "Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them."

I replied this part giving a benchmark that includes an 1P 6132 Xeon which is about one half cheaper than the EPYC chip, despite having better performance per core and better performance per thread on that bench.

What is more, the 6150 Xeon used in the 2P server is significantly cheaper than the 7601 EPYC used in the other 2P server.

I am not. Simple reason being, you don't get the same multi-threaded performance from Intel in single socket configs. I'm not denying Intel has the upper hand in a per-thread/core basis, but per socket and going wide, Intel has no answer to EPYC. Don't exclude the whole picture and go check all the tests and group them by socket.

Plus, don't forget, FOSS that goes wide gives AMD a much better bargain status.

Cheers!

 

[aldaia]

Spoiler

The review brought up a point worth considering at the end: Per-core licensing. Thats something we need to consider when talking about cost/performance of EPYC.

Yes and it's something interesting to touch on. A lot of the licensing agreements now are being done on a per-socket basis as well. Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them.

The dual socket EPYC performed more or less the same than dual socket Gold Xeon on a relevant bench as compiling. The Xeon is cheaper, and it can save additional licensing costs due to having 77% less cores/threads

Spoiler

And single socket Gold Xeon is not doing it bad either

Spoiler

Thank you for graciously ignoring my second paragraph where I say the exact same thing.

I don't get as your second paragraph contradicts the first paragraph saying "Which is a huge advantage for 1P systems using EPYC and Intel has no answer (cheap one) for them."

I replied this part giving a benchmark that includes an 1P 6132 Xeon which is about one half cheaper than the EPYC chip, despite having better performance per core and better performance per thread on that bench.

What is more, the 6150 Xeon used in the 2P server is significantly cheaper than the 7601 EPYC used in the other 2P server.

I am not. Simple reason being, you don't get the same multi-threaded performance from Intel in single socket configs. I'm not denying Intel has the upper hand in a per-thread/core basis, but per socket and going wide, Intel has no answer to EPYC. Don't exclude the whole picture and go check all the tests and group them by socket.

Plus, don't forget, FOSS that goes wide gives AMD a much better bargain status.

Cheers!

I think that the ideal comparison should be against xeon 8160 and not 6150. I hope they do so in the upcoming parts.

Anyway, besides highlighting the (few) negative aspects of AMD processors, we must also draw special attention to the strong points. E.g. that EPYC performance is 2,5 to 3 X that of 6150 on c-ray, about 50% faster on 7-zip and 60% on NAMD.

Speaking of NAMD, one of the arguments against Anandtech EPYC review was precisely that NAMD numbers where wrong for Xeon E5-2699 v4. Now we have a second review to check. Speed-up of EPYC 7601 over Xeon E5-2699 v4 on Anantech 50%, speed-up on STH 50%. It's hard to believe (except for Intel fanboys and AMD haters) that both are wrong.

 

[juanrga]

Anyway, besides highlighting the (few) negative aspects of AMD processors, we must also draw special attention to the strong points. E.g. that EPYC performance is 2,5 to 3 X that of 6150 on c-ray, about 50% faster on 7-zip and 60% on NAMD.

Speaking of NAMD, one of the arguments against Anandtech EPYC review was precisely that NAMD numbers where wrong for Xeon E5-2699 v4. Now we have a second review to check. Speed-up of EPYC 7601 over Xeon E5-2699 v4 on Anantech 50%, speed-up on STH 50%. It's hard to believe (except for Intel fanboys and AMD haters) that both are wrong.

C-ray is not a good benchmark, and the reason why the performance gap increases from 50--60% in the other benches to 150--200% in C-ray: : "Whetstone is much better than c-ray 😉"

Spoiler

C-ray:

1) Contains huge number of totally unnecessary sqrt and div calculations.
1a) These could be precalculated outside the loop.
1b) These could be (pre)calculated by much faster newton-rhapsody method.

2) Calculates everything, even pixel color values, with 64-bit doubles
3) Does not use any SIMD.

AFAIK, whetstone does test more than scalar div and sqrt. c-ray practically does not.

EPYC must be about 50% faster on 7-zip and 60% on NAMD than 6150 Xeon, but this is a 18-core SKL Xeon that costs less than the EPYC CPU tested.

People criticized Anandtech review because results for the gcc.403 subtest of the SPEC suite provided scores slower than real scores for the Xeon servers. In their EPYC review the Xeon E5-2699 v4 was 40% slower than measured by other people including a former review of the same Broadwell Xeon by Anandtech.

In the STH review the 18-core Broadwell Xeon is much closer to the 32-core EPYC. In the STH review the old E5-2699 v4 is only about 15% behind in 7-zip, and about 5% behind in the Linux compile test.

 

[aldaia]

Besides the Tom's review here are a couple of very interesting reviews for 1950x and 1920x
http://www.guru3d.com/articles-pages/amd-ryzen-threadripper-1950x-review.html
http://www.guru3d.com/articles-pages/amd-ryzen-threadripper-1920x-review.html

I found particularly interesting the IPC figure. It shows how close it's AMD to Intel and the huge step versus the previous microarchitecture.

 

[aldaia]

1) Contains huge number of totally unnecessary sqrt and div calculations.
1a) These could be precalculated outside the loop.
1b) These could be (pre)calculated by much faster newton-rhapsody method.

2) Calculates everything, even pixel color values, with 64-bit doubles
3) Does not use any SIMD.

1a) Intel can do the same
1b) Intel can do the same
2) And the problem is? There are plenty of applications that use 64-bit doubles
3) And the problem is? There are plenty of applications that cannot be vectorized and therefore SIMD cannot be used.

None of the arguments explains why there is such a difference

People criticized Anandtech review because results for the gcc.403 subtest of the SPEC suite provided scores slower than real scores for the Xeon servers. In their EPYC review the Xeon E5-2699 v4 was 40% slower than measured by other people including a former review of the same Broadwell Xeon by Anandtech.

You are running in circles like a headless chicken here. I already pointed out that "Something strange happens with 403.gcc, and not only in Johan's review. Official SPEC submissions for Xeon 8176 have scores varying by up to 40%." And your answer was "It is not only the gcc subtest, NAMD scores in the review are also weird."
None of the two arguments is strong enough to dismiss anand's review.

 

[juanrga]

Besides the Tom's review here are a couple of very interesting reviews for 1950x and 1920x
http://www.guru3d.com/articles-pages/amd-ryzen-threadripper-1950x-review.html
http://www.guru3d.com/articles-pages/amd-ryzen-threadripper-1920x-review.html

I found particularly interesting the IPC figure. It shows how close it's AMD to Intel and the huge step versus the previous microarchitecture.

Broadwell-E same IPC than Zen is not correct.

 

[aldaia]

Broadwell-E same IPC than Zen is not correct.

Of course is correct. Anyone assuming that measuring performance is an exact science has no clue about computer architecture and performance measurement. Between one figure and the other there is only a 3% difference in IPC for broadwell. I frequently measure performance of multiple IDENTICAL systems under IDENTICAL conditions including room temperature since i benchmark up to 10 systems simultaneously, guess what, the variations between them exceed 3%.

Here we are talking different systems, different processors (6900k vs 6950x) probably with different RAM configs and many other variables (including room temperature) that may influence results. Have you seen that between the top and the bottom Ryzen processor the single thread difference in IPC is 10%? What makes you think that Broadwell is different?

Also the figure you are quoting is from March 2. Do you know that Ryzen has received several microcode improvements that have improved performance in a significant way since then? If there is something wrong here is the outdated figure you posted. Please stop using outdated data to support your fanboyism.