DPS‑8 Performance, Part II

As we noted in the previous post about performance, there are some conflicting performance claims about these systems, depending on where you look. Since the last post was more popular than expected, we’ve decided to do a short follow-up, this time relying on details from various Honeywell sales documents, such as the Series 6000 Summary Description (1971) and Honeywell Large Systems Product Guide (1984).

The Product Guide helpfully includes a Systems Performance Summary chart for various machines, but instead of using absolute numbers, performance is reported relative to other Honeywell machines (such as the Level‑66/80, DPS‑8/47, and DPS‑7/65), so if you aren’t familiar with the Family Tree of GE/Honeywell/Bull 600/6000-series systems, you should take a look.

Thankfully, these documents do have some absolute numbers for select systems scattered throughout (although within the document there still are some conflicts, so we’ve used the numbers that appeared most often—and which made the most sense—when compiling this data). These absolute numbers for the DPS‑8 series are reported as:

While there were other models models and sub‑models introduced later, such as the DPS‑8/46, these five systems represent the core DPS‑8 product line. Fortunately, this performance data aligns with the DPS‑8 numbers from our last post.

The DPS‑8/47 and DPS‑8/49 were built using the ELS or Entry Level System architecture. These were cost‑reduced models manufactured using FAST (Fairchild Advanced Schottky TTL) technology. The traditional DPS‑8/52, DPS‑8/62, and DPS‑8/70 systems were referred to as the DPS‑E (Enhanced or Enterprise) system architecture.

The DPS‑8/47 could be upgraded to the DPS‑8/49 in the field (hinting that the DPS‑8/47 was a limited DPS‑8/49) while the DPS‑8/52 and DPS‑8/62 were known to be impaired versions of the DPS‑8/70 (and were also field upgradable).

For anyone yearning for a deep dive into the advertised multiprocessing efficiency (and other minutiae) of these systems, the Product Guide is worth reviewing. For example, they list the DPS‑8/70 (with a single CPU) as 1.7× a Level‑66/80, the DPS‑8/70‑2 (with 2 CPUs) as 3.1×, a DPS‑8/70‑3 (with 3 CPUs) as 4.5× and a DPS‑8/70‑4 (with 4 CPUs) as 5.9×. Scaling these results indicates the 2‑CPU machine was considered to be 1.82× faster than the single CPU model, the 3‑CPU machine was 2.65× faster, and the 4‑CPU machine was 3.47× faster.

For the purposes of this post, we’ll focus on the single‑CPU performance of the various DPS‑8 and Level‑66 models. It should be noted that while there isn’t specific Level‑68 performance numbers in these documents, the Level‑66 performance numbers can be assumed to be representative of the Level‑68 (Multics) machines.

In the last post we mentioned there were hardware refreshes during the Level‑66/Level‑68 era that affected performance, but we didn’t have more specific data — it turns out the Large Systems Product Guide does provide some of this data (listed in the table below as the P models).

Examining the tables in these documents allows us to make some extrapolations and estimate the absolute performance the single‑CPU models (assuming, of course, that the benchmarks used by Honeywell were mostly CPU bound, which at least appears to be the case, and that the DPS‑8/47 truly is 0.720 MIPS as reported).

We’ve used our calculators so you won’t have to…

As you can see, this estimation calculates 1.884 MIPS for the DPS‑8/70 (very close to the 1.8 MIPS we previously reported), while the 6080 comes in at 0.787 MIPS, a bit below the 0.87 MIPS we previously reported for the 6180.

The introductory (and flagship) Level‑66 model, the Level‑66/80, comes in at 1.108 MIPS, which is right on target with the 1.1 MIPS we previously reported as the performance of the Level‑68.

The first Level‑66 hardware refresh, the Level‑66/80P, improved performance by 16% to 1.274 MIPS.

The refreshed Level‑66/P models could also support NSA, the New System Architecture (or New Six‑Thousand Architecture) extensions which were needed for GCOS‑8 (and CP‑6) via an add‑on from Honeywell. The NSA extensions were always included in DPS‑8 machines.

It should be noted that the 6025 and Level‑66/05 models were low‑performing cost‑reduced “budget” machines. The DPS‑8/46 machine was another lower‑end release which seemingly replaced the DPS‑8/47, but we’re unsure exactly when that happened.

Later on in the same document, the DPS‑8 systems are shown again, but this time with performance relative to the DPS‑8/47 (rather than the Level‑66/80). Using the same methodology, we can produce the following table, which reassuringly tracks closely with our very first table:

Please be aware that these tables are non‑exhaustive. There were many other models produced with various marketing designations such as the Level‑66/07, Level‑66/17, and Level‑66/27, as well as the various Level‑66/DPS and Level‑68/DPS (and even CP‑6 specific configurations) that I don’t know enough about safely attempt to extrapolate performance estimations for. We’ve also excluded the Level‑66/DPS machines from the table since it isn’t clear how many processors these machines were configured with. For instance, this Datapro reference sheet for the series (1982) indicates all Level‑66/DPS and Level‑68/DPS models were sold as multiprocessor configurations, which complicates attempts at scaling relative comparisons to a single CPU, and we’d rather this post be incomplete than incorrect!

We’ve also not (yet) found any similarly specific performance claims for the DPS‑8000, DPS‑90, or DPS‑9000 series.

Readers should be aware that in every known case, the machines sold to run Multics were always the flagship models, that is, the H6180, the Level‑66/80 (or Level‑66/80P), and the DPS‑8/70M. Low‑end configurations, while of course possible, were likely never actually sold. The corresponding flagship models are shown with extra emphasis in tables above.

• DPS8M Performance